Daily Research Digest
arXiv Papers
2026-10-20
478
Papers
9
Categories
97
Translated
收藏清单 0
精选 · Favorites
98
cs.AI / 1 / 2610.25441
Stepwise Reasoning Verification: Don't Generate, Classify!
逐步推理验证：不要生成，要分类！
large language model
大语言模型相关
Abstract
Verifying the reasoning of large language models is increasingly important as they are applied to complex mathematical and scientific problems. Existing work focuses on large decoder models and reports mixed results for generative and discriminative verifiers, leaving unclear which paradigm is more effective and whether large models are necessary. In this work, we cast stepwise verification as a natural language inference task: the problem and partial solution prefix form the premise, while the next reasoning step serves as the hypothesis to be classified as entailed or non-entailed. This formulation enables error localisation at individual reasoning steps and direct comparison of encoder classification, decoder classification, and decoder verdict generation across model scales and annotation budgets. Across five reasoning datasets, a 0.15B encoder classifier outperforms most generative verifiers with up to 4B parameters. On identical decoder backbones, classification outperforms generation from 0.5B to 4B parameters and matches it on the strongest 7B backbone---reaching comparable performance with roughly one tenth as many annotated problems. Our findings establish classification as a parameter- and data-efficient approach to stepwise verification and show that strong verification does not require large generative decoders.
Chinese Translation
随着大语言模型被应用于复杂的数学和科学问题，验证其推理正变得越来越重要。现有工作聚焦于大型解码器模型，并报告了生成式验证器和判别式验证器的混合结果，这使得哪种范式更有效以及大模型是否必要仍不清楚。在这项工作中，我们将逐步验证视为一项自然语言推理任务：问题和部分解前缀构成前提，而下一个推理步骤作为假设，需被分类为蕴含或不蕴含。这一形式化设定使得能够在单个推理步骤上进行错误定位，并能够在不同模型规模和标注预算下直接比较编码器分类、解码器分类和解码器判定生成。在五个推理数据集上，一个0.15B的编码器分类器优于大多数参数量高达4B的生成式验证器。在相同的解码器骨干网络上，分类在0.5B到4B参数范围内优于生成，并在最强的7B骨干网络上与之匹敌——以大约十分之一的标注问题数量达到可比的性能。我们的发现确立了分类作为一种在参数和数据上都高效的逐步验证方法，并表明强验证并不需要大型生成式解码器。
cs.AI / 2 / 2610.25456
System X: An LLM Decision Model That Replies from Its Own KV Cache
System X：一种从自身 KV 缓存进行回复的 LLM 决策模型
large language model
大语言模型相关
Abstract
Decision models return probabilities over supplied options, but a deployed agent may also need to reply in words. The present study introduces System X, a large language model (LLM) decision model that adds a text-reply option to the LLM-as-Jev readout. When this option wins, or confidence is low, the same model continues from the decision's key-value (KV) cache, avoiding a repeated prompt prefill. Qwen3-0.6B has been fine-tuned with LoRA on Reason-9k and Bitext reply rows. JevBench accuracy shows no statistically significant change (55.4% to 54.5%, p=0.87), while expected calibration error falls from 0.292 to 0.169. Held-out Bitext routing accuracy reaches 98.3%. Cache reuse saves 15-37% of latency on an Apple M4 Pro, with token-identical replies in the FP32 demonstration cases. Scoring options on copies of the prompt cache matches full-sequence scoring to within 1.4e-6 in probability. Adding dialogue, industry and complaint data (System X+) improves support-task results but lowers IFEval accuracy from 55.5% to 45.0%. System X v4 uses Qwen3-1.7B, rebalanced data and teacher-rewritten reply targets. Against its base LLM, JevBench accuracy rises from 55.0% to 61.0% (p=0.04), with no statistically significant IFEval change (p=0.71). A matched latency test on unadapted Qwen3-1.7B saves 17-39%. One LLM judge prefers v4 replies to an unadapted 1.7B second call (31 to 11 pairs, p=0.003); the other's preference is not significant. These reply comparisons also differ in prompts and generation limits. State conflicts and unsupported reply details remain failure modes. Further studies with several seeds, human raters and real support data are recommended.
Chinese Translation
决策模型会返回在所提供选项上的概率，但已部署的智能体也可能需要以文字形式回复。本研究引入了 System X，这是一种大型语言模型（LLM）决策模型，它在 LLM-as-Jev 读出中增加了一个文本回复选项。当该选项胜出，或置信度较低时，同一模型会从该决策的键值（KV）缓存继续生成，从而避免重复的提示预填充。Qwen3-0.6B 已在 Reason-9k 和 Bitext 回复行上通过 LoRA 进行微调。JevBench 准确率没有统计学显著变化（55.4% 到 54.5%，p=0.87），而期望校准误差从 0.292 降至 0.169。留出 Bitext 路由准确率达到 98.3%。在 Apple M4 Pro 上，缓存复用节省了 15-37% 的延迟，并且在 FP32 演示案例中回复逐 token 相同。在提示缓存的副本上对选项进行评分，与全序列评分在概率上的差异不超过 1.4e-6。加入对话、行业和投诉数据（System X+）会改善支持任务结果，但会将 IFEval 准确率从 55.5% 降至 45.0%。System X v4 使用 Qwen3-1.7B、重新平衡的数据和教师重写的回复目标。与其基础 LLM 相比，JevBench 准确率从 55.0% 升至 61.0%（p=0.04），而 IFEval 无统计学显著变化（p=0.71）。在未适配的 Qwen3-1.7B 上进行的匹配延迟测试节省了 17-39%。一名 LLM 评判员相比未适配的 1.7B 第二次调用更偏好 v4 回复（31 对 11 对，p=0.003）；另一名的偏好不显著。这些回复比较在提示词和生成限制上也存在差异。状态冲突和无依据的回复细节仍然是失败模式。建议进一步开展使用多个随机种子、人工评分者和真实支持数据的研究。
cs.AI / 3 / 2610.25501
Formal Raising: Embedding Verifiable Ethical Properties into the Development of AI
形式化养育：将可验证的伦理属性嵌入人工智能的开发
large language model
大语言模型相关
Abstract
Alan Turing proposed constructing a ``child machine'' and educating it rather than directly programming an adult-like intelligent machine. Recent work on model raising revives this developmental view for Large Language Models by arguing that moral values should accompany capability acquisition from the beginning rather than be added only through post-hoc alignment. This paper proposes Formal Raising (FR): selected moral values are represented as explicit, machine-checkable properties that participate in development, evaluation, and action selection. We present a C-based reasoning framework in which candidate actions are evaluated against prioritized ethical properties. Conflicts are resolved w.r.t. a given order, strict rules are non-defeasible, and underdetermination is reported explicitly. Exceptions are not encoded as special cases but emerge from the comparison of alternatives. We provide examples that cover conflicts involving life, truthfulness, harm, privacy, autonomy, and strict rules. We outline a property layer inspired by a Property Specification Language (PSL) for state and temporal requirements and discuss its integration into a raising process.
Chinese Translation
艾伦·图灵提出构建一个``儿童机器''并对其进行教育，而不是直接编程一个类似成人的智能机器。近期关于模型养育的研究为大型语言模型重新复兴了这一发展性观点，主张道德价值观应当从一开始就伴随能力的获取，而不是仅通过事后的对齐来添加。本文提出形式化养育（Formal Raising, FR）：选定的道德价值被表示为显式的、机器可检验的属性，这些属性参与开发、评估与行动选择。我们提出一个基于 C 的推理框架，在其中候选行动依据经过优先排序的伦理属性进行评估。冲突依据给定的顺序加以解决，严格规则是不可废止的，而欠定性会被明确报告。例外情况并非被编码为特例，而是从备选方案的比较中涌现出来。我们提供了示例，涵盖涉及生命、真实性、伤害、隐私、自主性以及严格规则的冲突。我们概述了一个受属性规约语言（Property Specification Language, PSL）启发的属性层，用于表达状态与时序需求，并讨论了其融入养育过程的方式。
cs.AI / 4 / 2610.25522
When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice
当共情错失目标：LLM 建议中目标置换的基准
large language model
大语言模型相关
Abstract
Conversational large language models can respond with warmth and reassurance while failing to advance the user's stated goal. We conceptualize this failure as $\textbf{goal displacement}$, the substitution of persona-maintenance content for goal-relevant substance. To measure it, we introduce $\textbf{GDB}$ (Goal Displacement Benchmark), comprising 230 scenarios across nine domains and $4{,}134$ human-scored responses from 18 models. A $2 \times 2$ taxonomy crosses hard constraints with emotional distress, distinguishing failures to confront infeasible goals or combine empathy with practical guidance from reverse displacement, where unsolicited problem-solving replaces requested emotional support. A deletion-test rubric assesses whether substantive help remains after purely supportive sentences are removed. Independent raters achieve weighted $κ= 0.82$ against author scores for displacement severity. Trap displacement ranges from $0.0\%$ to $37.7\%$ and is strongly negatively correlated with Chatbot Arena Elo ($ρ= -0.706$, $n = 15$). Reverse displacement reaches $41.9\%$, including substantial rates in models with negligible trap displacement. A factorial probe supports a distinction from social sycophancy. Human-scored responses in the high-complexity, low-pressure condition exhibit $35.0\%$ sycophancy but $7.5\%$ displacement, with little association between the constructs. Matched OLMo-2 checkpoints show that mixed empathy-with-constraints displacement remains at $52.3\%$ after supervised fine-tuning and falls to $13.8\%$ after direct preference optimization. GDB provides a reproducible diagnostic of whether responses advance the operationalized user goal, supporting model comparison, post-training evaluation, and regression testing. We release the dataset, annotations, and rubric under CC-BY-4.0 and the evaluation code under MIT.
Chinese Translation
对话式大型语言模型可以在回应用户时表现出温暖与安抚，却未能推进用户所陈述的目标。我们将这种失败概念化为 $\textbf{goal displacement}$，即用维持人设的内容替代与目标相关的实质内容。为了测量它，我们引入 $\textbf{GDB}$（Goal Displacement Benchmark），它包含九个领域的 230 个场景，以及来自 18 个模型的 $4{,}134$ 条人工评分回复。一个 $2 \times 2$ 分类法将硬性约束与情绪困扰交叉，区分出两类失败：未能直面不可行目标，或未能将共情与实践指导相结合；并将其与反向置换区分开来，后者指未经请求的问题解决取代了所请求的情感支持。一个删除测试评分标准评估在移除纯支持性句子后是否仍保留实质性帮助。独立评分者在置换严重程度上相对于作者评分达到加权 $κ= 0.82$。陷阱置换的范围从 $0.0\%$ 到 $37.7\%$，并与 Chatbot Arena Elo 呈强负相关（$ρ= -0.706$，$n = 15$）。反向置换达到 $41.9\%$，包括在陷阱置换可忽略的模型中也有可观的比例。一项因子探针支持其与社会性谄媚相区分。在高复杂度、低压力条件下的人工评分回复表现出 $35.0\%$ 的谄媚，但只有 $7.5\%$ 的置换，且这两个构念之间几乎没有关联。匹配的 OLMo-2 检查点显示，混合共情与约束型置换在监督微调后仍为 $52.3\%$，并在直接偏好优化后降至 $13.8\%$。GDB 提供了一种可复现的诊断方法，用于判断回复是否推进了操作化的用户目标，并支持模型比较、训练后评估和回归测试。我们以 CC-BY-4.0 许可发布数据集、标注和评分标准，并以 MIT 许可发布评估代码。
cs.AI / 5 / 2610.25563
Learning Which Device to Call: Online Tool Grounding for LLM Agents in IoT
学习调用哪台设备：面向物联网中 LLM 智能体的在线工具接地
large language model
大语言模型相关
Abstract
The Internet of Things (IoT) gives large language model (LLM) agents access to physical observations, but one requested capability is usually offered by several devices, and the agent's plan does not say which one to call. Existing work retrieves relevant tools and checks plan feasibility, without learning which eligible device returns a useful observation. Unlike previous works, we study online tool grounding: binding an already specified tool call to one physical endpoint under service cost and device participation constraints. We propose a framework that learns which physical device should execute each agent-issued call. It formulates selection as a contextual bandit problem and combines adaptive context partitions with hierarchical endpoint trees, so that execution feedback is reused across related requests and devices. Selection balances task success against service cost while respecting device capability, authorization, battery reserves, and availability. Gaussian reward aggregation provides an optional mechanism for protecting individual feedback values. Theoretical results establish a finite-horizon regret bound and differential privacy for released reward statistics under fixed, reward-independent assignments. In a simulated loading-dock workload, the proposed router achieves higher step success and better cost-adjusted service, at lower simulated device energy and latency, than nearest-device, advertised-quality, UCB1, and LinUCB selection.
Chinese Translation
物联网（IoT）使大型语言模型（LLM）智能体能够获取物理观测，但一项被请求的能力通常由多台设备提供，而智能体的计划并未说明应调用哪一台。现有工作检索相关工具并检查计划可行性，却没有学习哪一台符合条件的设备能返回有用的观测。与以往工作不同，我们研究在线工具接地：在服务成本和设备参与约束下，将已经指定的工具调用绑定到一台物理端点。我们提出一个框架，用于学习应由哪台物理设备执行每个由智能体发出的调用。该框架将选择建模为上下文多臂老虎机问题，并结合自适应上下文划分与分层端点树，从而使执行反馈能够在相关请求和设备之间复用。选择过程在尊重设备能力、授权、电池储备和可用性的同时，平衡任务成功与服务成本。高斯奖励聚合提供了一种可选机制，用于保护单个反馈值。理论结果在固定且与奖励无关的分配下，为所发布的奖励统计量建立了有限时域遗憾界和差分隐私。在模拟的装卸货码头工作负载中，与最近设备、宣称质量、UCB1 和 LinUCB 选择方法相比，所提出的路由器以更低的模拟设备能耗和延迟，实现了更高的单步成功率和更好的成本调整后服务。
cs.AI / 6 / 2610.25640
Think, Compress, Answer: Reinforcement Learning for Multimodal Grounding via Reasoning Compression
思考、压缩、回答：通过推理压缩实现多模态定位的强化学习
large language model
大语言模型相关
Abstract
Extended reasoning has substantially improved multimodal large language models (MLLMs) on tasks requiring compositional and relational inference. However, multimodal grounding imposes a distinct requirement: reasoning must ultimately converge on a target-specific compact representation that supports precise localization. As reasoning trajectories grow longer, they often expand into broad scene narratives and peripheral relational details, which can dilute grounded perceptual evidence and lead to ambiguous or incorrect predictions. Motivated by the information bottleneck principle, we propose Think-Compress-Answer (TCA), a reinforcement learning framework that decomposes multimodal grounding into three phases: free-form reasoning, target-aware context compression, and answer generation conditioned on the compressed context. TCA is optimized through a multi-stage Group Relative Policy Optimization (GRPO) with progressive rewards that establish the structured output protocol, minimize reasoning-specific information while preserving target-relevant content, and directly improve localization accuracy. This staged design enables MLLMs to retain the benefits of extended reasoning while distilling it into the minimal context required for grounding. Experiments across six multimodal grounding benchmarks show that TCA attains substantial zero-shot performance gains, demonstrating the effectiveness of reasoning compression.
Chinese Translation
扩展推理显著提升了多模态大语言模型（MLLMs）在需要组合推理和关系推理的任务上的表现。然而，多模态定位提出了一个不同的要求：推理最终必须收敛到一种支持精确定位的、目标特定的紧凑表示。随着推理轨迹变长，它们往往会扩展为宽泛的场景叙述和外围关系细节，这可能稀释与定位相关的感知证据，并导致模糊或错误的预测。受信息瓶颈原理启发，我们提出了 Think-Compress-Answer（TCA），这是一个将多模态定位分解为三个阶段的强化学习框架：自由形式推理、目标感知上下文压缩，以及以压缩上下文为条件的答案生成。TCA 通过多阶段组相对策略优化（GRPO）进行优化，并采用渐进式奖励，以建立结构化输出协议、在保留目标相关内容的同时最小化推理特定信息，并直接提升定位准确率。这种分阶段设计使 MLLMs 能够保留扩展推理的优势，同时将其蒸馏为定位所需的最小上下文。在六个多模态定位基准上的实验表明，TCA 取得了显著的零样本性能提升，证明了推理压缩的有效性。
cs.AI / 7 / 2610.25675
LIRQuant: An Efficient Locally Isolated Rotation Quantization Method for Large Language Models
LIRQuant：一种面向大语言模型的高效局部隔离旋转量化方法
large language model
大语言模型相关
Abstract
Post-Training Quantization (PTQ) enables efficient deployment of Large Language Models (LLMs), yet activation outliers remain a key challenge for low-bit quantization. Existing global transformation methods overlook heterogeneous channel sensitivity to quantization noise, leading to suboptimal accuracy-efficiency trade-offs. To address this, we propose \textbf{Locally Isolated Rotation Quantization (LIRQuant)}, a sensitivity-aware framework that selectively applies transformations. LIRQuant identifies sensitive channels using a noise-weighted sensitivity metric, applying expressive learnable rotations only to these channels and lightweight fixed rotations to the remaining ones. This decoupled design reduces optimization complexity without sacrificing outlier smoothing capability. Extensive experiments across the LLaMA-2/3, Mistral, Qwen3, and Mixtral families demonstrate that LIRQuant achieves state-of-the-art accuracy under the challenging W4A4KV4 setting. Meanwhile, LIRQuant calibrates LLaMA-3-8B in only 25 minutes with 20.3 GB peak GPU memory, achieving the fastest calibration and the lowest memory footprint among recent learnable-rotation methods, and its fused inference kernels deliver up to $2.3\times$ prefill speedup and $3.5\times$ memory saving over FP16.
Chinese Translation
训练后量化（PTQ）能够实现大语言模型（LLMs）的高效部署，但激活异常值仍是低比特量化的关键挑战。现有的全局变换方法忽略了通道对量化噪声的异构敏感性，导致精度与效率的权衡次优。为解决这一问题，我们提出\textbf{局部隔离旋转量化（LIRQuant）}，一种选择性施加变换的敏感性感知框架。LIRQuant 使用噪声加权敏感性度量识别敏感通道，仅对这些通道施加具有表达能力的可学习旋转，而对剩余通道施加轻量级固定旋转。这种解耦设计降低了优化复杂度，同时不牺牲异常值平滑能力。在 LLaMA-2/3、Mistral、Qwen3 和 Mixtral 系列上的大量实验表明，LIRQuant 在具有挑战性的 W4A4KV4 设置下实现了最先进的精度。同时，LIRQuant 仅需 25 分钟、以 20.3 GB 峰值 GPU 内存即可完成 LLaMA-3-8B 的校准，在近期可学习旋转方法中实现了最快的校准速度和最低的内存占用，其融合推理内核相较 FP16 可带来最高 $2.3\times$ 的预填充加速和 $3.5\times$ 的内存节省。
cs.AI / 8 / 2610.25682
BitFlat: Error-Localized Bitstream Design for Robust Diffusion Image Compression
BitFlat：面向鲁棒扩散图像压缩的错误局部化比特流设计
diffusion
扩散模型相关
Abstract
Zero-shot diffusion image compression based on reverse channel coding achieves high perceptual quality at low bitrates without retraining, yet its bitstream remains vulnerable to bit flips. Prior work has shown that a joint combinatorial representation can amplify a single bit flip into substitutions of multiple atoms and that indexing each atom independently confines this error propagation. However, independent indexing reduces the number of atoms that fit within the bit budget and gives the decoder no indication of which atoms are corrupted. To address these issues, we propose BitFlat, a training-free error-localizing bitstream framework with two components. Blockwise Locality Alignment (BLA) decomposes the globally coupled atom selection at each diffusion step into independent selection units, each describing one signed atom from a dedicated codebook with a fixed-width field, so that any bit error affects only the unit in which it occurs. Reliability-Aware Diffusion Control (RADC) uses the bits saved by BLA to impose structured valid-codeword constraints within each field and converts the decoding state of each field into a binary reliability gate, which turns detected corrupted atoms into erasures at known positions and excludes them from noise synthesis. BitFlat requires neither target residuals nor additional network forward passes and leaves the diffusion model unchanged. Experiments on Kodak, DIV2K, and CLIC2020 show that BitFlat outperforms conventional, learned, and diffusion-based codecs at bit error rates from $10^{-4}$ to $10^{-2}$ and, even with $26.7\%$ fewer bits, surpasses Robust Turbo-DDCM.
Chinese Translation
基于反向信道编码的零样本扩散图像压缩无需重新训练即可在低比特率下实现高感知质量，然而其比特流仍然容易受到比特翻转的影响。先前工作表明，联合组合表示会将单个比特翻转放大为多个原子的替换，而独立索引每个原子可以限制这种错误传播。然而，独立索引减少了在比特预算内可容纳的原子数量，并且没有向解码器提供哪些原子已损坏的任何指示。为解决这些问题，我们提出BitFlat，一种无需训练的、具有两个组件的错误局部化比特流框架。块级局部性对齐（BLA）将每个扩散步骤中全局耦合的原子选择分解为独立的选择单元，每个选择单元用固定宽度字段描述来自专用码本的一个带符号原子，从而使任何比特错误仅影响其所在的单元。可靠性感知扩散控制（RADC）利用BLA节省的比特在每个字段内施加结构化的有效码字约束，并将每个字段的解码状态转换为二值可靠性门，从而将检测到的损坏原子转化为已知位置上的擦除，并将其排除在噪声合成之外。BitFlat既不需要目标残差，也不需要额外的网络前向传播，并且不改变扩散模型。在Kodak、DIV2K和CLIC2020上的实验表明，在从$10^{-4}$到$10^{-2}$的比特错误率下，BitFlat优于传统、学习和基于扩散的编解码器，并且即使比特数少$26.7\%$，也超过Robust Turbo-DDCM。
cs.AI / 9 / 2610.25683
When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight Perturbations
当安全集体失效：通过联合权重扰动实现抵御有害微调的鲁棒对齐
large language model
大语言模型相关
Abstract
Harmful fine-tuning can rapidly erode the safety alignment of large language models, creating substantial risks when aligned models are adapted to downstream tasks. Existing alignment-stage defenses use harmful examples and simulated harmful perturbations during safety alignment to prepare models for later harmful fine-tuning, yet the collective effects of harmful weight perturbations remain underexplored. We identify a collective failure pattern: disjoint parts of a harmful perturbation have limited effects in isolation, but their joint application increases the model's harmfulness beyond the sum of their individual effects. Guided by this observation, we propose Joint Weight Perturbations (JWP), an alignment-stage defense that masks different parts of a harmful perturbation estimated from the current model and trains the model to respond safely in the resulting states. This exposes the model to different combinations of harmful perturbations and reduces its sensitivity to their accumulation during subsequent harmful fine-tuning. Experiments across downstream tasks and model families show that JWP offers stronger resistance to harmful fine-tuning than prior defenses while largely preserving task performance.
Chinese Translation
有害微调可能迅速侵蚀大语言模型的安全对齐，当已对齐模型被适配到下游任务时会带来重大风险。现有的对齐阶段防御在安全对齐期间使用有害样本和模拟的有害扰动，以使模型为后续的有害微调做好准备，然而有害权重扰动的集体效应仍未得到充分探索。我们识别出一种集体失效模式：有害扰动的互不相交部分在单独作用时影响有限，但它们的联合施加会使模型的危害性增加超过其各自效应之和。受这一观察启发，我们提出联合权重扰动（JWP），一种对齐阶段防御方法，它屏蔽从当前模型估计出的有害扰动的不同部分，并训练模型在由此得到的状态下作出安全响应。这使模型接触到有害扰动的不同组合，并降低其在后续有害微调过程中对这些扰动累积的敏感性。跨下游任务和模型系列的实验表明，与先前的防御方法相比，JWP对有害微调提供了更强的抵抗力，同时基本保持了任务性能。
cs.AI / 10 / 2610.25697
CVaR-RMD: CVaR Selection Pressure for Robust Metaheuristic Discovery
CVaR-RMD：面向稳健元启发式发现的CVaR选择压力
large language model
大语言模型相关
Abstract
Large language models have recently been used to discover heuristics through iterative program search, offering a promising approach to automated algorithm design. However, candidates selected by average performance may degrade across instance families, random seeds, and problem scales because mean-based objectives can obscure rare but severe failures. We propose CVaR-RMD, a risk-sensitive evolutionary framework that uses an LLM to mutate executable heuristic programs and ranks candidates by empirical conditional value-at-risk over heterogeneous evaluation conditions. Across three combinatorial optimization domains, CVaR-based selection improves upper-tail performance and reduces failure rates while maintaining competitive mean performance. These gains persist on larger unseen instances and independently curated within-domain benchmarks
Chinese Translation
大型语言模型近来已被用于通过迭代式程序搜索来发现启发式算法，为自动化算法设计提供了一条有前景的途径。然而，按平均性能选出的候选算法可能会在不同实例族、随机种子和问题规模上出现性能退化，因为基于均值的目标可能掩盖罕见但严重的失败。我们提出CVaR-RMD，这是一个风险敏感的进化框架，它使用大语言模型对可执行的启发式程序进行变异，并依据在异构评估条件下的经验条件风险价值对候选进行排序。在三个组合优化领域上，基于CVaR的选择提升了上尾性能并降低了失败率，同时保持了有竞争力的平均性能。这些收益在更大的未见实例以及独立构建的领域内基准上依然成立。
cs.AI / 11 / 2610.25700
From Black Box to Controller: Inference-Time Steering of LLM via Feature Amplification
从黑箱到控制器：通过特征放大在推理时引导大语言模型
large language model
大语言模型相关
Abstract
Inference-time steering is crucial for deploying Large Language Models (LLMs) in real-world applications, where they must flexibly adjust behavior preferences. For instance, a model should dynamically adjust its caution level based on the risk of the scenario. This requirement introduces a challenge: modulating the intensity of LLM preferences. Prompt engineering can require task-specific prompt design, while fine-tuning requires additional optimization and data. These requirements can make repeated behavioral adjustments costly. To address this, we introduce NeuroSteer, a plug-and-play framework that enables flexible, quantitative adjustment of LLM preferences at inference time. NeuroSteer employs a Sparse Autoencoder (SAE) as its controller. Comparing neuron activation frequencies between positive and negative examples isolates "feature neurons" that correspond to target preferences such as politeness and positive sentiment. During inference, the activations of selected SAE features are amplified in the latent representations to steer the output distribution. Users can adjust the strength of this steering on a per-scenario or per-user basis, granting the LLM fine-grained, quantifiable trait modulation. Our experiments show that NeuroSteer can alter sentiment, toxicity, stance, and politeness at inference time. It achieves higher target-direction steering scores than the compared baselines in several settings, with task- and strength-dependent trade-offs in generation quality.
Chinese Translation
推理时引导对于在真实世界应用中部署大语言模型（LLMs）至关重要，在这些应用中模型必须灵活地调整行为偏好。例如，模型应当根据场景的风险动态调整其谨慎程度。这一要求带来了一个挑战：调节 LLM 偏好的强度。提示工程可能需要针对具体任务设计提示，而微调则需要额外的优化与数据。这些要求会使反复的行为调整代价高昂。为解决这一问题，我们提出 NeuroSteer，一个即插即用的框架，能够在推理时对 LLM 偏好进行灵活、定量的调整。NeuroSteer 采用稀疏自编码器（SAE）作为其控制器。通过比较正例与负例之间的神经元激活频率，可以分离出对应于目标偏好（如礼貌性和积极情感）的“特征神经元”。在推理过程中，所选 SAE 特征在潜在表示中的激活被放大，以引导输出分布。用户可以按场景或按用户调整这种引导的强度，从而赋予 LLM 细粒度、可量化的特质调节能力。我们的实验表明，NeuroSteer 能够在推理时改变情感、毒性、立场和礼貌性。在多种设置下，它取得了比所比较的基线更高的目标方向引导分数，同时在生成质量上存在依赖于任务和强度的权衡。
cs.AI / 12 / 2610.25712
PFedQFT: Personalized Federated Quantum-Inspired Fine-Tuning for Medical Large Language Models
PFedQFT：面向医学大语言模型的个性化联邦量子启发式微调
large language model
大语言模型相关
Abstract
Medical data remain fragmented across institutional silos because of privacy, governance, and ownership constraints. Federated learning enables medical institutions to adapt large language models (LLMs) collaboratively without transferring raw records. However, independently parameterized low-rank adapters incur repeated parameter and communication costs across Transformer projections, while uniform aggregation may limit adaptation to heterogeneous institutional data. We propose PFedQFT, a quantum-inspired framework for personalized federated fine-tuning with a compact shared adapter. Specifically, PFedQFT constructs reusable orthonormal adaptation bases from compact skew-symmetric generators using a truncated matrix-exponential approximation followed by QR orthonormalization. Compatible projections reuse these bases, while a globally shared full core preserves the flexibility of each projection. Each client further maintains a lightweight private residual operator, whose directions are learned globally and whose coefficients remain local. Experiments on seven public medical tasks and multiple LLM backbones show substantial reductions in trainable parameters and per-round adapter upload volume, with competitive predictive performance on most tasks.
Chinese Translation
由于隐私、治理和所有权方面的限制，医学数据仍然碎片化地分散在各机构孤岛中。联邦学习使医学机构能够在不传输原始记录的情况下协同适配大语言模型（LLMs）。然而，独立参数化的低秩适配器会在各 Transformer 投影中产生重复的参数和通信开销，而统一聚合可能会限制对异构机构数据的适配。我们提出 PFedQFT，一种采用紧凑共享适配器的量子启发式个性化联邦微调框架。具体而言，PFedQFT 使用截断矩阵指数近似并随后进行 QR 正交归一化，从紧凑的斜对称生成器构造可复用的标准正交适配基。兼容的投影复用这些基，而全局共享的完整核心保留每个投影的灵活性。每个客户端还维护一个轻量级私有残差算子，其方向在全局学习，而其系数保持本地。在七个公共医学任务和多个 LLM 骨干模型上的实验表明，可训练参数和每轮适配器上传量显著降低，同时在大多数任务上具有有竞争力的预测性能。
cs.AI / 13 / 2610.25777
Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills
站在巨人的肩膀上：利用人类专家技能增强用于自动事实核查的 LLM 智能体
large language model
大语言模型相关
Abstract
Automated fact-checking (AFC) systems powered by large language model (LLM) agents still face challenges in evidence retrieval and claim verification. Existing approaches improve their capabilities through computationally costly post-training or human-in-the-loop feedback that is difficult to scale. To transfer professional fact-checking expertise to LLM agents efficiently, we distill publicly available guidelines published by Agence France-Presse (AFP) Fact Check into a reusable, modular Professional Fact-Checking Skill (PFC-Skill). We then present SkillCheck, the first AFC system that gives LLM agents dynamic access to PFC-Skill to guide evidence retrieval and claim verification at inference time. Extensive experiments demonstrate that SkillCheck improves upon skill-free DEFAME baselines across multiple LLM backbones, with gains of 2.2--4.8 pp. in Accuracy and 1.5--6.3 pp. in Macro-F1 on the AVeriTeC benchmark. Notably, SkillCheck with GPT-4o-Mini exceeds DEFAME with GPT-4o by 3.4 pp. in Accuracy while cutting inference cost by roughly 61%. The performance gains also extend to open-web retrieval. Ablations show that incorporating the skill during evidence retrieval provides most of the gains, while incorporation into both stages yields the best overall performance. Case studies show that PFC-Skill enables more targeted search and closer scrutiny of evidence. To conclude, our work shows that LLM agents can "stand on the shoulders of giants" by drawing on the best practices of professional fact-checkers, substantially enhancing AFC performance without extra training or continuous human supervision.
Chinese Translation
由大型语言模型（LLM）智能体驱动的自动事实核查（AFC）系统在证据检索和声明验证方面仍面临挑战。现有方法通过计算成本高昂的后训练或难以规模化的人在回路反馈来提升其能力。为了高效地将专业事实核查专长迁移给 LLM 智能体，我们将法新社（Agence France-Presse, AFP）事实核查发布的公开指南提炼为可复用、模块化的专业事实核查技能（Professional Fact-Checking Skill, PFC-Skill）。随后，我们提出 SkillCheck，这是首个让 LLM 智能体在推理时动态访问 PFC-Skill 以指导证据检索和声明验证的 AFC 系统。大量实验表明，SkillCheck 在多个 LLM 主干模型上优于无技能的 DEFAME 基线，在 AVeriTeC 基准上准确率提升 2.2--4.8 个百分点，Macro-F1 提升 1.5--6.3 个百分点。值得注意的是，使用 GPT-4o-Mini 的 SkillCheck 在准确率上超过使用 GPT-4o 的 DEFAME 3.4 个百分点，同时将推理成本降低约 61%。性能提升也扩展到开放网络检索。消融实验表明，在证据检索阶段引入该技能带来了大部分收益，而在两个阶段均引入时总体性能最佳。案例研究表明，PFC-Skill 能够实现更有针对性的搜索和对证据更细致的审查。总之，我们的工作表明，LLM 智能体可以借鉴专业事实核查人员的最佳实践，“站在巨人的肩膀上”，在无需额外训练或持续人工监督的情况下显著增强 AFC 性能。
cs.AI / 14 / 2610.25819
When to Stop Asking: Factor Coverage Bias and LLM Determinability Judgments with Underspecified Information in Medicine
何时停止询问：医学中信息不充分时的因素覆盖偏差与LLM可判定性判断
large language model
大语言模型相关
Abstract
Clinical decision-making often begins with underspecified information and requires not only gathering missing information but also knowing when to stop. This calls for judging determinability (whether the available information is sufficient for a decision) by considering whether missing information could change the conclusion, not simply the amount of available decision-relevant information. Reliance on this amount as a cue for determinability, which we term \textit{factor coverage bias}, could undermine the trustworthiness of large language models (LLMs) through premature conclusions or excessive information gathering, such as unnecessary invasive tests. Yet whether LLMs exhibit this bias remains insufficiently studied. To address this gap, we construct a controlled pairwise benchmark in which LLMs identify the determinable case within each determinable--undeterminable pair. We add a decision-relevant factor to either side while keeping each case's determinability unchanged, and similarly add irrelevant text to distinguish this bias from context-length effects. Adding a relevant factor made LLMs more likely to select the augmented case as determinable even when it remained undeterminable, an effect that persisted with increased reasoning effort, while adding irrelevant text made them less likely to select the augmented case, an effect that reasoning substantially reduced. These findings identify factor coverage bias as a failure mode distinct from context-length effects and a potential risk to LLM trustworthiness. Our work provides an evaluation framework for LLM trustworthiness under underspecification.
Chinese Translation
临床决策常常始于信息不充分的情况，并且不仅需要收集缺失的信息，还需要知道何时停止。这要求通过考虑缺失信息是否可能改变结论来判断可判定性（即现有信息是否足以作出决策），而不是仅仅依据可获得的决策相关信息的数量。将这一数量作为可判定性的线索来依赖——我们将其称为\textit{因素覆盖偏差}——可能通过过早下结论或过度收集信息（例如不必要的侵入性检查）而损害大型语言模型（LLMs）的可信度。然而，LLMs是否会表现出这种偏差仍未得到充分研究。为填补这一空白，我们构建了一个受控的成对基准，其中LLMs在每一对可判定—不可判定的案例中识别出可判定的案例。我们在任一侧添加一个决策相关因素，同时保持每个案例的可判定性不变，并类似地添加无关文本，以将这种偏差与上下文长度效应区分开来。添加相关因素使LLMs更可能将增强后的案例选为可判定，即使该案例仍不可判定；这一效应在增加推理努力后仍然存在，而添加无关文本则使它们更不可能选择增强后的案例，这一效应被推理大幅削弱。这些发现将因素覆盖偏差识别为一种不同于上下文长度效应的失败模式，并且是对LLM可信度的一种潜在风险。我们的工作为信息不充分条件下LLM的可信度提供了一个评估框架。
cs.AI / 15 / 2610.25843
A Cue's Hidden Price: Auditing Seller Influence on AI Shoppers
线索的隐藏价格：审计卖家对 AI 购物者的影响
large language model
大语言模型相关
Abstract
Large language models are increasingly asked to shop on a user's behalf, which lets seller-controlled content such as badges, endorsements, and framing steer what they recommend. That poses the question at the core of generative engine optimization (GEO): not whether seller content shifts an agent's choice, but how much. The usual gauge of that influence, how often a cue wins the choice, has a blind spot: once a cue is strong enough to be picked almost every time, the choice rate can no longer show how much stronger one cue is than another. We introduce price-equivalent (PE) auditing, which removes this ceiling by pricing the cue in money. In a counterbalanced, position-controlled forced choice, we raise the cued product's price until it is chosen only half the time, and report that premium as the cue's price-equivalent: a formally defined, policy-conditioned parity threshold that makes very different cues comparable. Priced this way, cues that look equally strong at equal price differ by a wide margin, so choice-rate audits substantially under-resolve seller influence. We run the audit over 113,472 trials, seven models across families, and two product categories. To our knowledge, we also run the first price-equivalent audit of a decision-native model, Jev, which outputs a choice probability directly instead of generating text; strikingly, its profile is unlike any text model we tested, resisting most cues yet remaining susceptible to the strongest social-proof claims. We release the stimuli, logs, and code as a benchmark that reports each model's cue-price profile rather than a single score, giving developers and auditors a common, reproducible way to measure how content moves AI shopping agents.
Chinese Translation
大型语言模型越来越多地被要求代替用户购物，这使得卖家控制的内容（如徽章、背书和措辞框架）能够左右它们推荐什么。这就提出了生成引擎优化（GEO）的核心问题：不是卖家内容是否会改变智能体的选择，而是改变了多少。衡量这种影响的常用指标——即线索赢得选择的频率——存在一个盲点：一旦某个线索强到几乎每次都被选中，选择率就无法再显示一个线索比另一个强多少。我们引入价格等价（PE）审计，它通过用金钱为线索定价来消除这一上限。在一个平衡化、位置受控的强制选择中，我们提高被线索标记产品的价格，直到它只有一半的次数被选中，并将该溢价报告为线索的价格等价：一个形式化定义、以策略为条件的平价阈值，使得非常不同的线索可以相互比较。以这种方式定价，在同等价格下看起来强度相同的线索之间相差很大，因此选择率审计在很大程度上无法充分分辨卖家影响。我们在 113,472 次试验、跨模型家族的七个模型以及两个产品类别上运行了该审计。据我们所知，我们还首次对一个决策原生模型 Jev 进行了价格等价审计，该模型直接输出选择概率而不是生成文本；引人注目的是，它的画像与我们测试的任何文本模型都不同，它抵御了大多数线索，却仍然易受最强的社会认同类宣称的影响。我们发布刺激材料、日志和代码，作为一个报告每个模型的线索-价格画像而非单一分数的基准，为开发者和审计者提供一种共同、可复现的方式来衡量内容如何影响 AI 购物智能体。
cs.AI / 16 / 2610.25858
Are Self-Evolving Agents Continual Learners?
自进化智能体是持续学习者吗？
large language model
大语言模型相关
Abstract
Large language model agents are often deployed in environments where new knowledge, tools, and workflows keep arriving. Such an agent is therefore expected to adapt continually, learning new tasks as they arrive without forgetting earlier ones. This resembles the continual learning problem, but at a pace that conventional gradient updates cannot sustain. Nonparametric self-evolving agents are a natural alternative: they adapt by rewriting their own harness, such as prompts, memory, and skills, from task feedback while the model weights stay frozen. We study whether such agents are effective continual learners across four drifting task streams that shift in topic, workflow, or domain (built from BIRD, BFCL, and ScienceWorld). Evaluating three representative methods spanning experience retrieval and self-modification (ReasoningBank, GEPA, and Meta-Harness), we find that although self-evolution improves aggregate performance on most streams, higher plasticity consistently comes at the cost of stability. Specifically, self-modifying methods apply global edits to a shared prompt or workspace, interfering with unseen tasks and overwriting earlier knowledge as the stream shifts. Conversely, similarity-based memory retrieval localizes updates across distant domains, but still retrieves foreign memories on unseen tasks and suffers from retrieval collision when similar queries require conflicting workflows. These findings show that current self-evolving agents remain incomplete continual learners.
Chinese Translation
大型语言模型智能体通常被部署在新知识、新工具和新工作流不断涌现的环境中。因此，这样的智能体被期望能够持续适应，在到来时学习新任务，同时不遗忘早先的任务。这类似于持续学习问题，但其节奏是常规梯度更新无法维持的。非参数化自进化智能体是一种自然的替代方案：它们通过从任务反馈中重写自身的执行框架（harness），例如提示词、记忆和技能，来进行适应，而模型权重保持冻结。我们研究了这类智能体在四个漂移的任务流（分别在主题、工作流或领域上发生偏移，基于 BIRD、BFCL 和 ScienceWorld 构建）上是否是有效的持续学习者。我们评估了涵盖经验检索与自我修改的三类代表性方法（ReasoningBank、GEPA 和 Meta-Harness），发现尽管自进化在大多数任务流上提升了总体性能，但更高的可塑性始终以稳定性为代价。具体而言，自我修改方法对共享的提示词或工作区施加全局编辑，随着任务流的偏移，会干扰未见过的任务并覆盖早先的知识。相反，基于相似度的记忆检索将更新局部化于相距较远的领域，但在未见任务上仍会检索到外来记忆，并且当相似的查询需要相互冲突的工作流时，会遭遇检索碰撞。这些发现表明，当前的自进化智能体仍是不完整的持续学习者。
cs.AI / 17 / 2610.25949
OrderGPT: Generation Order as a Learnable Modeling Choice
OrderGPT：生成顺序作为一种可学习的建模选择
diffusion
扩散模型相关
Abstract
Generative modeling of discrete data has two popular paradigms: autoregressive models such as GPT, which factorize a sequence in a left-to-right (L2R) order, and masked diffusion models, which are essentially any-order autoregressive modeling. Rather than competing methods, these are the two extreme points of a single axis -- the generation order. For structured discrete domains without a natural traversal, both extremes may be suboptimal, yet the intermediate choice -- a single well-adapted order -- remains largely unexplored. We introduce OrderGPT, which builds on Any-Order GPT to learn a task-level generation order without canonical-order supervision. Our key observation is that predicting variables from randomly sampled reveal prefixes produces recurring attention patterns in the backbone that provide a model-dependent signal of predictive dependencies. We aggregate this signal into a reciprocal dependency graph and use a spectral teacher to derive continuous generation priorities. A lightweight OrderMLP integrated into the GPT backbone learns to predict these priorities from attention, incorporating order learning into the architecture within a single training run. Training gradually shifts toward the learned order, which is then frozen for subsequent autoregressive training and generation. The same mechanism adapts across modalities: adaptively it recovers near-L2R dependency structure in language, learns anatomically organized factorizations on human pose, and discovers a factorization on tabular data that places statistically and semantically related fields nearby, while improving downstream modeling metrics across all domains. The code is available at https://github.com/Chenhe-yang/OrderGPT/.
Chinese Translation
离散数据的生成建模有两种流行的范式：自回归模型（如 GPT）按从左到右（L2R）顺序对序列进行因子分解，以及掩码扩散模型，其本质是任意顺序自回归建模。这些并非相互竞争的方法，而是单一轴——生成顺序——上的两个极端点。对于没有自然遍历顺序的结构化离散领域，这两个极端可能都是次优的，而中间选择——一个良好适配的单一顺序——在很大程度上仍未被探索。我们提出 OrderGPT，它构建于 Any-Order GPT 之上，无需规范顺序监督即可学习任务级生成顺序。我们的关键观察是，从随机采样的揭示前缀预测变量会在骨干网络中产生反复出现的注意力模式，这些模式提供了预测依赖性的模型相关信号。我们将该信号聚合为互惠依赖图，并使用谱教师来推导连续的生成优先级。集成到 GPT 骨干中的轻量级 OrderMLP 学习从注意力中预测这些优先级，从而在单次训练运行中将顺序学习纳入架构。训练逐渐转向学习到的顺序，随后该顺序被冻结，用于后续的自回归训练和生成。同一机制可跨模态适配：它自适应地在语言中恢复接近 L2R 的依赖结构，在人体姿态上学习按解剖结构组织的因子分解，并在表格数据上发现一种将统计和语义相关字段放在附近的因子分解，同时在所有领域改善下游建模指标。代码可在 https://github.com/Chenhe-yang/OrderGPT/ 获取。
cs.AI / 18 / 2610.25957
Productive State-Space Exploration for Multi-turn RL in LLM Agents
面向LLM智能体多轮强化学习的富有成效的状态空间探索
large language model
大语言模型相关
Abstract
Reinforcement Learning (RL) for Large Language Model (LLM) agents in multi-turn, long-horizon environments remains challenging due to inefficient action selection and repetitive behavioral patterns. In particular, agents often become trapped in high-confidence yet unproductive behaviors, repeatedly executing invalid actions or entering action loops that hinder task completion. We observe that these frequently violate environment preconditions, revealing a mismatch between the agent's representation of the internal state and the dynamics of the underlying environment. In addition, repetitive high-confidence actions lead to cyclical behaviors that provide little to no progress towards the goal. We argue that addressing these challenges requires more productive task-aligned exploration to improve the agent's understanding of environment dynamics. Motivated by intrinsic curiosity-driven exploration, we propose a conditional state-space exploration framework that balances task completion with intrinsic information-seeking behavior. Specifically, we introduce an intrinsic reward that penalizes actions whose observation duplicates prior state memory. This provides the benefit of reducing redundant behaviors and encouraging state-space coverage without relying on a general information-theoretic measure. Strong experimental results demonstrate that our approach substantially improves policy learning in multi-turn environments, mitigating invalid actions and repetitive loops.
Chinese Translation
由于动作选择效率低下以及重复性的行为模式，面向多轮、长时程环境的大型语言模型（LLM）智能体的强化学习（RL）仍然具有挑战性。具体而言，智能体常常陷入高置信度但无成效的行为之中，反复执行无效动作，或进入阻碍任务完成的动作循环。我们观察到，这些行为经常违反环境前置条件，暴露出智能体对内部状态的表征与底层环境的动态之间存在不匹配。此外，重复的高置信度动作会导致循环行为，这些行为对目标的推进贡献甚微，甚至毫无贡献。我们认为，应对这些挑战需要更加富有成效、与任务对齐的探索，以增进智能体对环境动态的理解。受内在好奇心驱动探索的启发，我们提出了一种条件状态空间探索框架，在任务完成与内在信息寻求行为之间取得平衡。具体而言，我们引入了一种内在奖励，对观测结果与先前状态记忆重复的动作进行惩罚。此举的益处在于：无需依赖通用的信息论度量，即可减少冗余行为并鼓励对状态空间的覆盖。有力的实验结果表明，我们的方法显著改善了多轮环境中的策略学习，减少了无效动作与重复循环。
cs.AI / 19 / 2610.25996
Procedural Memory for Language Agents: A Neurosymbolic Perspective on Explicit Goal Structure
语言智能体的程序性记忆：关于显式目标结构的神经符号视角
large language model
大语言模型相关
Abstract
Large language models remain unreliable on long-horizon tasks requiring procedural reasoning. We argue that a key cause is the absence of explicit procedural memory: accessible, modifiable knowledge of how tasks are accomplished, distinct from semantic and episodic memory. Goal-Oriented Graphs offer such a representation, encoding goals and prerequisite dependencies as inspectable symbolic structure. We examine their construction, the challenges they pose, and principles for integrating them with neural reasoning.
Chinese Translation
大型语言模型在需要程序性推理的长时程任务上仍然不可靠。我们认为，一个关键原因是显式程序性记忆的缺失：即关于任务如何完成的、可访问且可修改的知识，它区别于语义记忆和情景记忆。目标导向图提供了这样一种表示，将目标和前置依赖关系编码为可检查的符号结构。我们考察其构建、它们所带来的挑战，以及将它们与神经推理相整合的原则。
cs.AI / 20 / 2610.26002
TimesJev: Thinking Fast on Time Series with System 1 Sufficiency
TimesJev：借助 System 1 充分性在时间序列上快速思考
large language model
大语言模型相关
Abstract
Large language models (LLMs) have recently shown strong potential for time series reasoning, while models that reason over time series are usually with long chains of thought. However, this process often consumes thousands of tokens, directly leading to substantial latency during inference. In contrast, System 1 models such as Jev make decisions instantly and can naturally be an alternative, but raw time series are not compatible with the embedding space of these models, thus hindering the reasoning performance, especially when input series contain thousands of numeric points. In this paper, we make a very early attempt to propose a time series System 1 model, namely TimesJev. Specifically, we first formalize System 1 sufficiency in time series reasoning, which motivates a compressed representation for time series with key frequency patterns preserved. To construct this representation, we derive a dimension reduction objective that combines (i) a sufficiency divergence that measures the core information lost during compression, and (ii) a realizability divergence that quantifies the incompatibility of the System 1 model to the optimal prediction given this representation. We optimize this objective by learning time series compression with a gating mechanism for sufficiency, and then aligning the selected representation with the embedding space for realizability. Finally, experiments on 13 datasets demonstrate that TimesJev outperforms reasoning models by over 62% with 62.9% less inference time than their fine-tuned versions on average.
Chinese Translation
大型语言模型（LLMs）最近在时间序列推理方面展现出强大潜力，而对时间序列进行推理的模型通常具有长思维链。然而，这一过程往往消耗数千个 token，直接导致推理期间出现大量延迟。相比之下，诸如 Jev 的 System 1 模型能够即时做出决策，自然可以成为一种替代方案，但原始时间序列与这些模型的嵌入空间不兼容，从而妨碍了推理性能，尤其当输入序列包含数千个数值点时。本文做出一个非常早期的尝试，提出一种时间序列 System 1 模型，即 TimesJev。具体而言，我们首先形式化时间序列推理中的 System 1 充分性，这促使我们采用一种保留关键频率模式的时间序列压缩表示。为了构建这种表示，我们推导出一个降维目标，它结合了（i）衡量压缩过程中丢失的核心信息的充分性散度，以及（ii）量化 System 1 模型在给定该表示时与最优预测之间不兼容性的可实现性散度。我们通过使用用于充分性的门控机制学习时间序列压缩，然后将所选表示与嵌入空间对齐以实现可实现性，来优化该目标。最后，在 13 个数据集上的实验表明，TimesJev 比推理模型高出超过 62%，并且平均比这些推理模型的微调版本减少 62.9% 的推理时间。
cs.AI / 21 / 2610.26185
MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching
MoveBench：一个用于评估运动与健康指导中大语言模型的基于评分细则的基准
large language model
大语言模型相关
Abstract
Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks. Existing general-purpose benchmarks (e.g., MMLU and HELM) primarily assess broad knowledge and reasoning capabilities, but fail to capture critical domain requirements such as exercise safety, physiological appropriateness, personalization, and actionable coaching quality. This gap is particularly concerning in fitness and health applications, where incorrect or unsafe recommendations can lead to physical harm. To address this limitation, we introduce MoveBench, the first comprehensive benchmarking framework specifically designed for evaluating LLMs in sports and health coaching. The proposed benchmark integrates heterogeneous task formats, including multiple-choice questions (MCQs) and open-ended question answering (OpenQA), grounded in real-world user scenarios and paired with structured rubrics emphasizing safety compliance, scientific validity, and training logic. Through large-scale evaluation of 15 diverse LLMs on 809 questions with 5,012 rubric-level criteria, our results reveal substantial performance variability, particularly in safety awareness, structured program design, and personalization capability, highlighting a critical gap between knowing and doing. Our work establishes a foundational evaluation standard for sports and health LLM applications and provides a reproducible framework for advancing safe, reliable, and context-aware AI-driven coaching systems. Open resources are available at https://github.com/SatoSakula/MoveBench.
Chinese Translation
尽管大语言模型（LLMs）取得了快速进展，但由于缺乏领域专用的评估基准，其在运动与健康领域的部署仍然具有高风险。现有的通用基准（如 MMLU 和 HELM）主要评估广泛的知识与推理能力，但未能捕捉诸如运动安全性、生理适当性、个性化以及可执行的指导质量等关键领域需求。这一差距在健身与健康应用中尤为令人担忧，因为不正确或不安全的建议可能导致身体伤害。为解决这一局限，我们提出 MoveBench，这是首个专门为评估运动与健康指导中的 LLMs 而设计的综合性基准框架。所提出的基准整合了异构的任务形式，包括多项选择题（MCQs）和开放式问答（OpenQA），其以真实世界用户场景为基础，并配以强调安全合规性、科学有效性和训练逻辑的结构化评分细则。通过对 15 个多样化 LLMs 在 809 个问题、5,012 条评分细则级标准上进行大规模评估，我们的结果揭示了显著的性能差异，尤其是在安全意识、结构化方案设计和个性化能力方面，凸显了“知”与“行”之间的关键差距。我们的工作为运动与健康 LLM 应用建立了基础性评估标准，并为推进安全、可靠且具有情境感知能力的 AI 驱动指导系统提供了一个可复现的框架。开放资源可在 https://github.com/SatoSakula/MoveBench 获取。
cs.AI / 22 / 2610.26226
Towards More Granular Reasoning: Mapping the DSA Transparency Database Taxonomy to Platform Terms and Conditions Using Large Language Models
迈向更细粒度的推理：使用大语言模型将 DSA 透明度数据库分类体系映射至平台条款与条件
large language model
大语言模型相关
Abstract
Article 24(5) of the Digital Services Act (DSA) requires providers of online platforms to submit every content moderation decision they make, with a corresponding Statement of Reasons (SoR), to the DSA Transparency Database (DSA-TDB). Such decisions are classified against seventeen high-level attribute categories that categorize the primary reason for the moderation decision. Because over 99.8% of statements cite a violation of the provider's own terms and conditions rather than a legal provision, the database is overwhelmingly a record of contractual enforcement using catch-all reporting; however, whether its categories actually describe those contractual rules well has not been examined. In this paper, we map the terms and conditions (T&C) of 31 designated online services against the DSA-TDB's high-level category taxonomy using a few-shot GPT-4.1 classifier over 447 policy documents. The analysis concerns how T&C rules are represented in the taxonomy, not how they are enforced. Two findings emerge. First, high-level category applicability varies sharply by service type: unsafe products account for 17.5% of e-commerce assignments and 0.2% of adult-platform assignments, and negative effects on civic discourse reaches 9.9% within the Meta family but is absent from adult platforms, Booking.com, and Wikipedia. Second, the categories have notable overlap among sub-category specifications, which makes assigning a single high-level category to a decision difficult. We argue that catch-all reporting is plausibly in part a consequence of the DSA-TDB's schema and attribute design, and recommend a schema that is differentiated by service type with more mandatory attributes derived inductively from platform governance language.
Chinese Translation
《数字服务法》（DSA）第24条第5款要求在线平台提供者将其作出的每一项内容审核决定，连同相应的理由说明（Statement of Reasons, SoR），提交至 DSA 透明度数据库（DSA-TDB）。此类决定依据十七个高层级属性类别进行分类，这些类别对审核决定的主要原因进行归类。因为超过99.8%的声明援引的是对提供者自身条款与条件的违反，而非法律规定，所以该数据库绝大多数是使用兜底式报告对合同执行情况的记录；然而，其类别是否真正很好地描述了这些合同规则，尚未得到考察。在本文中，我们使用一个在447份政策文档上运行的少样本 GPT-4.1 分类器，将31个指定在线服务的条款与条件（T&C）映射到 DSA-TDB 的高层级类别分类体系。该分析关注的是 T&C 规则在分类体系中如何被表示，而不是它们如何被执行。得到两项发现。首先，高层级类别的适用性因服务类型而差异显著：不安全产品占电子商务分配的17.5%，占成人平台分配的0.2%；对公民话语的负面影响在 Meta 系列中达到9.9%，但在成人平台、Booking.com 和 Wikipedia 中则不存在。其次，这些类别在子类别规范之间存在显著重叠，这使得为一项决定分配单一高层级类别变得困难。我们认为，兜底式报告在一定程度上很可能是 DSA-TDB 的模式与属性设计所导致的结果，并建议采用一种按服务类型区分、包含更多从平台治理语言中归纳得出的强制性属性的模式。
cs.AI / 23 / 2610.26362
GAIDA: Generative AI for Human-in-the-Loop Document Annotation
GAIDA：用于人在回路文档标注的生成式人工智能
large language model
大语言模型相关
Abstract
Document annotation is a fundamental yet labor-intensive step in building document information extraction systems. While Optical Character Recognition (OCR) and Large Language Models (LLMs) can automate parts of this process, we argue that annotation systems should be judged not by how completely they remove human effort, but by how effectively they redirect it toward review and correction. We propose a modular human-in-the-loop architecture that integrates OCR-based extraction, LLM-based semantic pre-annotation, and human validation in a single editable workflow, and present \textbf{GAIDA}, a research prototype serving as its reference implementation. GAIDA supports interchangeable OCR engines, LLM backends, and prompt configurations, and enables stage-specific measurement of annotation effort. In an initial single-annotator, stage-specific study on the FUNSD scanned-form and WildReceipt retail-receipt benchmarks, OCR assistance reduced region-creation and transcription time by up to \textbf{80.9\%}, and reviewing LLM pre-annotations reduced semantic-labeling time by up to \textbf{73.7\%}, despite imperfect pre-annotations of 78.86\% and 86.98\% Micro F1 on FUNSD and WildReceipt, respectively. Finally, we reuse the same implementation, adapting only OCR and prompt configuration, to build and openly release \textbf{AraReceipt}, a manually annotated dataset of 100 Arabic receipt images labeled with 25 key-information classes. The results indicate that imperfect automated outputs provide practical value when presented as editable suggestions under human control.
Chinese Translation
文档标注是构建文档信息抽取系统中的一个基础但劳动密集的步骤。尽管光学字符识别（OCR）和大语言模型（LLMs）可以自动化这一过程的部分环节，我们认为，评判标注系统不应依据其在多大程度上完全消除人工劳动，而应依据其如何有效地将人工劳动重新引导至审核与修正。我们提出一种模块化的人在回路架构，将基于OCR的抽取、基于LLM的语义预标注和人工验证集成到单一可编辑工作流中，并介绍\textbf{GAIDA}，一个作为其参考实现的研究原型。GAIDA支持可互换的OCR引擎、LLM后端和提示配置，并能够对标注工作量进行分阶段测量。在一项针对FUNSD扫描表单和WildReceipt零售收据基准的初始单标注者、分阶段研究中，OCR辅助将区域创建和转录时间最多减少了\textbf{80.9\%}，而审核LLM预标注将语义标注时间最多减少了\textbf{73.7\%}，尽管预标注并不完美，在FUNSD和WildReceipt上的Micro F1分别为78.86\%和86.98\%。最后，我们复用同一实现，仅调整OCR和提示配置，构建并公开发布\textbf{AraReceipt}，这是一个包含100张阿拉伯语收据图像、标注有25个关键信息类别的人工标注数据集。结果表明，不完美的自动化输出在以人工控制下的可编辑建议形式呈现时具有实用价值。
cs.AI / 24 / 2610.26369
Autonomous intelligence for mining productivity: a multi-agent AI approach to operational value creation
面向采矿生产率的自主智能：一种实现运营价值创造的多智能体人工智能方法
large language model
大语言模型相关
Abstract
Integrated Remote Operation Centres (IROCs) are emerging as the architectural frontier of Industry 4.0 for capital-intensive sectors such as mining, oil and gas, and energy. Yet two intertwined gaps constrain their evolution towards full autonomy. First, no integrated agentic architecture has been proposed for mining IROCs that orchestrates Supervisory Control and Data Acquisition systems, digital twins, predictive analytics and human-machine interfaces under safety-critical constraints. Second, no evaluation framework jointly addresses the real-time, Health-Safety-Environment, regulatory-auditability and cost-per-decision requirements that govern industrial agentic deployments. This paper makes four contributions. It synthesises industry needs and quantified gains documented in analogous industries that motivate an agentic approach to mining IROCs. It introduces EI2ROC, a three-tier multi-agent architecture combining a Large Language Model orchestrator with specialised monitoring and production-capacityprediction agents communicating through a Model Context Protocol-style tool interface. It provides a critical review of agentic evaluation along six orthogonal dimensions and identifies the industrial blind spot of current benchmarks. Finally, it proposes a multi-dimensional benchmarking framework for industrial agentic AI comprising scenario library, harness, metrics and humancalibration protocol. The paper states explicitly which components are implemented in the current prototype, which belong to the proposed reference design and which remain future work, and reports preliminary cost and latency observations from an operating site. Together, EI2ROC and the benchmarking framework constitute a methodological blueprint transferable across CAPEX intensive industries.
Chinese Translation
综合远程运营中心（IROC）正逐渐成为采矿、石油与天然气以及能源等资本密集型行业在工业4.0中的架构前沿。然而，有两个相互交织的缺口制约着其向完全自主化演进。第一，尚未有针对采矿IROC提出集成式智能体架构，能够在安全关键约束下编排监控与数据采集系统、数字孪生、预测分析和人机界面。第二，尚无评估框架能够同时应对支配工业智能体部署的实时性、健康-安全-环境、监管可审计性以及单决策成本等要求。本文作出四项贡献。它综合了同类行业中记录的行业需求与量化收益，这些需求与收益为面向采矿IROC的智能体方法提供了动机。它引入了EI2ROC，这是一种三层多智能体架构，将大语言模型编排器与专门的监测智能体和产能预测智能体相结合，这些智能体通过Model Context Protocol风格的工具接口进行通信。它沿六个正交维度对智能体评估进行了批判性综述，并指出现有基准的工业盲点。最后，它提出了一个面向工业智能体人工智能的多维基准测试框架，包括场景库、测试装置、指标和人类校准协议。本文明确说明了哪些组件已在当前原型中实现、哪些属于所提出的参考设计、哪些仍属未来工作，并报告了来自一个运行站点的初步成本与延迟观察结果。EI2ROC与基准测试框架共同构成了一套可迁移至CAPEX密集型行业的方法论蓝图。
cs.AI / 25 / 2610.26418
Local data augmentation for imbalanced binary classification
用于不平衡二分类的局部数据增强
diffusion
扩散模型相关
Abstract
Facing imbalanced binary classification, we aim at maximizing the recall while limiting the proportion of false detections in the minority class. For doing that, the present work studies the effect of several rebalancing strategies, from the uniform subsampling of the observations from the minority class, to a more involved local data augmentation scheme. To be more specific, we establish that only rebalancing class proportions cannot modify the classification performance, whereas local data augmentation is proved to have an impact on the final results. The local augmentation scheme is introduced by means of a two-step procedure. The first step randomly subsamples anchor points among the training observations. These anchor points are used in the second step, for drawing new augmentation data following an ''augmentation kernel'' to be specified. This enables us to decouple the impact of the anchor points subsampling from the one of drawing augmentation data accounting (or not) for the local geometry of the support. The experiments on synthetic data illustrate the practical performance of several strategies for subsampling anchor points combined with either a Gaussian or a diffusion-based augmentation kernel. Compared to SMOTE-like data augmentation procedures, a better performance is achieved when subsampling in the close neighborhood to the classification boundary combined with a Gaussian augmentation kernel. Furthermore in specific settings where the geometric structure of the support matters, the experimental results show that using a diffusion-based augmentation kernel outperforms the Gaussian one.
Chinese Translation
面对不平衡二分类问题，我们旨在最大化召回率，同时限制少数类中误检测的比例。为此，本文研究了几种重平衡策略的效果，从对少数类观测进行均匀子采样，到一种更复杂的局部数据增强方案。更具体地说，我们确立了仅重新平衡类别比例不能改变分类性能，而局部数据增强被证明会对最终结果产生影响。该局部增强方案通过一个两步过程引入。第一步在训练观测中随机子采样锚点。这些锚点用于第二步，以按照一个待指定的“增强核”抽取新的增强数据。这使我们能够将锚点子采样的影响与抽取增强数据时考虑（或不考虑）支撑集局部几何结构的影响解耦。在合成数据上的实验说明了若干锚点子采样策略与高斯增强核或基于扩散的增强核相结合时的实际性能。与类似 SMOTE 的数据增强过程相比，当在分类边界附近邻域内进行子采样并与高斯增强核结合时，取得了更好的性能。此外，在支撑集的几何结构起作用的特定设置中，实验结果表明使用基于扩散的增强核优于使用高斯增强核。
cs.AI / 26 / 2610.26511
RouterWorld: Scaling Conversational AI with Latency-Aware LLM Routing
RouterWorld：以延迟感知的LLM路由扩展对话式AI
large language model
大语言模型相关
Abstract
Conversational AI systems built on Large Language Models (LLMs) require routing mechanisms that balance answer quality and response latency at scale. However, existing LLM routing benchmarks assume full observability of all candidate model outcomes, under-emphasize latency as a first-class metric, and fragment evaluation across heterogeneous protocols misaligned with real-world settings. We introduce RouterWorld, a comprehensive benchmark for LLM routing in conversational GenAI. RouterWorld unifies 371k single-turn responses and 76k multi-turn agentic trajectories, adopts natural user interactions with a standardized LLM-as-a-judge protocol, and decouples efficiency from transient pricing via hardware-grounded latency annotations. Crucially, RouterWorld formalizes routing under partial observability through two setups: Cold-Start, which restricts learning to sparse, static logs, and Voyager, which models continuous data streams where routing decisions shape future training distributions. Evaluating 13 routing baselines, we find that most routers over-emphasize accuracy while ignoring latency under partial observability. Code and benchmark data are publicly available.
Chinese Translation
基于大语言模型（LLMs）构建的对话式AI系统需要能够在规模化条件下平衡答案质量与响应延迟的路由机制。然而，现有的LLM路由基准假设所有候选模型结果完全可观测，未充分重视延迟这一首要指标，并且将评估碎片化于与真实世界设置不一致的异构协议之中。我们提出RouterWorld，一个用于对话式生成式AI中LLM路由的综合基准。RouterWorld统一了371k条单轮响应和76k条多轮智能体轨迹，采用自然用户交互与标准化的LLM-as-a-judge协议，并通过基于硬件的延迟标注将效率与瞬时定价解耦。至关重要的是，RouterWorld通过两种设置形式化了部分可观测性下的路由：Cold-Start将学习限制在稀疏、静态日志上，Voyager则对连续数据流建模，其中路由决策塑造未来的训练分布。在评估13个路由基线后，我们发现大多数路由器过度强调准确性，而忽视部分可观测性下的延迟。代码和基准数据公开可用。
cs.AI / 27 / 2610.26524
Beyond Behavioral Accuracy: Evaluating Individual-Level Fidelity in LLM Investor Simulation
超越行为准确性：评估 LLM 投资者模拟中的个体层面保真度
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used as user simulators, but can they faithfully reproduce an individual's behavior and decision logic over time? We investigate this question in a longitudinal paper-trading study with 80 participants. Using rolling next-day predictions, we evaluate trading activity, buy-sell structure, asset selection, and portfolio outcomes, complemented by controlled input ablations and participant-rated explanation alignment. No evaluated LLM reliably outperforms a simple recent-activity persistence baseline for predicting trade occurrence, while buy-sell structure and asset selection remain difficult to reproduce. Similar activity-level performance can also yield substantially different portfolio trajectories. Ablations show that activity predictions depend primarily on recent trades, whereas asset choices are sensitive to multiple evidence sources. Participants rate explanations as more aligned with their recurring strategies than with their decision-specific rationales. These findings reveal a gap between coarse behavioral agreement and individual-level fidelity, motivating separate evaluation of behavioral accuracy, explanation alignment, and evidence dependence in personalized user simulation.
Chinese Translation
大型语言模型（LLMs）正越来越多地被用作用户模拟器，但它们能否随时间忠实地再现个体的行为和决策逻辑？我们通过一项包含 80 名参与者的纵向纸面交易研究来探讨这一问题。使用滚动式的次日预测，我们评估交易活动、买卖结构、资产选择和投资组合结果，并辅以受控的输入消融实验和参与者评定的解释一致性。在预测交易发生方面，没有哪个被评估的 LLM 能够可靠地优于一个简单的近期活动持续性基线，而买卖结构和资产选择仍然难以复现。相似的活动层面表现也可能产生显著不同的投资组合轨迹。消融实验表明，活动预测主要依赖于近期交易，而资产选择对多种证据来源敏感。参与者认为解释与其反复出现的策略的一致性高于与其特定决策理由的一致性。这些发现揭示了粗粒度行为一致性与个体层面保真度之间的差距，这促使在个性化用户模拟中分别评估行为准确性、解释一致性和证据依赖性。
cs.AI / 28 / 2610.26542
Breaking Diffusion: Parallel Reasoning with Draft Refinement
打破扩散：结合草稿精炼的并行推理
diffusion
扩散模型相关
Abstract
Drafting an answer allows a reasoning model to reconsider each decision in the context of the others. Yet repeated diffusion denoising does not ensure that individually plausible token updates form a better answer together. We introduce Draft Refinement Language Models (DRLMs), which use a single denoiser to propose parallel revisions and evaluate a candidate draft. We train the denoiser to reconstruct every token from uniformly corrupted drafts, teaching both preservation and repair. Its predictions specify revisions directly, without a diffusion reverse transition. Evaluation requires a different use of these predictions: confidence in an existing token reflects both contextual support and evidence from observing the token itself. Using the known corruption process, we correct for this self-token evidence to estimate how well each token is supported by the prompt and the remaining draft. Aggregating these estimates yields a cavity energy computed in one forward pass, without an additional scoring model. Each round proposes a revision, reads the completed candidate in its revised context, and retains it only if its energy is lower. Across Countdown, Sudoku-Extreme, and Maze-Hard, our method outperforms autoregressive and discrete diffusion baselines. With 6M parameters, DRLM reaches 44.4% on Countdown-5 and 93.9% on Sudoku-Extreme. The results establish complete-draft parallel refinement as an effective alternative to reverse diffusion for complex reasoning.
Chinese Translation
起草一个答案使推理模型能够在其他决策的背景下重新考虑每个决策。然而，重复的扩散去噪并不能确保单独看似合理的 token 更新能够共同构成更好的答案。我们引入草稿精炼语言模型（DRLM），它使用单个去噪器来提出并行修订并评估候选草稿。我们训练去噪器从均匀损坏的草稿中重建每个 token，同时教会其保留和修复。其预测直接指定修订，而无需扩散逆向转移。评估需要以不同方式使用这些预测：对现有 token 的置信度既反映上下文支持，也反映来自观察该 token 本身的证据。利用已知的损坏过程，我们对这种自 token 证据进行校正，以估计每个 token 受到提示和其余草稿支持的程度。聚合这些估计可得到在一次前向传播中计算出的空腔能量，而无需额外的评分模型。每一轮提出一个修订，在其修订后的上下文中读取已完成的候选，并且仅当其能量更低时才保留它。在 Countdown、Sudoku-Extreme 和 Maze-Hard 上，我们的方法优于自回归和离散扩散基线。在拥有 6M 参数的情况下，DRLM 在 Countdown-5 上达到 44.4%，在 Sudoku-Extreme 上达到 93.9%。这些结果确立了完整草稿并行精炼作为复杂推理中逆向扩散的有效替代方案。
cs.AR / 29 / 2610.26131
A Hybrid GPU and Processing-Near-Memory Architecture with Shared DRAM for Efficient Prefill-Decode Disaggregation
一种用于高效预填充-解码解耦的、具有共享 DRAM 的 GPU 与近存处理混合架构
large language model
大语言模型相关
Abstract
The growing adoption of large language models (LLMs) has significantly increased the compute and memory demands of inference serving. A key challenge arises from the mismatch between compute-bound prefill and memory-bound decoding: when executed together, the two phases interfere, while separating them via prefill-decode (PD) disaggregation leads to substantial resource underutilization. Specifically, prefill leaves memory bandwidth underutilized, whereas decoding underutilizes compute resources. Prior approaches, such as GPU Streaming Multiprocessor (SM) partitioning, attempt to improve utilization by spatially dividing GPU resources. However, these approaches remain fundamentally limited, as memory-bound decoding still requires a large number of SMs to saturate bandwidth, reducing the compute resources available for prefill and resulting in an inherent trade-off. To address this limitation, we propose DuoPD, a hybrid PD-disaggregation architecture that integrates lightweight Processing-Near-Memory (PNM) units into the HBM logic die. Unlike prior PIM/PNM designs that rely on costly internal bandwidth scaling and offload only partial decoding operations, DuoPD enables full decoding execution on PNM by leveraging bandwidth sharing with the GPU over a shared HBM interface. This allows compute-bound prefill and memory-bound decoding to execute concurrently with minimal interference, improving both compute and memory utilization. To realize this design, we develop a compact PNM microarchitecture, a TSV-level bandwidth sharing mechanism, and a hierarchical data mapping scheme tailored for hybrid GPU-PNM execution. Evaluations on server-scale LLMs demonstrate that DuoPD achieves up to 2.23x higher end-to-end throughput than conventional PD-disaggregation.
Chinese Translation
大语言模型（LLMs）的日益广泛采用显著增加了推理服务的计算和内存需求。一个关键挑战来自计算受限的预填充与内存受限的解码之间的不匹配：当两者一起执行时，这两个阶段会相互干扰，而通过预填充-解码（PD）解耦将它们分离则会导致显著的资源利用不足。具体而言，预填充使内存带宽未被充分利用，而解码则未充分利用计算资源。先前的方法，例如 GPU 流式多处理器（SM）划分，试图通过空间划分 GPU 资源来提高利用率。然而，这些方法仍然受到根本性限制，因为内存受限的解码仍需要大量 SM 来饱和带宽，从而减少了可用于预填充的计算资源，并导致固有的权衡。为解决这一限制，我们提出 DuoPD，一种将轻量级近存处理（PNM）单元集成到 HBM 逻辑裸片中的混合 PD 解耦架构。与依赖昂贵的内部带宽扩展且仅卸载部分解码操作的先前 PIM/PNM 设计不同，DuoPD 通过利用在共享 HBM 接口上与 GPU 的带宽共享，使完整解码执行能够在 PNM 上运行。这允许计算受限的预填充和内存受限的解码以最小干扰并发执行，从而提高计算和内存利用率。为实现这一设计，我们开发了紧凑的 PNM 微架构、TSV 级带宽共享机制，以及为 GPU-PNM 混合执行量身定制的分层数据映射方案。在服务器规模 LLMs 上的评估表明，DuoPD 相比传统 PD 解耦实现了最高达 2.23 倍的端到端吞吐量提升。
cs.CL / 30 / 2610.25463
DiaLong: Benchmarking Reminiscence in Long-Context Dialogues
DiaLong：在长上下文对话中对回忆能力进行基准测试
large language model
大语言模型相关
Abstract
How well can transformer-based Large Language Models (LLMs) retain and recall memories in long contexts of dialogue? LLMs are known to forget long-range factual information in their context, yet the field lacks a comprehensive benchmark to quantitatively assess these shortcomings. We present DiaLong, a long-context dialogue dataset and accompanying memory task that standardizes evaluation in this space. It consists of 100 unique conversations created by humans and then augmented with LLMs, extending beyond lengths found in existing datasets. Our selected memory task is fact recall: classify 40 candidate facts about a dialogue as true or false. This task directly probes the model's ability to actively retain and retrieve information, serving as a critical measure of memory within extended conversational contexts. In testing both the latest LLMs and proposed memory enhancements on DiaLong, we demonstrate that more work needs to be done to enable LLMs with better memory performance. By enhancing memory capabilities, LLMs can improve in maintaining coherence, relevance, and human-like interaction over extended conversation. Improving these capabilities will help LLMs fulfill diverse user needs and expectations in various conversational scenarios.
Chinese Translation
基于 Transformer 的大型语言模型（LLMs）在长对话上下文中能够多好地保留和回忆记忆？已知 LLMs 会在其上下文中遗忘长程事实信息，然而该领域仍缺乏一个全面的基准来定量评估这些不足。我们提出了 DiaLong，这是一个长上下文对话数据集以及配套的记忆任务，用于标准化这一领域的评估。它由 100 段由人类创建、随后用 LLMs 增强的独特对话组成，其长度超出了现有数据集中所见的长度。我们选定的记忆任务是事实回忆：将关于一段对话的 40 个候选事实分类为真或假。该任务直接探查模型主动保留和检索信息的能力，可作为扩展对话上下文中记忆的一项关键衡量指标。在 DiaLong 上测试最新的 LLMs 以及提出的记忆增强方法时，我们证明还需要开展更多工作才能使 LLMs 具备更好的记忆性能。通过增强记忆能力，LLMs 可以在扩展对话中更好地保持连贯性、相关性以及类人交互。提升这些能力将有助于 LLMs 在各种对话场景中满足多样化的用户需求与期望。
cs.CL / 31 / 2610.25596
Cognitive Map in Large Language Models
大语言模型中的认知地图
large language model
大语言模型相关
Abstract
It has long been proposed that animals construct cognitive maps to support flexible behavior in novel environments. We tested whether large language models form similar representations using relational reasoning tasks and causal "mosaic experiments" that manipulate internal activations in open-weight language models. Models beyond the ten-billion-parameter scale successfully inferred unstated relationships between entities and generalized to synthetic entity names. Across model and task families, models generated abstract representations of relational structure and combined them additively with representations of entity identity, binding entities to roles within that structure. Models stored observed relationships as pairwise links that they composed retrospectively when queried, rather than prospectively positioning each entity in a global coordinate system. Binding was mediated by a sparse, distributed set of key--value pairs; by perturbing these pairs, we implanted "false memories" that altered model behavior in predictable ways. Overall, our findings reveal an internal organization analogous to a cognitive map that explains how language models reason with limited information.
Chinese Translation
长期以来，人们一直提出动物会构建认知地图，以支持其在新环境中的灵活行为。我们利用关系推理任务以及因果性的“马赛克实验”（即对开放权重语言模型的内部激活进行干预），检验了大语言模型是否会形成类似的表征。参数量超过百亿规模的模型成功推断了实体之间未被明示的关系，并泛化到合成的实体名称。在多个模型系列与任务系列中，模型生成了关系结构的抽象表征，并将其与实体身份的表征以加性方式结合，从而将实体绑定到该结构中的角色上。模型将观察到的关系存储为成对链接，并在被查询时以回溯方式组合这些链接，而不是前瞻性地将每个实体定位在一个全局坐标系中。这种绑定由一组稀疏、分布式的键--值对介导；通过扰动这些键--值对，我们植入了“虚假记忆”，并以可预测的方式改变了模型的行为。总体而言，我们的发现揭示了一种类似于认知地图的内部组织，它解释了语言模型如何利用有限的信息进行推理。
cs.CL / 32 / 2610.25626
MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models
MapSteal：针对大语言模型的基于地图的自主知识窃取攻击
large language model
大语言模型相关
Abstract
Knowledge acquisition from black-box large language models (LLMs) offers a practical approach to adapting smaller student models. However, many existing methods rely on predefined queries or student feedback to guide acquisition, leaving autonomous expansion of knowledge targets an important challenge. We propose \textbf{MapSteal}, a student-independent framework for structured recursive knowledge acquisition. Starting from a small set of initial roots, MapSteal represents acquisition targets using structured states and recursively expands them through teacher-proposed directions. Question realization, proposal admission, and deduplication constrain the acquisition process, while root-aware scheduling organizes exploration under a fixed QA budget. The resulting question--answer pairs are used to adapt student models without requiring student feedback during acquisition. Experiments with Qwen3.5 students at 2B, 4B, and 9B scales show competitive downstream performance across six medical and cybersecurity benchmarks. On the 4B medical evaluation, Full MapSteal improves average accuracy by 1.17 percentage points over a simplified recursive acquisition variant. Cross-domain evaluations show that gains depend on the domain and evaluation panel. These results demonstrate the feasibility of structured, student-independent recursive acquisition while highlighting the distinction between generating acquisition targets and obtaining useful training supervision. Our code is available at \href{https://github.com/LabRAI/Mapsteal} {https://github.com/LabRAI/Mapsteal}.
Chinese Translation
从黑盒大语言模型（LLMs）中获取知识为适配更小的学生模型提供了一种实用方法。然而，许多现有方法依赖预定义查询或学生反馈来指导获取，使得知识目标的自主扩展成为一个重要挑战。我们提出 \textbf{MapSteal}，一个用于结构化递归知识获取的独立于学生的框架。从一小组初始根节点出发，MapSteal 使用结构化状态表示获取目标，并通过教师提出的方向递归地扩展这些目标。问题实现、提议准入和去重约束获取过程，而根感知调度在固定的问答（QA）预算下组织探索。由此得到的问答对用于适配学生模型，而无需在获取过程中要求学生反馈。使用 2B、4B 和 9B 规模的 Qwen3.5 学生模型进行的实验表明，在六个医学和网络安全基准上具有有竞争力的下游性能。在 4B 医学评估中，Full MapSteal 相比简化的递归获取变体将平均准确率提高了 1.17 个百分点。跨领域评估表明，增益取决于领域和评估面板。这些结果证明了结构化、独立于学生的递归获取的可行性，同时强调了生成获取目标与获得有用训练监督之间的区别。我们的代码可在 \href{https://github.com/LabRAI/Mapsteal} {https://github.com/LabRAI/Mapsteal} 获取。
cs.CL / 33 / 2610.25724
Inside the Scorer: Mechanistic Evidence on Rubric Control in LLM-Based Automated Essay Scoring from 25 Base Models
评分器内部：来自25个基础模型的基于LLM的自动作文评分中评分细则控制的机制证据
large language model
大语言模型相关
Abstract
Automated essay scoring (AES) using large language models (LLMs) offers promising use cases for language assessment, yet high-stakes testing requires scores that are both controllable and verifiable, which current LLM-based scorers lack. Given that, we investigate whether LLM-based scorers form distinct internal representations for individual scoring rubrics and whether these representations can be independently controlled. Probing and steering 25 open-weight base LLMs (0.13B-10.31B parameters; 14 families) on human-scored BESTEP English writing tasks, we find that holistic scores are decodable from model activations for the longer tasks (Parts 2 and 3) and only weakly for the shortest (Part 1). Furthermore, scoring rubrics have partly separable directions in representation space for the longer tasks, though not for short responses (Part 1). However, causal steering reveals a critical limitation: applying a matching intervention (e.g., steering "grammar" on a grammar-degraded response) yields only a negligible advantage (about 2% of the effect) over a mismatched intervention (e.g., steering "vocabulary"), and none that is detectable in any of the 10 testable comparisons across reliably measured rubrics. Rather than controlling distinct evaluative dimensions, the rubric directions share a dominant lever whose rubric-specific parts are not selective, which may mirror the halo structure of human ratings. Our contributions are two-fold. First, we show that analytic scoring rubrics designed for human raters do not behave as independent controls in base LLMs, even among the seven models whose scores agree best with human raters ($r \geq 0.6$). Second, we establish a methodology that shows that score agreement and decodability do not imply rubric-level control in LLM-based assessment.
Chinese Translation
基于大语言模型（LLM）的自动作文评分（AES）为语言评估提供了有前景的应用场景，但高风险测试要求分数既可控又可验证，而当前基于LLM的评分器缺乏这一点。鉴于此，我们考察基于LLM的评分器是否会针对各个评分细则形成不同的内部表征，以及这些表征是否可以被独立控制。我们在人工评分的BESTEP英语写作任务上，对25个开放权重的基础LLM（0.13B-10.31B参数；14个模型家族）进行探测和引导，发现对于较长的任务（第2部分和第3部分），整体分数可以从模型激活中解码出来，而对于最短的任务（第1部分）只能微弱解码。此外，对于较长的任务，评分细则在表征空间中具有部分可分离的方向，但对于短回答（第1部分）则并非如此。然而，因果引导揭示了一个关键局限：施加匹配干预（例如，对语法退化的回答引导“语法”）相较于不匹配干预（例如，引导“词汇”）仅带来可忽略的优势（约为效应的2%），并且在跨可靠测量的评分细则的10项可测试比较中的任何一项都检测不到这种优势。评分细则方向并未控制不同的评价维度，而是共享一个主导杠杆，其针对具体评分细则的部分并不具有选择性，这可能反映了人类评分的晕轮结构。我们的贡献有两个方面。第一，我们表明，为人类评分者设计的分析性评分细则在基础LLM中并不表现为独立控制，即使在分数与人类评分者一致性最高的七个模型中也是如此（$r \geq 0.6$）。第二，我们建立了一种方法，表明分数一致性和可解码性并不意味基于LLM的评估中具有评分细则层面的控制。
cs.CL / 34 / 2610.25728
No Matter Who You Are: Addressivity and Sycophancy in Large Language Models
无论你是谁：大型语言模型中的言说对象性与谄媚性
large language model
大语言模型相关
Abstract
Sycophancy, where an LLM becomes overly agreeable, deferential, or flattering toward the user, has been identified as a major failure mode. While post-training has been implicated as a cause, how it internally gives rise to sycophantic behavior remains insufficiently understood. Post-training transforms the model's output from text completion into utterances directed toward an addressee, thereby giving the output \textit{addressivity}. We hypothesize that the accompanying pressure to follow the addressee's instructions and align with their preferences links this addressivity to sycophancy. To test this, we use Sparse Autoencoders to extract addressivity directions from pre-trained and post-trained models and examine their relationship with sycophancy. Steering models with addressivity directions can amplify or suppress sycophancy, and this effect is observed only in post-trained models. Furthermore, comparing the internal mechanisms of both model types confirms that the association between addressivity and sycophancy is strengthened through post-training. Finally, through a controlled post-training experiment, we find that models are sycophantic not toward a specific addressee but toward any recognized addressee, suggesting that the presence of an addressee, rather than the addressee's identity, is what elicits sycophancy.
Chinese Translation
谄媚性，即 LLM 对用户变得过度顺从、恭顺或奉承，已被认定为一种主要失败模式。尽管后训练被认为是一个原因，但其在内部如何引发谄媚行为仍未被充分理解。后训练将模型的输出从文本补全转变为指向某个言说对象的表述，从而赋予输出以\textit{言说对象性}。我们假设，伴随而来的遵循言说对象指令并与其偏好保持一致的压力，将这种言说对象性与谄媚性联系起来。为检验这一点，我们使用稀疏自编码器从预训练模型和后训练模型中提取言说对象性方向，并考察它们与谄媚性的关系。用言说对象性方向对模型进行引导可以放大或抑制谄媚性，且这种效应仅在经过后训练的模型中观察到。此外，比较两类模型的内部机制证实，言说对象性与谄媚性之间的关联通过后训练得到增强。最后，通过一项受控的后训练实验，我们发现模型并非针对特定的言说对象表现出谄媚，而是针对任何被识别出的言说对象，这表明引发谄媚的是言说对象的存在，而非言说对象的身份。
cs.CL / 35 / 2610.25768
Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-makers
预算约束下的截断推理：评估 System-1 模型作为决策者
large language model
大语言模型相关
Abstract
Large language models (LLMs) often generate more reasoning tokens than needed. Limiting the reasoning budget reduces cost, but forcing an answer from an incomplete trace can hurt accuracy. System-1 decision models select among candidate options without producing a reasoning response. We evaluate Jev as a decision reader of truncated reasoning across seven LLM reasoners and five multiple-choice benchmarks (35 pairs). Jev achieves higher mean accuracy than forced answering at tested budgets from 500 to 16,000 tokens, including the estimated costs of reasoning and answering. At a 1,500-token budget, it improves accuracy by 4.1 pp overall and 8.6 pp on truncated questions, with significant gains in 17 pairs under Holm-corrected bootstrap tests. Offline replay estimates 34.5% of the full-run cost, with a 1.13 pp accuracy loss. A policy selected on training folds saves 71.3% of the estimated full-run cost on held-out folds, with a 0.58 pp accuracy loss. We also find that Jev shows strong question-only accuracy, exceeding 500-token forced answering on average (68.0% vs. 65.5%), although this advantage varies by benchmark. These findings suggest that System-1 decision models can complement LLM reasoning and reduce the need for costly reasoning completion.
Chinese Translation
大型语言模型（LLMs）经常生成比所需更多的推理 token。限制推理预算可降低成本，但从完整推理轨迹不完整时强制给出答案可能损害准确率。System-1 决策模型在候选选项中进行选择，而不产生推理响应。我们在七个 LLM 推理器和五个多项选择基准（35 对）上，将 Jev 作为截断推理的决策读取器进行评估。在从 500 到 16,000 个 token 的测试预算下，Jev 的平均准确率高于强制作答，且这些预算包含推理和作答的估计成本。在 1,500 个 token 的预算下，它总体将准确率提高 4.1 个百分点，在截断问题上提高 8.6 个百分点，并在 Holm 校正的 bootstrap 检验下在 17 对中取得显著提升。离线回放估计占完整运行成本的 34.5%，并伴随 1.13 个百分点的准确率损失。在训练折上选择的策略在留出折上节省了估计完整运行成本的 71.3%，准确率损失为 0.58 个百分点。我们还发现，Jev 表现出很强的仅问题准确率，平均超过 500-token 强制作答（68.0% 对 65.5%），尽管这一优势因基准而异。这些发现表明，System-1 决策模型可以补充 LLM 推理，并减少对昂贵推理完成的需求。
cs.CL / 36 / 2610.25835
Diffusion Language Models Change Speed Across Tasks and Languages
扩散语言模型在不同任务和语言中的速度会变化
diffusion
扩散模型相关
Abstract
Diffusion language models can write several tokens in one forward pass, and they decide during generation how many passes to spend. Two outputs of the same token length can therefore vary a lot in cost. To understand the underlying factors of this cost variation, we analyze two diffusion models across 10 tasks. We find that the number of forward passes per token depends heavily on the task. Text copied or lightly edited from the input is cheap, while open-ended writing is costly. In DiffusionGemma, generating a token for story writing requires more than three times as many forward passes as for grammar correction. Two output properties explain much of this gap: how much of it is new relative to the input, and how rare its words are within the task. The same holds within a single response when thinking is involved. The final answer is cheap when it repeats the model's reasoning trace and costly when it adds something new. In multilingual generation, we find that highly tokenized languages also require more passes per token, adding to the cost they already pay with more tokens. Finally, cost tracks correctness: on the same math problem, wrong answers take more passes per token than correct ones, and weighting candidate answers by decoding speed beats majority voting by up to 4.2 points at no extra cost. Taken together, our results show that diffusion's computational advantage depends on how much of the output is already determined, while its decoding effort offers a practical signal for making inference more efficient and reliable.
Chinese Translation
扩散语言模型可以在一次前向传播中写出若干个 token，并且它们在生成过程中决定要花费多少次前向传播。因此，两个 token 长度相同的输出在成本上可能差异很大。为了理解这种成本变化的潜在因素，我们在 10 个任务上分析了两个扩散模型。我们发现，每个 token 所需的前向传播次数在很大程度上取决于任务。从输入中复制或轻度编辑的文本成本较低，而开放式写作成本较高。在 DiffusionGemma 中，为故事写作生成一个 token 所需的前向传播次数是语法纠错的三倍以上。两个输出属性解释了这一差距的很大一部分：输出相对于输入有多少是新内容，以及其词语在任务内的稀有程度。当涉及思考时，在单个响应内部也同样如此。当最终答案重复模型的推理轨迹时成本较低，而当它添加新内容时成本较高。在多语言生成中，我们发现高度 token 化的语言每个 token 也需要更多前向传播次数，这进一步增加了它们因使用更多 token 而已经付出的成本。最后，成本与正确性相关：在同一个数学问题上，错误答案每个 token 所需的前向传播次数比正确答案更多，而按解码速度对候选答案进行加权，在不增加额外成本的情况下，比多数投票最多高出 4.2 个百分点。综上所述，我们的结果表明，扩散的计算优势取决于输出中有多少部分已经确定，而其解码工作量提供了一个实用信号，可用于使推理更高效、更可靠。
cs.CL / 37 / 2610.25846
Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models
大语言模型中置信信号的测量效度与表征动态
large language model
大语言模型相关
Abstract
Whether a language model "knows what it knows" matters for trustworthy deployment and for deciding when it should abstain. Prior work usually measures this through one channel: verbalized confidence, a linear probe on hidden states, or agreement among sampled answers (semantic entropy), and often treats these as interchangeable. We compare all three in a pre-registered audit of five Qwen2.5 models (0.5B-7B, including a base/instruct pair at 7B) on factual retrieval (PopQA, SimpleQA) and multi-step reasoning (GSM8K, MATH). First, probes trained on prefill states to predict whether sampled answers will disagree become informative earlier for retrieval (about 7% of layer depth on average) than for reasoning (about 28%). The pre-registered five-layer test supported this only marginally; a finer 21-layer follow-up, designed after those results, supports it more clearly. In most retrieval cells, however, a prompt-only TF-IDF classifier matches the probe, so the early signal may reflect surface features of the question. Second, verbalized confidence is unstable: slightly different prompt templates yield scores that barely correlate. Third, Expected Calibration Error is misleading: verbal confidence looks well calibrated because it clusters near the base rate while discriminating poorly, whereas sampling agreement has worse ECE but about twice the resolution. Of five pre-registered hypotheses, two were falsified, one was partly supported, one was not supported, and one could not be tested because the base model's verbal confidence was degenerate.
Chinese Translation
语言模型是否“知道自己知道什么”对于可信部署以及决定它何时应当弃答很重要。先前工作通常通过单一渠道测量这一点：言语化置信度、对隐藏状态的线性探针，或采样答案之间的一致性（语义熵），并且常常将它们视为可互换。我们在一次预注册审计中，在事实检索（PopQA、SimpleQA）和多步推理（GSM8K、MATH）任务上比较这三者，涉及五个Qwen2.5模型（0.5B-7B，包括7B的一个基座/指令配对）。第一，在预填充状态上训练、用于预测采样答案是否会不一致的探针，在检索任务中比在推理任务中更早变得有信息量（平均约在层深度的7%处，而推理任务约为28%）。预注册的五层检验仅边际支持这一点；在这些结果之后设计的更精细的21层后续检验则更清楚地支持了这一点。然而，在大多数检索单元中，仅使用提示的TF-IDF分类器与探针表现相当，因此早期信号可能反映问题的表面特征。第二，言语化置信度不稳定：略微不同的提示模板会产生几乎不相关的分数。第三，期望校准误差（ECE）具有误导性：言语置信度看起来校准良好，因为它聚集在基础率附近，同时区分能力很差；而采样一致性具有更差的ECE，但分辨率约为其两倍。在五个预注册假设中，两个被证伪，一个得到部分支持，一个未得到支持，还有一个无法检验，因为基座模型的言语化置信度是退化的。
cs.CL / 38 / 2610.25914
FisherEdit: Fisher Information Guided Knowledge Editing for Large Language Models
FisherEdit：面向大语言模型的Fisher信息引导知识编辑
large language model
大语言模型相关
Abstract
Sequential knowledge editing in Large Language Models (LLMs) often leads to a degradation of general capabilities due to cumulative parameter interference. In this paper, we revisit prominent projection-based methods like AlphaEdit and reveal a surprising mismatch: their robustness is primarily driven by update regularization rather than the intended null-space projection. Based on this insight, we propose FisherEdit, a theoretically grounded method built on Fisher geometry, which replaces coarse activation-space projections with importance-aware regularization in parameter space. It penalizes changes to sensitive parameters more strongly while retaining flexibility for new facts, and admits an efficient row-wise closed-form solution. Experiments on Llama-3, Qwen2.5, and Olmo-3 across multiple benchmarks demonstrate that FisherEdit achieves state-of-the-art performance while better preserving model stability and downstream task accuracy.
Chinese Translation
大语言模型（LLMs）中的序列知识编辑往往因累积的参数干扰而导致通用能力退化。本文重新审视 AlphaEdit 等主流基于投影的方法，并揭示出一个令人意外的错配：其鲁棒性主要由更新正则化驱动，而非其预期的零空间投影。基于这一洞见，我们提出 FisherEdit，一种建立在 Fisher 几何之上的具有理论依据的方法，它用参数空间中的重要性感知正则化取代粗粒度的激活空间投影。它对敏感参数的改变施加更强惩罚，同时为新事实保留灵活性，并具有高效的行式闭式解。在 Llama-3、Qwen2.5 和 Olmo-3 上跨多个基准的实验表明，FisherEdit 在取得最先进性能的同时，更好地保持了模型稳定性和下游任务准确率。
cs.CL / 39 / 2610.25934
MITRA-lexicon: Lexica of attested translations from Sanskrit into English, Tibetan and Chinese
MITRA-lexicon：梵语译入英语、藏语和汉语的有据可查译文词典
large language model
大语言模型相关
Abstract
The standard Sanskrit lexica are monumental and indispensable, but they cite their attestations as abbreviated references to editions, without links to the digitized texts now widely available and without a translation of the passage, and they naturally do not reflect translations made after them. To address these gaps, we present MITRA-lexicon, a family of corpus-grounded lexica based on span annotations of translation equivalents with Sanskrit as the pivot: Sanskrit-English, derived from the Mitrasamgraha corpus, as well as Sanskrit-Tibetan and Sanskrit-Chinese, derived from the MITRA-parallel corpus of Buddhist literature. In total, the resource comprises 2.18 million headword entries in six directional lexica. Headwords are lemmas and compounds from silver annotations that follow the Digital Corpus of Sanskrit (DCS). Senses are ranked by frequency, and each of the 10.4 million attestations includes its sentence pair, with character offsets of the word and its equivalent. We propose a multi-step pipeline based on grammatical annotation models, fine-tuned large language models, and a teacher-student setup to produce the alignments. We evaluate the lexica against reference dictionaries and manually corrected gold alignments. We release the lexica and the fine-tuned span annotation model, together with a web interface for looking up headwords and their attested senses.
Chinese Translation
标准梵语词典是里程碑式且不可或缺的，但它们将例证以对版本的缩写参考文献形式引用，既不链接到如今广泛可用的数字化文本，也不提供段落的译文，并且它们自然无法反映在其之后产生的翻译。为弥补这些空白，我们提出 MITRA-lexicon，这是一系列以语料库为基础的词典，基于以梵语为枢轴的翻译对应项的跨度标注：梵语—英语词典源自 Mitrasamgraha 语料库，梵语—藏语和梵语—汉语词典则源自佛教文献的 MITRA-parallel 语料库。总体而言，该资源包含六部方向性词典中的 218 万条词目条目。词目是来自遵循 Digital Corpus of Sanskrit（DCS）的银标注中的词元和复合词。义项按频率排序，且 1040 万条例证中的每一条都包含其句对，并带有该词及其对应项的字符偏移。我们提出一个多步骤流水线，基于语法标注模型、微调后的大语言模型以及教师—学生设置来生成这些对齐。我们参照参考词典和人工校正的金标准对齐来评估这些词典。我们发布这些词典和微调后的跨度标注模型，并提供一个用于查询词目及其有据义项的网页界面。
cs.CL / 40 / 2610.25950
AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments
AdaExploreBench：面向演化环境中主动探索与测试时适应的LLM智能体基准测试
large language model
大语言模型相关
Abstract
Large Language Model (LLM) agents are increasingly deployed in unfamiliar and evolving environments, requiring them to proactively explore, understand, and adapt to environmental dynamics. However, existing benchmarks rarely test whether agents can proactively discover how an environment works and adapt as it evolves. To address this gap, we introduce AdaExploreBench, a benchmark for proactive exploration and test-time adaptation in evolving environments. By instantiating environmental dynamics as hidden rules, AdaExploreBench requires agents to actively acquire evidence and infer the rules. Under implicit rule drift, the latent rule changes without announcement, requiring agents to detect from degraded performance that prior knowledge is stale and re-explore. Experiments on eight leading LLMs show that proactive and adaptive exploration remain challenging: GPT-6 Sol achieves only 61.9% calibrated accuracy, while providing sufficient evidence or removing rule drift improves average performance significantly by close to 10%. These results reveal a critical bottleneck in current LLM agents: they can often reason effectively from informative evidence once it is available, but struggle to actively acquire the right evidence and recognize when previously acquired evidence is no longer valid.
Chinese Translation
大语言模型（LLM）智能体正越来越多地被部署在不熟悉且不断演化的环境中，这要求它们主动探索、理解并适应环境动态。然而，现有基准很少检验智能体能否主动发现环境如何运作，并在环境演化时进行适应。为弥补这一空白，我们提出了AdaExploreBench，一个面向演化环境中主动探索与测试时适应的基准。通过将环境动态实例化为隐藏规则，AdaExploreBench要求智能体主动获取证据并推断规则。在隐式规则漂移下，潜在规则会在没有公告的情况下发生变化，要求智能体从性能下降中检测到先验知识已经过时并重新探索。对八个领先LLM的实验表明，主动且自适应的探索仍然具有挑战性：GPT-6 Sol仅达到61.9%的校准准确率，而提供充分证据或消除规则漂移可使平均性能显著提升近10%。这些结果揭示了当前LLM智能体的一个关键瓶颈：一旦有信息丰富的证据可用，它们通常能够有效地进行推理，但难以主动获取正确的证据，并难以识别先前获取的证据何时不再有效。
cs.CL / 41 / 2610.26040
Balance-Aware Pivot Selection for Parallel Diffusion LLM Decoding
面向并行扩散大语言模型解码的平衡感知枢轴选择
diffusion
扩散模型相关
Abstract
Diffusion language models enable parallel token generation, offering a promising alternative to sequential autoregressive decoding. Existing acceleration methods use predictive confidence to enable parallel token commitment while preserving generation quality. However, confidence-based selection prioritizes reliable token predictions without explicitly considering how token positions shape the remaining decoding schedule. Moreover, such ordering does not always reflect inter-token dependencies, suggesting that confidence need not be the sole criterion for token selection. Motivated by this limitation, we propose Balance-Aware Pivot Decoding (BPD), which favors central pivots within contiguous masked intervals and decodes the selected positions in parallel. Balanced pivots yield logarithmic decoding depth in the interval length. Our practical method accounts for model uncertainty by combining confidence-based prefilling with confidence-thresholded center selection. We evaluate the resulting quality-efficiency tradeoff on LLaDA and Dream across five mathematical reasoning, code generation, and instruction-following benchmarks. Our implementation is available at: https://anonymous.4open.science/r/DLM_random_pivot-EE62/
Chinese Translation
扩散语言模型能够实现并行 token 生成，为顺序自回归解码提供了一种有前景的替代方案。现有的加速方法利用预测置信度来实现并行 token 提交，同时保持生成质量。然而，基于置信度的选择优先考虑可靠的 token 预测，而没有明确考虑 token 位置如何影响剩余的 decoding schedule。此外，这种排序并不总能反映 token 间的依赖关系，这表明置信度未必是 token 选择的唯一标准。受这一局限性的启发，我们提出了平衡感知枢轴解码（Balance-Aware Pivot Decoding, BPD），它偏好连续掩码区间内的中心枢轴，并并行解码所选位置。平衡枢轴可在区间长度上产生对数级的解码深度。我们的实用方法通过将基于置信度的预填充与置信度阈值化的中心选择相结合，来考虑模型的不确定性。我们在 LLaDA 和 Dream 上，跨五个数学推理、代码生成和指令遵循基准，评估了由此产生的质量-效率权衡。我们的实现可在以下网址获取：https://anonymous.4open.science/r/DLM_random_pivot-EE62/
cs.CL / 42 / 2610.26296
Answering While You Type: Optimizing Time-to-Answer with Streaming Generation
边输入边作答：利用流式生成优化答案时延
diffusion
扩散模型相关
Abstract
Work on efficient large-language-model inference has concentrated almost entirely on reducing the cost of decoding once a request is complete. The quantity a user experiences, however, is the interval between submitting a request and receiving a usable response, and that interval can be shortened without making decoding itself faster. While a request is being composed it is already partially specified, yet conventional deployments begin no computation until submission. We propose and study streaming generation, a decoding regime in which the model begins producing a response from an incomplete request and revises it as the remaining input arrives, so that only a small residue of the answer remains to be computed once the request is submitted. We evaluate the regime on autoregressive and diffusion language models, characterise the conditions under which it reduces user-visible latency, and analyse the mechanism responsible.
Chinese Translation
关于高效大语言模型推理的研究几乎完全集中在降低请求完成后的解码成本上。然而，用户所体验到的量是提交请求与收到可用响应之间的时间间隔，而即使不加快解码本身，这一间隔也可以被缩短。当请求正在被编写时，它已经被部分地确定，然而传统部署直到提交才开始任何计算。我们提出并研究流式生成，一种解码范式，其中模型从尚不完整的请求开始生成响应，并在剩余输入到达时对其进行修正，从而在请求提交时只剩少量答案残余需要计算。我们在自回归和扩散语言模型上评估该范式，刻画其降低用户可见延迟的条件，并分析其中的机制。
cs.CL / 43 / 2610.26474
AfriGuard: Safety and Robust Refusal For African Languages
AfriGuard：面向非洲语言的安全与稳健拒绝
large language model
大语言模型相关
Abstract
Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content. However, ensuring appropriate responses becomes more challenging in diverse linguistic and cultural contexts, where safety risks may be expressed or interpreted differently. Building safe AI systems therefore requires detecting safety risks, understanding their nature, and responding appropriately, while accounting for linguistic and cultural variations that add complexity to each stage. In this work, we introduce AfriGuard, a culturally grounded multilingual dataset for studying harmful intent, safety-category understanding, and refusal behavior across 10 African languages and English. Our findings reveal that guardrail models underperform both open- and closed-source LLMs on safety-intent and safety-category classification, while models show limited understanding of appropriate safety behavior, scoring below 60% F1. We further find that models often fail to respond appropriately to unsafe prompts not because they explicitly refuse, but because they produce linguistically inadequate responses that result in incidental non-assistance rather than proper refusals. Finally, using AfriGuard, we develop efficient guardrail models that achieve performance competitive with models up to 30x larger, introduce refusal detectors that improve average performance by 43.9% over existing detectors, and demonstrate that supervised fine-tuning can substantially improve overall safety performance.
Chinese Translation
旨在与用户交互的大型语言模型（LLM）通常经过训练，以对有害或有毒内容作出恰当回应。然而，在多样化的语言与文化语境中，确保恰当的回应变得更具挑战性，因为安全风险可能以不同方式被表达或解读。因此，构建安全的 AI 系统需要检测安全风险、理解其性质并作出恰当回应，同时还要考虑语言与文化差异，这些差异为每个阶段都增添了复杂性。在本工作中，我们提出 AfriGuard，这是一个具有文化根基的多语言数据集，用于研究 10 种非洲语言和英语中的有害意图、安全类别理解以及拒绝行为。我们的发现表明，在安全意图和安全类别分类任务上，防护栏（guardrail）模型的表现不如开源和闭源 LLM，而模型对恰当安全行为的理解有限，F1 分数低于 60%。我们进一步发现，模型往往无法对不安全的提示作出恰当回应，其原因并非它们明确拒绝，而是因为它们生成了在语言上不充分的回应，从而导致偶然性的不提供帮助，而非恰当的拒绝。最后，借助 AfriGuard，我们开发了高效的防护栏模型，其性能可与规模最多大 30 倍的模型相竞争，引入了拒绝检测器，相较现有检测器平均性能提升 43.9%，并证明监督微调能够显著提升整体安全性能。
cs.CL / 44 / 2610.26531
ReAnchor: Reusing Rejected Verifier States for Diffusion Drafting
ReAnchor：复用被拒绝的验证器状态进行扩散草拟
diffusion
扩散模型相关
Abstract
Block diffusion drafters accelerate speculative decoding by generating multiple draft tokens in a single forward pass. However, each pass produces only a fixed-length block from a single target-generated anchor, limiting how far the draft can extend. Meanwhile, target verification computes hidden states even at positions beyond the first rejection, but these states are typically discarded rather than reused. We propose ReAnchor, a training-free method that reanchors a pretrained drafter at shifted positions using discarded verifier states to generate parallel draft blocks. These blocks improve predictions within the original block and extend coverage beyond its endpoint. At sampling temperature of T=0, ReAnchor achieves its largest average gains on Gemma 4-12B, improving acceptance length by 19.4% and decoding speed by 17.9% across eight benchmarks. Averaged equally over four target models and eight benchmarks, it improves acceptance length by 15.2% and decoding speed by 13.4% with pretrained DFlash drafters.
Chinese Translation
块扩散草稿器通过在单次前向传播中生成多个草稿 token 来加速推测解码。然而，每次传递仅从单个由目标模型生成的锚点产生一个固定长度的块，限制了草稿能够延伸的距离。与此同时，目标验证甚至在首个拒绝位置之后的位置上计算隐藏状态，但这些状态通常被丢弃而不是被复用。我们提出 ReAnchor，一种无需训练的方法，它使用被丢弃的验证器状态将预训练草稿器在偏移位置重新锚定，以生成并行的草稿块。这些块改进了原始块内的预测，并将覆盖范围扩展到其端点之外。在采样温度为 T=0 时，ReAnchor 在 Gemma 4-12B 上取得其最大的平均增益，在八个基准上将接受长度提高了 19.4%，并将解码速度提高了 17.9%。在四个目标模型和八个基准上取等权平均，它使用预训练的 DFlash 草稿器将接受长度提高了 15.2%，并将解码速度提高了 13.4%。
cs.CL / 45 / 2610.26601
BERT Is Already a Denoiser: Cheaply Unlocking Iterative Generation in Masked Language Models
BERT 已经是一个去噪器：廉价地释放掩码语言模型中的迭代生成能力
diffusion
扩散模型相关
Abstract
Masked language models (MLMs) such as BERT are typically used for representation learning despite their denoising pretraining. We test whether this capability supports conditional generation without diffusion training. Our decoder repeatedly predicts masked tokens and commits high-confidence predictions using a cosine schedule. It requires no diffusion timestep or reverse-diffusion objective. Frozen BERT already exhibits substantial denoising capability, reconstructing 61.0%, 37.5%, and 24.4% of masked tokens at 15%, 45%, and 60% corruption, respectively, although generation deteriorates sharply as visible context disappears. This capability alone does not yield task-directed generation: frozen BERT reaches only 11.0% on ARC-Easy and 4.2% on MMLU. Ordinary MLM adaptation, however, rapidly converts the same architecture into a usable generator. Full adaptation reaches 28.2% and 24.6%, respectively, while parameter-efficient adaptation substantially reduces unmatched outputs. On open-ended conditional generation, full MLM adaptation reaches BLEU/ROUGE-L of 0.163/0.583 on QQP and 0.107/0.327 on Quasar-T, placing it in the same broad performance range as published BERT models trained through diffusion without explicit diffusion training. Finally, applying the same generative interface to stronger masked representations substantially improves performance: RoBERTa-large reaches 44.2% on ARC-Easy and 33.8% on MMLU. These results separate three components of masked iterative generation: pretraining supplies the representation and denoising operator, task adaptation determines what the model generates, and an iterative commitment rule supplies the decoding process. Explicit diffusion training is therefore not necessary to obtain effective iterative conditional generation from pretrained MLMs in the settings studied here.
Chinese Translation
诸如 BERT 之类的掩码语言模型（MLM）尽管经过去噪预训练，通常仍被用于表示学习。我们检验这种能力能否在不进行扩散训练的情况下支持条件生成。我们的解码器反复预测被掩码的 token，并使用余弦调度提交高置信度的预测。它不需要扩散时间步，也不需要反向扩散目标。冻结的 BERT 已经展现出可观的去噪能力，在 15%、45% 和 60% 的破坏率下分别重建了 61.0%、37.5% 和 24.4% 的被掩码 token，尽管随着可见上下文消失，生成质量会急剧下降。仅凭这一能力并不能产生面向任务的生成：冻结的 BERT 在 ARC-Easy 上仅达到 11.0%，在 MMLU 上仅达到 4.2%。然而，普通的 MLM 适配能迅速将同一架构转变为可用的生成器。完全适配分别达到 28.2% 和 24.6%，而参数高效的适配则大幅减少了不匹配的输出。在开放式条件生成任务上，完全 MLM 适配在 QQP 上达到 0.163/0.583 的 BLEU/ROUGE-L，在 Quasar-T 上达到 0.107/0.327，使其处于与已发表的、通过扩散训练、但没有显式扩散训练的 BERT 模型相同的宽泛性能区间。最后，将相同的生成接口应用于更强的掩码表示可显著提升性能：RoBERTa-large 在 ARC-Easy 上达到 44.2%，在 MMLU 上达到 33.8%。这些结果区分了掩码迭代生成的三个组成部分：预训练提供表示与去噪算子，任务适配决定模型生成什么，而迭代提交规则提供解码过程。因此，在此处所研究的设定下，要从预训练 MLM 获得有效的迭代条件生成，显式扩散训练并非必要。
cs.CL / 46 / 2610.26730
EpiLens: Lightweight RAG for Epidemiological Knowledge Acquisition
EpiLens：用于流行病学知识获取的轻量级 RAG
large language model
大语言模型相关
Abstract
To prepare for future outbreaks, researchers need to know what studies have been done and what data they use. Mapping a large literature by hand is slow, while large language models alone can misstate or conflate sources. We present EpiLens, a lightweight, open-source retrieval-augmented generation package that applies a fixed retrieval template and output schema to every paper in a corpus and links outputs to source passages. We build a corpus of 2,297 papers from the sources of a multi-pathogen parameter study and evaluate EpiLens's paper-classification workflow against 214 manually labelled papers. Without task-specific training data, the classifier matches or exceeds the tested supervised and zero-shot baselines across four tasks. The study's early extraction sheets cite published sources for 327 estimates. We link each estimate to a cited document or to corpus studies included by a review. Then, we use the classifier's labels to compare the papers associated with different parameters and pathogens. EpiLens labels more papers linked to reproduction numbers than fatality risks as model-based (54% versus 22%, compared with 62% versus 24% in the manual labels). The gap is consistent with the frequent use of transmission models to estimate reproduction numbers and persists within each of five pathogens compared. Beyond the contrast between parameters, the model-based share of papers linked to incubation-period estimates differs by pathogen: 55% for SARS-CoV-2 versus 6% for mpox.
Chinese Translation
为了为未来的疫情暴发做好准备，研究人员需要了解已经开展了哪些研究以及这些研究使用了哪些数据。手工梳理大量文献很慢，而仅依赖大语言模型可能会错误陈述或混淆来源。我们提出 EpiLens，这是一个轻量级、开源的检索增强生成软件包，它对语料库中的每篇论文应用固定的检索模板和输出模式，并将输出链接到来源段落。我们基于一项多病原体参数研究的来源构建了包含 2,297 篇论文的语料库，并针对 214 篇人工标注论文评估 EpiLens 的论文分类工作流。在没有任务特定训练数据的情况下，该分类器在四项任务上达到或超过了所测试的监督式基线和零样本基线。该研究的早期提取表为 327 个估计值引用了已发表来源。我们将每个估计值链接到一篇被引文献，或链接到一篇综述所纳入的语料库研究。然后，我们使用分类器的标签来比较与不同参数和病原体相关的论文。EpiLens 将更多与再生数相关的论文而非与病死风险相关的论文标记为基于模型的（54% 对 22%，人工标签中为 62% 对 24%）。这一差距与频繁使用传播模型来估计再生数相一致，并且在所比较的五种病原体中每一种内部都持续存在。除了参数之间的对比之外，与潜伏期估计值相关的论文中基于模型的论文所占比例因病原体而异：SARS-CoV-2 为 55%，而 mpox 为 6%。
cs.CL / 47 / 2610.26809
Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability
安全中期训练：将大语言模型安全性内化为一项基础能力
large language model
大语言模型相关
Abstract
Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come at the cost of general capability. This tension suggests that safety and general reasoning should not be treated as separate training goals. We instead view safety as an LLM capability that can be learned through a reasoning structure shared with general tasks. Both require the model to understand intent or task goals, apply relevant knowledge, and reason about potential consequences. We introduce safety mid-training, a stage between pre-training and post-training that jointly develops safety and general reasoning through this shared structure. Given safety and utility prompts, the base model generates structured documents and reasoning traces that are then used for mid-training. Safety mid-training reduces average attack success rate (ASR) by 72% to 81% relative to Qwen3-1.7B, 4B, and 8B base models, while also improving general capability. The benefits persist through subsequent post-training. After general-only SFT, mid-trained models largely retain their safety, achieving lower ASR than the baselines without sacrificing general capability. They also retain safety under general-only GRPO without safety rewards, while subsequent RL reaches a better safety-utility frontier when initialized from the mid-trained models. Together, these results suggest that learning safety and general reasoning within a shared mid-training stage can reduce the safety-utility trade-off and make safety more robust to subsequent training.
Chinese Translation
大语言模型（LLMs）的安全对齐主要在训练后阶段进行，而近期的工作也开始将安全数据纳入预训练。然而，当后续训练不再包含安全数据时，安全性可能会退化，而且安全性的提升往往以通用能力为代价。这种张力表明，安全与通用推理不应被视为彼此分离的训练目标。我们转而将安全性视为一种 LLM 能力，它可以通过与通用任务共享的推理结构来习得。二者都要求模型理解意图或任务目标、应用相关知识，并对潜在后果进行推理。我们提出安全中期训练，这是一个位于预训练与训练后之间的阶段，通过这一共享结构联合发展安全性与通用推理。给定安全提示与效用提示，基座模型生成结构化文档与推理轨迹，这些内容随后被用于中期训练。相对于 Qwen3-1.7B、4B 和 8B 基座模型，安全中期训练将平均攻击成功率（ASR）降低 72% 至 81%，同时还提升了通用能力。这些收益在随后的训练后阶段依然保持。在仅使用通用数据的 SFT 之后，经过中期训练的模型在很大程度上保留了其安全性，在不牺牲通用能力的情况下实现了低于基线的 ASR。在没有安全奖励的仅通用 GRPO 下，它们同样保持了安全性；而当从经中期训练的模型初始化时，后续的强化学习能够达到更优的安全—效用前沿。综合来看，这些结果表明，在共享的中期训练阶段中学习安全性与通用推理，可以缓解安全—效用权衡，并使安全性对后续训练更加鲁棒。
cs.CL / 48 / 2610.26819
ASTOR: Benchmarking and Improving Autoformulation for Soft-Constraint Optimization with a Formal Grammar
ASTOR：使用形式文法对面向软约束优化的自动形式化进行基准测试与改进
large language model
大语言模型相关
Abstract
Real-world decision problems combine two qualitatively different kinds of constraints: requirements that must be satisfied, called hard constraints, and preferences that can be violated, called soft constraints. Because these problems are often posed to decision-makers in natural language, large language models offer a natural tool for assisting with their solution. This has given rise to autoformulation, which focuses on using LLMs to translate natural-language descriptions of decision problems into executable mathematical optimization models. However, this research often neglects soft constraints like those proposed by consumers, managers, and planners, despite the fact that many practical problems require modeling them explicitly. We introduce ASTOR, an inference-time framework that maps a natural-language decision problem to a grammar-guided abstract syntax tree (AST) before deterministically compiling the resulting representation into executable optimization code. We also introduce SoftOR, a unified benchmark of natural language problems which focuses on soft constraint formulations. On SoftOR, ASTOR achieves the best aggregate accuracy among comparable single-request and multi-agent LLM methods while remaining competitive on standard autoformulation benchmarks that include only hard constraints. Code and data are available at https://anonymous.4open.science/r/astor-softor-iclr-E34C/README.md.
Chinese Translation
现实世界的决策问题结合了两类性质不同的约束：必须被满足的要求，称为硬约束；以及可以被违反的偏好，称为软约束。由于这些问题通常以自然语言形式呈现给决策者，大型语言模型为协助求解这些问题提供了一种自然的工具。这催生了自动形式化，其重点在于使用 LLMs 将决策问题的自然语言描述转换为可执行的数学优化模型。然而，这项研究常常忽视消费者、管理者和规划者所提出的软约束，尽管许多实际问题需要显式地对它们建模。我们介绍了 ASTOR，一个推理时框架，它先将自然语言决策问题映射为文法引导的抽象语法树（AST），然后将所得表示确定性地编译为可执行的优化代码。我们还介绍了 SoftOR，一个统一的自然语言问题基准，其聚焦于软约束形式化。在 SoftOR 上，ASTOR 在可比的单请求和多智能体 LLM 方法中取得了最佳总体准确率，同时在仅包含硬约束的标准自动形式化基准上保持竞争力。代码和数据可在 https://anonymous.4open.science/r/astor-softor-iclr-E34C/README.md 获取。
cs.CL / 49 / 2610.26840
LatComm: Reusable Context Compression for Latent Communication between Large Language Models
LatComm：用于大型语言模型之间潜在通信的可复用上下文压缩
large language model
大语言模型相关
Abstract
Systems that combine several large language models (LLMs) often use one context for many questions. When the context is passed between models as text, the receiving model processes all of its tokens again for every question. Avoiding this cost requires a context representation that is compact, reusable across questions, and readable by a different LLM. Existing compression and latent communication methods each provide only some of these properties. We introduce LatComm, a Latent Communication interface that provides all three by separating encoding from reading: the context is encoded once without the question, and the receiver selects what to read for each question. A frozen context encoder LLM compresses the context into a fixed-size packet and sends it to a different frozen receiver LLM, which reads the packet through cross-attention, with gates computed from its prompt. On 12 multiple-choice question-answering datasets, LatComm improves average accuracy over single-vector baselines by up to 10.2 points. On QuALITY, it reduces the receiver's input by 57x and answers the questions of an article 3x faster.
Chinese Translation
将多个大型语言模型（LLM）组合起来的系统常常将同一上下文用于许多问题。当上下文以文本形式在模型之间传递时，接收模型会针对每个问题重新处理其全部 token。要避免这一开销，需要一种紧凑、可跨问题复用并且能被另一个 LLM 读取的上下文表示。现有的压缩方法和潜在通信方法各自只提供了其中部分特性。我们提出 LatComm，一种潜在通信接口，它通过将编码与读取分离来同时提供上述三项特性：上下文在没有问题的情况下被编码一次，而接收方针对每个问题选择读取什么。一个冻结的上下文编码器 LLM 将上下文压缩成一个固定大小的数据包，并将其发送给另一个冻结的接收器 LLM，后者通过交叉注意力读取该数据包，其中的门控由其提示计算得到。在 12 个多项选择问答数据集上，LatComm 相较于单向量基线将平均准确率最多提高了 10.2 个百分点。在 QuALITY 上，它将接收方的输入减少了 57 倍，并使回答一篇文章问题的速度快 3 倍。
cs.CL / 50 / 2610.26860
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
OpenVLHarness：具备多模态记忆的真实世界视觉智能体工具使用
large language model
大语言模型相关
Abstract
Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply increasing the inference budget of the base model yields only modest gains. Such tasks demand grounded perception, external knowledge, and numerical measurements obtained through heterogeneous tools, whose outputs must be made interpretable to the model and retained across reasoning steps. We introduce OpenVLHarness, a multimodal reasoning harness organized around a capability layer and persistent multimodal memory. The capability layer combines task-level tool interfaces with evidence rendering, providing consistent access to heterogeneous backends and converting their outputs into model-readable observations. Persistent multimodal memory retains the corresponding images, structured data, and web references as addressable artifacts throughout each query. A coding interface executes generated programs over stored data, enabling computations that combine results from multiple tools. Together, these mechanisms support an iterative process of acquiring, inspecting, and composing multimodal evidence. Evaluated on 22 benchmarks with eight open and proprietary backbones, OpenVLHarness improves average scores by 9.7-12.6 points over matched no-tool baselines, including +11.2 with Qwen3-VL-8B, +9.7 with Kimi K3, and +9.9 with GPT-6 Astra. With GPT-6 Luna at medium reasoning effort, OpenVLHarness outperforms the base model, Codex, and Visual Sketchpad at maximum effort by 7.7-8.5 points while incurring at most 37% of their model-token cost.
Chinese Translation
多模态大语言模型（MLLMs）在视觉推理方面取得了快速进展，然而许多任务所需的证据难以仅通过内部推理获得：我们在 22 个基准上的对照研究表明，仅仅增加基础模型的推理预算只能带来有限的收益。此类任务需要有依据的感知、外部知识以及通过异构工具获得的数值测量，而这些工具的输出必须可被模型解释，并在推理步骤之间得到保留。我们提出 OpenVLHarness，一种围绕能力层与持久多模态记忆组织的多模态推理框架。能力层将任务级工具接口与证据渲染相结合，提供对异构后端的一致访问，并将其输出转换为模型可读的观测结果。持久多模态记忆在每个查询过程中将相应的图像、结构化数据和网络引用保留为可寻址制品。编码接口在存储的数据上执行生成的程序，从而支持组合多个工具结果的计算。这些机制共同支持获取、检查和组合多模态证据的迭代过程。在 22 个基准上使用八个开源与专有骨干模型进行评估，OpenVLHarness 相比匹配的无工具基线将平均得分提高 9.7–12.6 分，其中使用 Qwen3-VL-8B 提高 +11.2 分，使用 Kimi K3 提高 +9.7 分，使用 GPT-6 Astra 提高 +9.9 分。在 GPT-6 Luna 以中等推理强度运行时，OpenVLHarness 以 7.7–8.5 分的优势优于以最大推理强度运行的基础模型、Codex 和 Visual Sketchpad，同时产生的模型 token 成本至多为其 37%。
cs.CR / 51 / 2610.25705
How Effectively Do Security Skills Guide Coding Agents in Vulnerability Detection in Linux Kernel?
安全技能在多大程度上有效指导编码智能体进行 Linux 内核漏洞检测？
large language model
大语言模型相关
Abstract
Large language model (LLM) coding agents, such as Claude Code, are increasingly used for security auditing, with task-specific skills providing reusable vulnerability-analysis procedures. However, their effectiveness in detecting real vulnerabilities and the reasons for their failures remain unclear. We conduct the first large-scale empirical study of 37 security skills using Claude Code with four underlying LLMs on 100 real Linux kernel vulnerabilities across ten classes, totaling 14800 runs. We evaluate vulnerability detection at the function, statement, and vulnerability-class levels. Results show that 24.4 percent of runs locate the correct function, 18.3 percent identify both the function and responsible statement, and 15.7 percent correctly classify the vulnerability. Our failure analysis reveals limitations in evidence acquisition, state-relation reasoning, code interpretation, candidate verification, investigation coverage, and reporting. Based on these findings, we propose SKILL-SIF, a self-improving framework that iteratively evolves security skills by learning from missed vulnerabilities and failure patterns. We evolve five top-performing skills using the original benchmark and evaluate them on 100 previously unseen vulnerabilities. SKILL-SIF improves fully correct vulnerability detection by up to eight times (from 2 percent to 16 percent) on unseen vulnerabilities, while aggregate accuracy across four LLMs increases by 1.85 times (from 6.05 percent to 11.20 percent). These findings highlight the limitations of existing security skills and demonstrate the potential of failure-driven skill evolution to improve vulnerability detection in coding agents.
Chinese Translation
大型语言模型（LLM）编码智能体（如 Claude Code）正越来越多地被用于安全审计，任务特定的技能提供了可复用的漏洞分析流程。然而，它们在检测真实漏洞方面的有效性以及失败的原因仍不清楚。我们开展了首个大规模实证研究，使用 Claude Code 与四种底层 LLM，针对十类共 100 个真实 Linux 内核漏洞，对 37 项安全技能进行了评估，总计 14800 次运行。我们从函数、语句和漏洞类别三个层面评估漏洞检测。结果表明，24.4% 的运行定位到了正确的函数，18.3% 同时识别出函数和应负责的语句，15.7% 正确地对漏洞进行了分类。我们的失败分析揭示了在证据获取、状态关系推理、代码解释、候选验证、调查覆盖范围和报告方面的局限性。基于这些发现，我们提出了 SKILL-SIF，一个通过从漏检漏洞和失败模式中学习来迭代演化安全技能的自改进框架。我们使用原始基准演化五项表现最佳的技能，并在 100 个此前未见过的漏洞上对它们进行评估。SKILL-SIF 在未见过的漏洞上将完全正确的漏洞检测提升了最多八倍（从 2% 提升到 16%），而四个 LLM 的总体准确率提升了 1.85 倍（从 6.05% 提升到 11.20%）。这些发现凸显了现有安全技能的局限性，并展示了失败驱动的技能演化在改进编码智能体漏洞检测方面的潜力。
cs.CR / 52 / 2610.25791
Towards Architectural-Level Safety Control: Relocating Hazardous Knowledge into Removable Experts in Mixture-of-Experts Models
迈向架构级安全控制：将危险知识迁移到混合专家模型中的可移除专家
large language model
大语言模型相关
Abstract
Large language models acquire broad knowledge from web-scale corpora during pretraining, including knowledge that could be hazardous if misused. Removing such knowledge from the model itself without degrading general capability is desirable but difficult, as it is deeply entangled with general capability within the same parameters. Existing safety alignment and unlearning methods typically adjust these shared parameters, tending to suppress the expression of hazardous knowledge rather than remove it. As a result, the knowledge remains in the model and is readily recoverable by subsequent fine-tuning. To address this limitation, we propose REX (Relocating into Removable EXperts), a knowledge relocation method that separates hazardous knowledge from general capability at the architectural level. Exploiting the modular structure of Mixture-of-Experts models, REX expands the model with a set of new, physically removable experts designated to hold hazardous knowledge. Through expert expansion, selective gradient masking and routing differentiation, it relocates hazardous knowledge out of the original experts and concentrates it in the new ones. Excising these new experts yields a model suitable for public release. This removal holds consistently across Intern-S2, Qwen3.5-35B-A3B and Gemma-4-26B-A4B; on Qwen3.5-35B-A3B, for instance, excision lowers WMDP-Bio accuracy by 48.4 points while MMLU changes by less than 3 points. The excised model matches the strongest of nine unlearning baselines in removing hazardous knowledge, and it is the only one that preserves this removal under subsequent fine-tuning. After 250M tokens of standard fine-tuning on Qwen3.5, the best baseline rebounds on WMDP-Bio from 44.1% to 78.3%, whereas REX rises only from 32.4% to 42.8%.
Chinese Translation
大型语言模型在预训练过程中从网络规模语料中获取广泛知识，其中包括一旦被误用就可能造成危险的知识。在不降低通用能力的情况下从模型本身移除此类知识是理想的做法，但却十分困难，因为它在同一组参数中与通用能力深度纠缠。现有的安全对齐与遗忘方法通常调整这些共享参数，往往只是抑制危险知识的表达，而非将其移除。因此，这些知识仍留在模型中，并且很容易通过后续微调恢复。为解决这一局限，我们提出 REX（Relocating into Removable EXperts），一种在架构层面将危险知识与通用能力分离的知识迁移方法。REX 利用混合专家模型的模块化结构，通过一组新的、物理上可移除的专家来扩展模型，这些专家被指定用于承载危险知识。通过专家扩展、选择性梯度掩蔽和路由差异化，它将危险知识从原始专家中迁移出去，并将其集中到新专家中。切除这些新专家可得到一个适合公开发布的模型。这种移除效果在 Intern-S2、Qwen3.5-35B-A3B 和 Gemma-4-26B-A4B 上均一致成立；例如，在 Qwen3.5-35B-A3B 上，切除使 WMDP-Bio 准确率下降 48.4 个百分点，而 MMLU 变化不到 3 个百分点。经切除后的模型在移除危险知识方面与九个遗忘基线中最强者相当，并且是唯一一个在后续微调下仍能保持这种移除效果的模型。在 Qwen3.5 上进行 250M 个 token 的标准微调后，最佳基线在 WMDP-Bio 上从 44.1% 回升到 78.3%，而 REX 仅从 32.4% 上升到 42.8%。
cs.CR / 53 / 2610.25897
Budgeted Attack Scheduling for LLM Red-Teaming: The Roles of Static Priors, Semantic Guidance, and Online Feedback
面向大语言模型红队评估的预算约束攻击调度：静态先验、语义引导与在线反馈的作用
large language model
大语言模型相关
Abstract
As jailbreak attacks against large language models (LLMs) become increasingly diverse, red-teaming has come to rely on large libraries of jailbreak prompts and attack strategies to evaluate the robustness of victim models and their defense settings. However, attack effectiveness is highly condition-dependent: an attack that succeeds under one victim-defense setting may fail under another, while exhaustively evaluating the full attack library for each setting incurs substantial token and monetary costs. This motivates studying how jailbreak attacks should be prioritized under a limited attempt budget for efficient robustness evaluation. In this work, we formalize this problem as budgeted attack scheduling over a fixed attack library and investigate three complementary signals under a fixed victim-defense condition: (1) static historical attack success rate (ASR) priors; (2) semantic guidance from LLMs; and (3) online feedback from earlier queries. Building on these signals, we design four scheduling strategies and evaluate them across three victim models and three defense settings on GPT-5.2. Extensive experiments clearly characterize the conditions under which each signal is most effective. Historical ASR provides a strong initialization, but its reliability degrades under defense-induced distribution shifts. Semantic guidance consistently improves or matches static prior-based scheduling in the cold-start regime, while online feedback becomes particularly valuable when transferred priors become stale. Semantic guidance and online feedback can also partially compensate for missing historical priors. Taken together, these findings characterize budgeted red-teaming as an attack scheduling problem in which historical priors, semantic guidance, and online feedback play complementary roles.
Chinese Translation
随着针对大语言模型（LLM）的越狱攻击日益多样化，红队评估已逐渐依赖大规模的越狱提示与攻击策略库，以评估受害模型及其防御设置的鲁棒性。然而，攻击的有效性高度依赖于具体条件：在某一受害模型—防御设置下取得成功的攻击，在另一设置下可能失败，而针对每种设置穷尽评估整个攻击库会带来可观的 token 开销与金钱成本。这促使我们研究：在有限的尝试预算下，应如何对越狱攻击进行优先级排序，以实现高效的鲁棒性评估。在本工作中，我们将该问题形式化为在固定攻击库上的预算约束攻击调度问题，并在固定的受害模型—防御条件下考察三种互补信号：（1）静态的历史攻击成功率（ASR）先验；（2）来自 LLM 的语义引导；（3）来自先前查询的在线反馈。基于这些信号，我们设计了四种调度策略，并在 GPT-5.2 上跨三个受害模型与三种防御设置对它们进行评估。大量实验清晰地刻画了每种信号最为有效的条件。历史 ASR 提供了强有力的初始化，但在防御引起的分布偏移下其可靠性会下降。在冷启动阶段，语义引导能够持续优于或持平于基于静态先验的调度，而当迁移的先验变得过时时，在线反馈则变得尤为有价值。语义引导与在线反馈还可以部分弥补历史先验的缺失。综合来看，这些发现将预算约束下的红队评估刻画为一个攻击调度问题，其中历史先验、语义引导与在线反馈发挥互补作用。
cs.CR / 54 / 2610.25937
RAGLeak: Supply-Chain Privacy Backdoors in Retrieval-Augmented Generation
RAGLeak：检索增强生成中的供应链隐私后门
large language model
大语言模型相关
Abstract
Retrieval-Augmented Generation (RAG) enables large language models to access external knowledge during inference, but its reliance on external knowledge bases also introduces new attack surfaces. Existing data extraction attacks against RAG systems rely on either prompt injection, which heavily depends on the generator's instruction-following capability, or backdoors that reproduce or paraphrase retrieved documents but struggle to balance extraction fidelity and stealthiness. In this paper, we propose RAGLeak, a supply-chain privacy backdoor attack that decouples information leakage from direct textual exposure. When triggered, the generator reversibly obfuscates retrieved private documents for high-fidelity recovery by an auxiliary LLM. Across three model families and three application domains, RAGLeak achieves up to $100\%$ ASR with limited impact on benign utility and at most a $10$-percentage-point ASR drop under representative detection-based defenses. The backdoor remains effective after clean fine-tuning and across model scales and versions, revealing an underexplored supply-chain privacy risk in RAG systems. The source code is available at https://anonymous.4open.science/r/RAGLeak-4B67.
Chinese Translation
检索增强生成（RAG）使大语言模型能够在推理期间访问外部知识，但其对外部知识库的依赖也引入了新的攻击面。现有针对 RAG 系统的数据提取攻击要么依赖提示注入——这高度依赖生成器的指令遵循能力——要么依赖后门，这些后门能够复现或改写检索到的文档，但难以在提取保真度与隐蔽性之间取得平衡。在本文中，我们提出 RAGLeak，一种供应链隐私后门攻击，它将信息泄露与直接文本暴露解耦。当被触发时，生成器会对检索到的私有文档进行可逆混淆，以便由辅助 LLM 高保真恢复。在三个模型系列和三个应用领域中，RAGLeak 实现了高达 $100\%$ 的 ASR，同时对良性效用的影响有限，并且在代表性的基于检测的防御下 ASR 最多下降 $10$ 个百分点。该后门在干净微调之后以及跨模型规模和版本仍然有效，揭示了 RAG 系统中一个尚未被充分探索的供应链隐私风险。源代码可在 https://anonymous.4open.science/r/RAGLeak-4B67 获取。
cs.CR / 55 / 2610.26001
SPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in Generated Code
SPLASH：用于检测和定位生成代码中安全幻觉的对比审计压力
large language model
大语言模型相关
Abstract
Code generated by large language models often passes every functional test while remaining insecure. These security hallucinations are common, written with high confidence, and repeated across samples, so neither the generator's own uncertainty nor static analyzers reliably catch them. We introduce SPLASH, which is guided by a new measure called audit pressure. It contrasts teacher-forced log-probabilities for the generated program under a security review and a style review. Aggregated to statements and programs, and grounded in a probabilistic model of competing edits, audit pressure allows SPLASH to choose the secure program among a model's candidate programs and locate the vulnerable statement, with no model training, CWE-specific rules, or internal access to the generating model. On 63,684 programs generated by 12 models for 899 tasks, SPLASH picks secure code in 74% of cases, places the vulnerable statement first in 58% of insecure programs, and chooses better than a frontier LLM judge.
Chinese Translation
大型语言模型生成的代码常常通过所有功能测试，却仍然不安全。这些安全幻觉很常见，以高置信度写出，并在样本间重复出现，因此生成器自身的不确定性以及静态分析器都无法可靠地捕捉它们。我们引入 SPLASH，它以一种称为审计压力的新度量为指导。它对比生成程序在安全审查和风格审查下的教师强制对数概率。聚合到语句和程序层面，并建立在竞争性编辑的概率模型之上，审计压力使 SPLASH 能够在模型的候选程序中选择安全程序，并定位易受攻击的语句，而无需模型训练、特定于 CWE 的规则，也无需对生成模型的内部访问。在 12 个模型为 899 个任务生成的 63,684 个程序上，SPLASH 在 74% 的情况下选择了安全代码，在 58% 的不安全程序中将易受攻击的语句排在首位，并且选择效果优于前沿 LLM 评判器。
cs.CR / 56 / 2610.26438
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
假新闻，真实崩盘：针对LLM金融交易智能体的信息注入攻击
large language model
大语言模型相关
Abstract
Large language model (LLM) agents now trade real capital using news as evidence, yet they cannot reliably tell fabricated headlines from real news. Because they read the same feeds under near-identical policies, a single fabricated post can push them into the same trade at the same moment. We present a systematic attack benchmark for information injection against LLM trading agents. Unlike prompt injection, which targets the model's instructions, information injection targets the evidence the agent trusts. The benchmark spans five injection vectors and seven event categories, evaluated over ~10,000 decisions on 15 U.S. stocks. Five findings emerge. (i) Prompt injection achieves 14% attack success rate (ASR), but fabricated news reaches 62% and two 30-word templates reach 100% on all 15 stocks, inverting the usual LLM-security intuition. (ii) For fact-poisoning templates, surface form matters more than content, since removing credibility cues such as a named source or a "BREAKING" tag drops ASR from 100% to 0% while the fake claim remains identical. (iii) Success depends on the agent's state, with a fake indicator panel scoring 91% against an agent about to hold but 12% against one about to buy. (iv) Attacks generalize across model vendors and remain effective when blended, unlabeled, into real news feeds. (v) Most attacks evade standard defenses, since model scaling stops shallow vectors but not fact poisoning, peer debate fails, and only hierarchical aggregation reduces ASR.
Chinese Translation
大语言模型（LLM）智能体现在使用新闻作为证据来交易真实资本，但它们无法可靠地区分伪造标题与真实新闻。由于它们在近乎相同的策略下阅读相同的信息流，一条伪造帖子就能推动它们在相同时刻进行相同的交易。我们提出了一个针对LLM交易智能体的信息注入系统性攻击基准。与针对模型指令的提示注入不同，信息注入针对的是智能体所信任的证据。该基准涵盖五种注入向量和七类事件，在15只美国股票上的约10,000次决策中进行了评估。由此得出五项发现。（i）提示注入达到14%的攻击成功率（ASR），但伪造新闻达到62%，并且两个30词模板在全部15只股票上达到100%，颠覆了通常的LLM安全直觉。（ii）对于事实投毒模板，表面形式比内容更重要，因为移除诸如具名来源或“BREAKING”标签之类的可信度线索会使ASR从100%降至0%，而虚假声明保持不变。（iii）成功与否取决于智能体的状态：伪造指标面板对即将持有的智能体得分为91%，但对即将买入的智能体为12%。（iv）攻击可泛化到不同模型供应商，并且在未加标签地混入真实新闻信息流时仍然有效。（v）大多数攻击能够规避标准防御，因为模型扩展能阻止浅层向量但不能阻止事实投毒，同行辩论失败，且只有分层聚合能降低ASR。
cs.CR / 57 / 2610.26589
Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks
超越攻击成功率：用于评估迭代式LLM越狱的成本感知指标
large language model
大语言模型相关
Abstract
Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce $\mathrm{ASR}_{\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments. Our evaluation considers GCG~\cite{zou2023universal}, PAIR~\cite{chao2025jailbreaking} with a step budget, and a single-step PAP~\cite{zeng2024johnny} baseline on JailBreakBench across three model families. GCG's scores decrease by 7.0--8.7 percentage points after discounting, while PAP's remain unchanged. While for PAIR~\cite{chao2025jailbreaking}, we see a wider discounting range between 3.4-12 percentage points, showing that the "step" quantity differs per iterative strategy. Reporting both metrics makes the discount due to repeated victim-model interaction explicit. Code is available \hyperlink{https://github.com/SuELS-Group/beyond_asr}{here}.
Chinese Translation
攻击成功率（ASR）在评估针对大语言模型的越狱时，并未考虑对受害模型的重复调用。然而，每一次额外调用都会让已部署系统获得又一次检测攻击并进行干预的机会。我们引入 $\mathrm{ASR}_{\mathrm{STEP}}$，它在保留二元成功判定的同时，对额外的受害模型调用进行惩罚。我们的评估在 JailBreakBench 上跨三个模型系列，考虑了 GCG~\cite{zou2023universal}、带有步数预算的 PAIR~\cite{chao2025jailbreaking}，以及单步 PAP~\cite{zeng2024johnny} 基线。经折扣后，GCG 的分数下降 7.0--8.7 个百分点，而 PAP 的分数保持不变。而对于 PAIR~\cite{chao2025jailbreaking}，我们观察到更宽的折扣范围，介于 3.4-12 个百分点之间，这表明“step”这一数量因迭代策略而异。同时报告这两个指标，使得因重复的受害模型交互而产生的折扣变得明确。代码可在 \hyperlink{https://github.com/SuELS-Group/beyond_asr}{here} 获取。
cs.CR / 58 / 2610.26713
Proof-of-Machine-Learning-Inference: Mining Knowledge, Not Hashes
机器学习推理证明：挖掘知识，而非哈希
diffusion
扩散模型相关
Abstract
Bitcoin-style Proof-of-Work (PoW) secures permissionless blockchains but spends substantial computation on hash puzzles which have no utility other than extending the chain. At the same time, we have seen that the rise of generative AI has created significant demand for computational resources that are unfortunately available only to a few players. We propose a solution that solves both problems at the same time. In particular, we present Proof-of-ML-Inference (PoML), a consensus protocol in which miners perform user-requested machine-learning inference as part of their consensus algorithm. We identify exactly what computational assumptions are needed for ML inference to safely replace hash work, formalize those assumptions, and give evidence that LLMs and diffusion models satisfy them. We then formally analyse the security of our protocol by reducing the security of our protocol to the security of the Bitcoin Backbone model. Finally, we empirically evaluate block-time stability, showing that, in a simulated environment, PoML maintains a block production rate close to its calibrated block-time target.
Chinese Translation
比特币风格的工作量证明（PoW）保障了无需许可区块链的安全，但会把大量计算花费在哈希难题上，而这些难题除了延长链之外没有其他用途。与此同时，我们已经看到，生成式AI的兴起对计算资源产生了巨大需求，而遗憾的是，这些资源只有少数参与者能够获得。我们提出一种同时解决这两个问题的方案。特别地，我们提出Proof-of-ML-Inference（PoML），这是一种共识协议，其中矿工将用户请求的机器学习推理作为其共识算法的一部分来执行。我们准确识别了机器学习推理要安全替代哈希工作需要哪些计算假设，将这些假设形式化，并给出证据表明LLMs和扩散模型满足这些假设。然后，我们通过将协议的安全性归约到Bitcoin Backbone模型的安全性，来形式化分析我们协议的安全性。最后，我们通过实验评估区块时间稳定性，表明在模拟环境中，PoML维持的出块率接近其校准的区块时间目标。
cs.AI / 59 / 2610.25481
Multimodal-Conditioned Latent Diffusion Models for Spatiotemporal Brain MRI Forecasting
用于时空脑部MRI预测的多模态条件潜在扩散模型
diffusion
扩散模型相关
Abstract
Spatiotemporal forecasting of 3D anatomical changes is a critical generative vision challenge for modeling Alzheimer's Disease (AD) progression. While brain MRI forecasting from single-modality data has advanced, other key complementary multimodal biomarkers remain underutilized. In this work, we propose the Multimodal-Conditioned Latent Diffusion Model (MC-LDM), a novel framework for patient-specific brain MRI synthesis. MC-LDM leverages multimodal biomarkers---MRI, tau-PET, Clinical Dementia Rating (CDR), and APOE4---during training, but relies only on baseline MRI, APOE4, and patient metadata at inference, reflecting realistic clinical availability. Our contributions are threefold: (1) a Latent Diffusion Model (LDM) fine-tuned via Low-Rank Adaptation (LoRA) for data-efficient learning; (2) a multimodal conditioning that integrates molecular and clinical signals to model heterogeneous disease trajectories; and (3) auxiliary predictors estimating unobserved multimodal biomarkers from MRI features and patient covariates to condition the diffusion model, removing the need for these assessments at inference. Trained on 1,355 visits from 852 subjects in the Alzheimer's Disease Neuroimaging Initiative (ADNI), MC-LDM reduces average regional volumetric MAE by 18% over the second-best baseline on an independent test set of 664 visits. To our knowledge, MC-LDM represents the first diffusion-based framework to leverage multimodal priors for MRI forecasting. This conceptual advancement enables patient-specific trajectory generation that captures aspects of AD heterogeneity not observable from structural anatomy alone, bridging the gap between rich multimodal supervision and the strict constraints of routine clinical deployment. Code is available at: https://github.com/antonioscardace/MC-LDM.
Chinese Translation
3D解剖变化的时空预测是建模阿尔茨海默病（AD）进展的一项关键生成式视觉挑战。虽然基于单模态数据的脑部MRI预测已取得进展，但其他关键的互补多模态生物标志物仍未得到充分利用。在本工作中，我们提出了多模态条件潜在扩散模型（MC-LDM），这是一个用于患者个体化脑部MRI合成的新框架。MC-LDM在训练期间利用多模态生物标志物——MRI、tau-PET、临床痴呆评定量表（CDR）和APOE4——但在推理时仅依赖基线MRI、APOE4和患者元数据，反映了现实中临床可获得的资料。我们的贡献有三方面：（1）通过低秩适配（LoRA）微调的潜在扩散模型（LDM），以实现数据高效学习；（2）整合分子与临床信号以建模异质性病程轨迹的多模态条件化；（3）辅助预测器，从MRI特征和患者协变量估计未观测到的多模态生物标志物，以对扩散模型进行条件化，从而在推理时无需这些评估。MC-LDM在阿尔茨海默病神经影像学倡议（ADNI）中852名受试者的1,355次访视上训练，在包含664次访视的独立测试集上，相较次优基线将平均区域体积MAE降低18%。据我们所知，MC-LDM是首个利用多模态先验进行MRI预测的基于扩散的框架。这一概念性进展使得能够生成患者个体化轨迹，捕捉仅凭结构解剖无法观测到的AD异质性方面，弥合了丰富的多模态监督与常规临床部署的严格约束之间的差距。代码见：https://github.com/antonioscardace/MC-LDM。
cs.AI / 60 / 2610.25533
Modaliff: A Unified Token Diffusion Model for Any-to-Any Generation
Modaliff：一种用于任意到任意生成的统一 Token 扩散模型
diffusion
扩散模型相关
Abstract
Any-to-any generation seeks a single model that translates flexibly among text, image, and audio, yet existing approaches either require fully aligned multimodal tuples or a shared representation that suppresses modality-specific structure. We introduce Modaliff, a novel token-based diffusion model that learns from widely available bimodal pairs while preserving each modality in its native latent sequence. Modaliff assembles conditions and targets into a unified token sequence, with per-token attributes to disambiguate token identities and role-based positional embedding to preserve each modality's internal structure across sequence layouts. This design enables a single denoiser to support diverse generation tasks without task-specific architectural changes or a shared cross-modal latent space. Beyond one-to-one generation, Modaliff supports multiple conditions either by combining single-condition predictions at inference or by lightweight finetuning on captions that jointly describe multiple inputs. Experiments across text, image, and audio show that Modaliff achieves the best fidelity among any-to-any models on most generation tasks while using substantially fewer parameters and training data.
Chinese Translation
任意到任意生成旨在寻求一个单一模型，能够在文本、图像和音频之间灵活转换，然而现有方法要么需要完全对齐的多模态元组，要么需要一种会抑制模态特定结构的共享表示。我们提出了 Modaliff，一种新颖的基于 token 的扩散模型，它能够从广泛可用的双模态数据对中学习，同时将每种模态保留在其原生潜在序列中。Modaliff 将条件和目标组装成一个统一的 token 序列，并使用逐 token 属性来消除 token 身份的歧义，以及基于角色的位置嵌入来在序列布局之间保留每种模态的内部结构。这一设计使单个去噪器能够支持多样化的生成任务，而无需针对特定任务的架构改动或共享的跨模态潜在空间。除了一对一生成之外，Modaliff 还支持多条件生成：既可以在推理时组合单条件预测，也可以通过对联合描述多个输入的标题进行轻量级微调来实现。在文本、图像和音频上的实验表明，Modaliff 在大多数生成任务上实现了任意到任意模型中最佳的保真度，同时使用的参数和训练数据显著更少。
cs.LG / 61 / 2610.25634
Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base
Clean Forcing：基于冻结基座的抗漂移自回归视频扩散
diffusion
扩散模型相关
Abstract
Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect outputs, and errors compound until quality collapses within seconds (drift). Existing methods either require cluster-scale base retraining, apply training-free inference corrections with inconsistent results, or train frozen-base correctors for sampling and cache errors. We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is ~95% systematic across noise seeds. Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts. Because the clean histories can be the base model's own single-shot generations, training needs no real videos, and the merged corrector adds no inference cost. In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data. Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.
Chinese Translation
自回归（AR）视频扩散模型通过将每个新的帧块条件于已经生成的帧，生成任意长度的流式视频。然而，它们受到暴露偏差的影响：在干净上下文上训练的模型必须从自身不完美的输出继续生成，误差会累积，直到质量在数秒内崩溃（漂移）。现有方法要么需要集群规模的基座模型重新训练，要么采用无训练的推理校正但结果不一致，要么针对采样误差和缓存误差训练冻结基座校正器。我们转而发现，漂移在很大程度上是确定性的，因此是可学习的，因为在保持噪声状态固定而仅改变历史时，会揭示出一个反事实速度间隙，该间隙在不同噪声种子下约有 ~95% 是系统性的。基于这一观察，我们提出 Clean Forcing，它在冻结的基座模型上训练一个 5.9M 参数的 LoRA 校正器来回归这一间隙，并带有一个闭环阶段，该阶段加入 DAgger 风格的聚合和漂移收缩目标，以使校正器暴露于其自身的自生成轨迹。由于干净历史可以是基座模型自身的单次生成结果，训练不需要真实视频，并且合并后的校正器不增加推理成本。在 50 秒文本到视频生成中，Clean Forcing 在没有真实视频时将 Δ-drift 从 +11.3 降低到 +1.65，并在使用 40 个真实片段时降低到 -0.16，在使用数据量少 150 倍以上的情况下改进了已发表的最佳结果 +3.57。大量实验表明，Clean Forcing 在长时程视频生成中在美学质量和数据效率方面优于先前方法，并具有降低的漂移。
cs.AI / 62 / 2610.26022
FrameGRPO: Incentivizing Video Reasoning via Multi-Frame Rollouts and Temporal Credit
FrameGRPO：通过多帧 rollout 与时间信用激励视频推理
large language model
大语言模型相关
Abstract
Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has improved video reasoning in multimodal large language models. However, standard video GRPO varies only linguistic trajectories while conditioning every rollout in a group on the same sampled frames, which can produce homogeneous rewards and weaken group-relative learning when the shared visual evidence is insufficient or distracting. We propose \textbf{FrameGRPO}, a simple and effective reinforcement-learning framework that requires neither auxiliary networks nor changes to the model architecture. Specifically, FrameGRPO combines two lightweight components: (1) \textbf{Multi-Frame Rollouts}, which use only uniform temporal subsampling of a single maximal-frame decode to construct aligned views that diversify visual evidence and reasoning trajectories; and (2) \textbf{Gated Temporal Efficiency Credit}, which selectively reinforces a correct lower-ratio trajectory when all rollouts in a fixed full-ratio reference subset fail, without imposing an unconditional preference for fewer frames. Extensive experiments across diverse video understanding and reasoning benchmarks demonstrate robust overall improvements over standard GRPO, highlighting temporal-view diversification as an effective mechanism for exploration and credit assignment in video MLLMs.
Chinese Translation
带可验证奖励的强化学习，尤其是组相对策略优化（GRPO），已提升了多模态大语言模型中的视频推理能力。然而，标准的视频 GRPO 仅变化语言轨迹，而将一组中的每次 rollout 都以相同的采样帧为条件，当共享的视觉证据不足或具有干扰性时，这可能产生同质化的奖励并削弱组相对学习。我们提出 \textbf{FrameGRPO}，一个简单且有效的强化学习框架，既不需要辅助网络，也不需要改变模型架构。具体而言，FrameGRPO 结合了两个轻量级组件：（1）\textbf{多帧 rollout}，其仅通过对单次最大帧解码进行均匀时间子采样来构建对齐视图，从而多样化视觉证据与推理轨迹；（2）\textbf{门控时间效率信用}，当固定的全比例参考子集中的所有 rollout 均失败时，有选择地强化一条正确的低比例轨迹，而不对更少帧施加无条件的偏好。在多样化的视频理解与推理基准上的大量实验表明，相较于标准 GRPO 具有稳健的整体提升，凸显了时间视图多样化作为视频 MLLMs 中探索与信用分配的有效机制。
cs.AI / 63 / 2610.26025
TamedFlow: Trajectory-Aware Distillation\\for Teacher-Assisted Diffusion Sampling
TamedFlow：面向教师辅助扩散采样的轨迹感知蒸馏
diffusion
扩散模型相关
Abstract
Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two acceleration strategies are increasingly important: model compression makes each evaluation cheaper, while adaptive inference avoids selected teacher computation through reuse, forecasting, or speculative execution. Their combination, however, is brittle: directly applying the evaluated adaptive methods to a compressed student can degrade quality, an observed failure mode we call the \textbf{Combination Gap}. We link this gap to a mismatch between pointwise distillation and short-horizon execution: a student trained at teacher states need not remain reliable when its own predictions are chained. We propose \textbf{TamedFlow}, which trains a compressed DiT for this deployed role through cross-architecture feature alignment, a near-identity Velocity Alignment Head, and randomized rollout supervision. At inference, a lightweight Learnable Trajectory Estimator routes completed drafts: accepted drafts bypass the teacher, while rejected drafts roll back and receive one teacher rescue step. Distilling DiT-XL/2 into DiT-B/2, TamedFlow reaches FID $2.58$ at $2.25\times$ measured speedup over DDIM-50 on ImageNet $256\times256$. Ablations support trajectory supervision, and routing controls favor LTE over random decisions at nearly matched latency. Evaluations on ImageNet $512\times512$, MS-COCO, and Latte/FaceForensics extend the empirical scope.
Chinese Translation
扩散Transformer（DiTs）能够实现高保真的图像与视频生成，但其推理需要大量顺序的模型评估。两种加速策略日益重要：模型压缩使每次评估更廉价，而自适应推理通过复用、预测或推测执行来避免选定的教师计算。然而，二者的结合是脆弱的：将所评估的自适应方法直接应用于压缩后的学生模型可能会降低质量，我们将这一观察到的失效模式称为\textbf{Combination Gap}（组合鸿沟）。我们将这一鸿沟与逐点蒸馏和短时程执行之间的不匹配联系起来：在教师状态上训练的学生模型，当其自身预测被链式串联时，未必仍保持可靠。我们提出\textbf{TamedFlow}，它通过跨架构特征对齐、一个近似恒等的速度对齐头（Velocity Alignment Head）以及随机化展开监督，为这一部署角色训练压缩后的DiT。在推理时，一个轻量级的可学习轨迹估计器对已完成的草稿进行路由：被接受的草稿绕过教师，而被拒绝的草稿则回滚并接受一次教师救援步骤。将DiT-XL/2蒸馏为DiT-B/2时，TamedFlow在ImageNet $256\times256$上以相对于DDIM-50的$2.25\times$实测加速达到FID $2.58$。消融实验支持了轨迹监督，并且在几乎匹配的延迟下，路由控制表明LTE优于随机决策。在ImageNet $512\times512$、MS-COCO以及Latte/FaceForensics上的评估扩展了实证范围。
cs.LG / 64 / 2610.26244
RECAP: Reciprocal Sparse Attention and Hierarchical Caching for Efficient Video Diffusion
RECAP：互惠稀疏注意力与分层缓存的高效视频扩散
diffusion
扩散模型相关
Abstract
Diffusion transformers generate high-quality videos but remain computationally expensive. Sparse attention and temporal caching reduce computation within and across denoising steps, respectively, offering a natural composition. However, predictions produced with sparse attention become sources for temporal reuse, coupling their errors along the sampling trajectory. We observe that improving the accuracy of these sources recovers fidelity without changing the reuse schedule, and that sparsification errors remain directionally aligned across nearby steps, suggesting that correction information can be shared rather than recomputed. These observations motivate RECAP, a training-free method that couples sparse attention with hierarchical caching through reciprocal correction, adaptively coordinated by online error feedback. Cached dense refinement corrects omitted interactions in current sparse estimates, while fresh sparse statistics track input changes and regulate the aging correction. The resulting predictions then renew the sources for whole-step prediction reuse, allowing prediction history to advance between dense refreshes. We evaluate RECAP on multiple video diffusion models against sparse-attention methods, caching methods, and their direct compositions. On HunyuanVideo at 720p 5-second generation, RECAP achieves a 3.05x speedup and outperforms all tested direct compositions in both fidelity and inference speed.
Chinese Translation
扩散Transformer能够生成高质量视频，但计算开销仍然高昂。稀疏注意力与时间缓存分别减少去噪步内与跨去噪步的计算量，二者构成一种自然的组合。然而，由稀疏注意力产生的预测会成为时间复用的来源，使其误差沿采样轨迹相互耦合。我们观察到，提高这些来源的精度可以在不改变复用调度的情况下恢复保真度，并且稀疏化误差在相邻步骤之间保持方向一致，这表明校正信息可以被共享而不是被重新计算。这些观察促成了RECAP，一种免训练方法，它通过互惠校正将稀疏注意力与分层缓存耦合起来，并由在线误差反馈进行自适应协调。缓存的稠密细化校正当前稀疏估计中被省略的交互，而新鲜的稀疏统计量跟踪输入变化并调节逐渐老化的校正。由此得到的预测随后更新用于整步预测复用的来源，使预测历史能够在稠密刷新之间推进。我们在多个视频扩散模型上评估RECAP，并与稀疏注意力方法、缓存方法及其直接组合进行比较。在720p、5秒生成的HunyuanVideo上，RECAP实现了3.05倍的加速，并在保真度和推理速度上均优于所有测试过的直接组合。
cs.CR / 65 / 2610.26682
BrushOff: Stroke-Based Re-rendering as a Watermark Removal Attack on Painting
BrushOff：基于笔触的重渲染作为针对绘画的水印移除攻击
diffusion
扩散模型相关
Abstract
Invisible watermarks are used to mark who owns a painting. We show that a simple attack removes the most robust one we test. The attack redraws the watermarked painting with brush strokes, using a pretrained stroke-based renderer. We call this attack BrushOff. It needs no training, no knowledge of the watermark, and no generative model. We test four watermarks on 500 paintings from five styles. VINE, a recent watermark built to survive image editing, survives all 14 conventional distortions we test, with at least 97.8% detection. Repainting lowers its detection to 1.8%, while the output keeps a no-reference quality score close to that of the clean paintings. Diffusion regeneration can also remove VINE, but at the same removal level it moves the image twice as far from the painting in DreamSim distance, and it changes what the painting shows. Because the renderer draws a countable number of strokes, we can control the attack strength step by step. Finally, we test what kind of signal survives. A simple probe that stores bits in the relative brightness of neighboring regions survives repainting, while a probe that uses absolute brightness does not. The same probe fails under diffusion regeneration, while VINE fails under repainting. Robustness to these two attacks is therefore a separate property, and benchmarks should test both.
Chinese Translation
不可见水印被用来标记绘画的拥有者。我们表明，一种简单的攻击就能移除我们所测试的最稳健的水印。该攻击使用预训练的基于笔触的渲染器，以笔触重绘带水印的绘画。我们将此攻击称为 BrushOff。它无需训练、无需知晓水印，也不需要生成模型。我们在来自五种风格的 500 幅绘画上测试了四种水印。VINE 是一种近期提出的、旨在经受图像编辑的水印，它在我们测试的全部 14 种常规失真下都能存活，检测率至少为 97.8%。重绘将其检测率降至 1.8%，同时输出的无参考质量分数与干净绘画的质量分数接近。扩散再生成也能移除 VINE，但在相同移除水平下，它使图像在 DreamSim 距离上偏离原绘画的距离达到两倍，并且会改变绘画所呈现的内容。由于渲染器绘制可计数的笔触数量，我们可以逐步控制攻击强度。最后，我们测试何种信号能够存活。一种将比特存储在相邻区域相对亮度中的简单探针能够在重绘后存活，而使用绝对亮度的探针则不能。同一探针在扩散再生成下失效，而 VINE 在重绘下失效。因此，对这两种攻击的鲁棒性是相互独立的属性，基准测试应同时测试两者。
cs.AI / 66 / 2610.26645
Reasoning-Aware Multicalibrated Reliability Predictors for Language Models
面向语言模型的推理感知多校准可靠性预测器
large language model
大语言模型相关
Abstract
As people increasingly rely on large language models (LLMs) for information, model providers need to communicate meaningful estimates of whether individual responses are correct or, more generally, reliable. In the context of hallucinations, existing work uses multicalibration to make such reliability predictions meaningful across application-relevant groups, such as prompt topics. Topic groups alone, however, can conceal systematic errors: within the same topic and confidence level, responses with different chain-of-thought (CoT) properties may exhibit opposing errors that cancel on average. In this work, we introduce reasoning-aware multicalibration, a framework that supplements application-relevant topic groups with groups derived from observable properties of CoT traces. We highlight that, theoretically, groupings based on CoT are "internally important", i.e., groupings for which the multicalibration guarantee is not societally crucial as it is for high-risk topics, but whose inclusion in the multicalibration procedure could easily and significantly improve predictive accuracy of the final reliability predictor. Empirically, we apply our framework to hallucination prediction and unsafe-response detection across a variety of benchmarks, and find that adding CoT groups to topic-based calibration reduces Brier score in both tasks. Interestingly, we find that multicalibrating with respect to CoT groups alone (without any topic information) already improves calibration across topics, suggesting that the predictor's errors depend on how the model reasoned, rather than what it was asked about in the prompt.
Chinese Translation
随着人们日益依赖大语言模型（LLMs）获取信息，模型提供者需要传达有意义的估计，说明单个回答是否正确，或者更一般地说，是否可靠。在幻觉背景下，现有工作使用多校准来使此类可靠性预测在应用相关群体（例如提示主题）上具有意义。然而，仅靠主题群体可能掩盖系统性错误：在同一主题和置信水平内，具有不同思维链（CoT）属性的回答可能表现出相反的错误，从而在平均意义上相互抵消。在这项工作中，我们引入推理感知多校准，这是一个框架，它用由 CoT 轨迹的可观察属性推导出的群体来补充应用相关主题群体。我们从理论上强调，基于 CoT 的分组是“内部重要的”，即对于这些分组，多校准保证并不像对高风险主题那样具有社会关键性，但将其纳入多校准过程可以轻松且显著地提高最终可靠性预测器的预测准确性。在实证上，我们将该框架应用于多种基准上的幻觉预测和不安全回答检测，并发现将 CoT 分组加入基于主题的校准会在两项任务中降低 Brier 分数。有趣的是，我们发现仅针对 CoT 分组进行多校准（不含任何主题信息）已经能够改善跨主题的校准，这表明预测器的错误取决于模型如何推理，而不是提示中问了什么。
cs.AR / 67 / 2610.26118
Superposed analog memory for vector storage in neural networks
用于神经网络中向量存储的叠加模拟存储器
large language model
大语言模型相关
Abstract
The growing memory demands of artificial neural networks pose a major challenge for energy-efficient computing, particularly for models that rely on large collections of high-dimensional weight and embedding vectors. Conventional memory architectures store these vectors individually, requiring many memory cells per vector and limiting achievable storage density. Here we demonstrate an analog memory architecture that encodes multiple vectors in a distributed superposition within a single memory array. Individual vectors are recovered through in-memory matrix-vector multiplication, allowing the degree of superposition to be tuned to balance storage density against recall fidelity. This approach introduces a new scaling dimension for analog memory, in which effective storage density can be increased without a corresponding increase in the number of physical memory cells. The distributed representation also provides resilience to device-level noise. We implement the architecture on a prototype phase-change memory chip and experimentally evaluate it using embedding tables from deep learning recommendation models and large language models. The results demonstrate inference accuracy comparable to full-precision software baselines, establishing analog superposition as a promising route towards dense, energy-efficient memory for artificial intelligence.
Chinese Translation
人工神经网络不断增长的存储需求对高能效计算构成了重大挑战，尤其是对于那些依赖大量高维权重向量和嵌入向量的模型而言。传统存储架构将这些向量单独存储，每个向量需要许多存储单元，从而限制了可实现的存储密度。在此，我们展示了一种模拟存储架构，该架构在单个存储阵列内以分布式叠加的方式编码多个向量。各个向量通过存内矩阵-向量乘法被恢复，使得可以调节叠加程度，以在存储密度与召回保真度之间取得平衡。这种方法为模拟存储器引入了一个新的缩放维度，在该维度中，可以在不相应增加物理存储单元数量的情况下提高有效存储密度。这种分布式表示还提供了对器件级噪声的鲁棒性。我们在原型相变存储器芯片上实现了该架构，并使用来自深度学习推荐模型和大型语言模型的嵌入表对其进行了实验评估。结果表明，其推理准确性与全精度软件基线相当，从而确立了模拟叠加作为一种有前景的途径，可用于构建面向人工智能的密集、高能效存储器。
cs.LG / 68 / 2610.25662
Knowledge Manifold Optimization for Scientific Discovery
面向科学发现的知识流形优化
large language model
大语言模型相关
Abstract
The rapid expansion and specialization of scientific literature make it increasingly difficult to identify not only what is known, but where new and still supportable knowledge may emerge. We introduce Knowledge Manifold Optimization (KMO), a discovery-oriented extension of the Knowledge Manifold framework that treats scientific discovery as a multiobjective search over a continuous, literature-derived knowledge space. Candidate locations are evaluated using four complementary objectives: persistent knowledge-regime boundary strength, band-pass white-space novelty, semanticgradient magnitude, and evidential support. Rather than collapsing these criteria into a single weighted score, KMO identifies Pareto-optimal Supported Knowledge Gaps, defined as corpus-relative unoccupied regions that remain sufficiently supported by surrounding literature. The method is demonstrated on a 99-paper corpus in aerodynamics and aerodynamic design. On a 21 x 21 grid, exact Pareto analysis produced 19 non-dominated points, of which 16 formed the practical front after excluding points with zero white-space score. These solutions exhibited distinct discovery modes, including methodology bridges, aero-structural topology/FSI integration, functional aerodynamic-device integration, and rear-wake/vehicle flow-control integration. A distance-maximization baseline preferentially selected extrapolative regions with substantially lower support and reliability, demonstrating why emptiness alone is insufficient. Representative Pareto locations were converted into auditable evidence packages and then verbalized by a constrained large language model into candidate research hypotheses. The results show that scientific discovery can be formulated as structured multi-objective search in an explicit knowledge space, while keeping novelty corpus-relative and separating numerical discovery from downstream language generation.
Chinese Translation
科学文献的迅速扩张与专业化，使人们不仅越来越难以识别已知的内容，也越来越难以识别新的、且仍然可被支持的知识可能在何处涌现。我们引入知识流形优化（Knowledge Manifold Optimization, KMO），这是知识流形框架的一个面向发现的扩展，它将科学发现视为在连续的、由文献导出的知识空间上进行的多目标搜索。候选位置通过四个互补目标进行评估：持久知识体制边界强度、带通空白空间新颖性、语义梯度幅值以及证据支持。KMO 并不将这些标准压缩为单一加权分数，而是识别帕累托最优的受支持知识缺口（Supported Knowledge Gaps），其定义为相对于语料库未被占据、但仍得到周围文献充分支持的区域。该方法在空气动力学与气动设计领域的 99 篇论文语料库上进行了演示。在 21 x 21 网格上，精确帕累托分析产生了 19 个非支配点，其中在排除空白空间得分为零的点后，有 16 个构成了实际前沿。这些解表现出不同的发现模式，包括方法学桥梁、气动-结构拓扑/FSI 集成、功能性气动装置集成以及后尾流/车辆流动控制集成。一个距离最大化基线优先选择了外推区域，这些区域的支持度和可靠性显著较低，从而表明仅有空白性是不够的。具有代表性的帕累托位置被转换为可审计的证据包，然后由受约束的大语言模型表述为候选研究假设。结果表明，科学发现可以被表述为显式知识空间中的结构化多目标搜索，同时使新颖性保持为相对于语料库的，并将数值发现与下游语言生成分离。
cs.LG / 69 / 2610.25447
Enhanced Masked Diffusion Unlearning
增强型掩码扩散遗忘
diffusion
扩散模型相关
Abstract
Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens with bidirectional context. As these models reach real-world deployment, machine unlearning becomes essential. This process aims to remove the influence of a designated "forget set" without retraining from scratch, so that sensitive, copyrighted, or restricted content can no longer be elicited while preserving retained capabilities. Masked Diffusion Unlearning (MDU) is the first publicly available unlearning method customized for these models: it tunes the model $p_θ$ to match a frozen prompt-masked anchor $p_{\mathrm{anchor}}$ on the forget set by minimizing the anchor-last divergence $\mathrm{KL}(p_θ\Vert p_{\mathrm{anchor}})$. We identify two failure modes of MDU. First, random answer masking can leave key forget tokens visible to the anchor even when the question is fully masked. Second, at a memorized (near-one-hot) prediction, the anchor-last logit gradient vanishes while the anchor-first divergence $\mathrm{KL}(p_{\mathrm{anchor}} \Vert p_θ)$ logit gradient does not. As a result, MDU training stalls on the tokens we most need to forget. Backed by theoretical and empirical analyses, we propose Enhanced Masked Diffusion Unlearning (EMDU): anchor-first KL divergence toward a frozen anchor whose entire answer span is masked. On TOFU and RWKU benchmarks using LLaDA-8B and Dream-7B, our EMDU matches or exceeds the forgetting of MDU and adapted AR baselines, while better preserving retain utility.
Chinese Translation
掩码扩散语言模型（MDLMs）通过双向上下文迭代地解除令牌掩码来生成文本。随着这些模型进入真实世界部署，机器遗忘变得至关重要。该过程旨在在不从头重新训练的情况下移除指定“遗忘集”的影响，从而使敏感的、受版权保护的或受限制的内容不再能够被诱导生成，同时保留模型原有的保留能力。掩码扩散遗忘（MDU）是首个公开可用的、为这类模型定制的遗忘方法：它通过最小化锚点在后散度 $\mathrm{KL}(p_θ\Vert p_{\mathrm{anchor}})$，在遗忘集上调整模型 $p_θ$ 以匹配一个冻结的、提示被掩码的锚点 $p_{\mathrm{anchor}}$。我们识别出 MDU 的两种失效模式。第一，随机答案掩码可能在问题被完全掩码时仍使关键的遗忘令牌对锚点可见。第二，在记忆化（近 one-hot）预测处，锚点在后（anchor-last）的 logit 梯度消失，而锚点在前散度 $\mathrm{KL}(p_{\mathrm{anchor}} \Vert p_θ)$ 的 logit 梯度并不消失。因此，MDU 训练在我们最需要遗忘的令牌上停滞。在理论分析与实证分析的支持下，我们提出增强型掩码扩散遗忘（EMDU）：面向一个其整个答案跨度均被掩码的冻结锚点，采用锚点在前（anchor-first）的 KL 散度。在使用 LLaDA-8B 和 Dream-7B 的 TOFU 与 RWKU 基准上，我们的 EMDU 达到或超过了 MDU 以及经适配的自回归（AR）基线的遗忘效果，同时更好地保持了保留效用。
cs.LG / 70 / 2610.25480
Consistent LLM Call Handling Across Java Frameworks: A Reference Implementation and Benchmark
跨 Java 框架的一致 LLM 调用处理：参考实现与基准测试
large language model
大语言模型相关
Abstract
Java applications reach large language models (LLMs) through an integration framework, and rely on it to handle a failing provider: to retry what can be retried and nothing else, to wait as long as the provider asks, never to present a truncated answer as complete, and to record each failure in the application's metrics. We state these expectations as a six-clause reliability contract and measure Spring AI 2.0.1 and LangChain4j 1.21.0 against it with a scripted OpenAI-compatible provider: fifteen failure scenarios, synchronous and streaming, three repetitions. LangChain4j meets it in 47 of 77 applicable cases with its defaults and in 51 of 83 when configured explicitly, Spring AI in 72 of 77 and 76 of 83; both deliver a stream that ended without its terminator as a complete answer. We then place one implementation of the contract beneath both frameworks, at the HTTP boundary - an OkHttp interceptor for Spring AI, an HttpClient for LangChain4j, and a java.net.http transport for code without a framework - sharing a 725-line core that depends only on the JDK and the Micrometer Observation API. With it, all three clients behave identically and meet the contract in 90 of 91 cases; the remaining case is a policy trade-off between retrying a hung provider and waiting for a slow first token. Against a local model server the layer rejected none of 700 calls, and it adds a median of 32-217 microseconds to a call.
Chinese Translation
Java 应用通过集成框架访问大语言模型（LLM），并依赖该框架来应对提供商故障：只重试可以重试的内容而不重试其他内容，按提供商要求的时间等待，绝不把被截断的答案当作完整答案呈现，并把每次失败记录到应用的指标中。我们将这些期望表述为一份包含六项条款的可靠性契约，并使用一个脚本化的、兼容 OpenAI 的提供商，以十五个故障场景、同步与流式两种模式、三次重复，来衡量 Spring AI 2.0.1 和 LangChain4j 1.21.0 对该契约的符合程度。LangChain4j 在默认配置下于 77 个适用案例中的 47 个满足该契约，在显式配置后于 83 个案例中的 51 个满足；Spring AI 则分别为 77 个中的 72 个和 83 个中的 76 个；两者都会把未收到终止符便结束的流当作完整答案交付。随后，我们将该契约的一个实现置于两个框架之下、位于 HTTP 边界处——为 Spring AI 提供 OkHttp 拦截器，为 LangChain4j 提供 HttpClient，为不使用框架的代码提供 java.net.http 传输层——它们共享一个 725 行的核心，该核心仅依赖 JDK 和 Micrometer Observation API。借助该实现，三个客户端的行为完全一致，并在 91 个案例中的 90 个满足契约；剩下的一个案例是在重试挂起的提供商与等待缓慢的首个 token 之间的一项策略权衡。在针对本地模型服务器的测试中，该层未拒绝 700 次调用中的任何一次，并且它为每次调用增加的中位开销为 32–217 微秒。
cs.LG / 71 / 2610.25495
Probabilistic Limit Order Book Forecasting with Conditional Diffusion Models
基于条件扩散模型的概率限价订单簿预测
diffusion
扩散模型相关
Abstract
Modeling limit order book (LOB) dynamics is central to market forecasting and execution-strategy evaluation. While diffusion models achieve strong empirical performance, existing LOB snapshot models typically lack a learned recurrent encoder and formal statistical guarantees. We introduce LOBGrad, a conditional DDPM framework for probabilistic LOB forecasting that combines an autoregressive RNN encoder with a diffusion-based generator. In particular, we establish an excess-risk oracle inequality for the one-step DDPM noise-prediction objective under exponential mixing and regularity assumptions. We further show that, under contractive dynamics, the dependence on the conditioning-window length can be reduced to a dependence on the latent-state dimension. Experiments on BTC/USDT, GOOGL, and META show that LOBGrad improves on its encoder-free diffusion ablation across all reported queue metrics and obtains competitive results against other baselines. It also reproduces the key stylized facts of LOB microstructure, and yields reasonably calibrated queue-based predictive intervals. A stylized downstream task illustrates the potential value of the generated forecasts.
Chinese Translation
对限价订单簿（LOB）动态进行建模是市场预测与执行策略评估的核心。尽管扩散模型取得了很强的实证性能，但现有的 LOB 快照模型通常缺乏可学习的循环编码器以及形式化的统计保证。我们提出 LOBGrad，一个用于概率性 LOB 预测的条件 DDPM 框架，它将自回归 RNN 编码器与基于扩散的生成器相结合。特别地，我们在指数混合与正则性假设下，为一步 DDPM 噪声预测目标建立了一个超额风险 oracle 不等式。我们进一步表明，在收缩动力学下，对条件窗口长度的依赖可以约化为对潜在状态维度的依赖。在 BTC/USDT、GOOGL 和 META 上的实验表明，LOBGrad 在所有报告的队列指标上都优于其无编码器的扩散消融版本，并在与其他基线方法比较时取得了有竞争力的结果。它还重现了 LOB 微观结构的关键典型事实，并给出了校准合理的基于队列的预测区间。一个简化的下游任务说明了所生成预测的潜在价值。
cs.LG / 72 / 2610.25502
Steered Not Stirred: Effective Reward Tilting using Surrogate Reward Mixing
被引导而非被搅动：使用代理奖励混合的有效奖励倾斜
diffusion
扩散模型相关
Abstract
Continuous-time generative models, from diffusion models, flow matching, and now flow maps, owe their practical value to high-fidelity generation and steerability. In many applications, a ground-truth latent ``oracle'' reward $R^*$ is unavailable or expensive to evaluate, e.g., human judgments or experimental measurements. Although the oracle reward is not directly observable, it can often be approximated through multiple cheaper surrogates that, individually, provide incomplete characterizations but collectively reveal substantial information about the downstream task. In this paper, we present a method for composing such surrogates to steer a pretrained model toward the distribution tilted by an unavailable $R^*$. Given a small feedback set of oracle rankings or numerical rewards, we introduce Mixing And Reward Tilting for INference-time Improvement (MARTINI), which learns a surrogate composition for plug-and-play use with standard inference-time steering methods. Our theory characterizes surrogate information limits and relates population recovery error to ideal-target KL at small tilt strength. For preferences, the bound additionally accounts for an irreducible mismatch between the Plackett-Luce population target and the conditional mean. Experiments demonstrate improved oracle-tilt fidelity in language generation and ImageNet alignment, and improved joint pose accuracy and validity in protein-ligand cofolding, without inference-time oracle queries.
Chinese Translation
连续时间生成模型——从扩散模型、流匹配到如今的流映射——其实际价值归功于高保真生成和可引导性。在许多应用中，真值潜在“oracle”奖励 $R^*$ 不可获得或评估代价高昂，例如人类判断或实验测量。尽管 oracle 奖励不能直接观测，但通常可以通过多个更廉价的代理来近似；这些代理各自提供不完整的刻画，但合在一起能揭示关于下游任务的大量信息。本文提出一种方法，用于组合此类代理，以将预训练模型引导至由不可获得的 $R^*$ 所倾斜的分布。给定由 oracle 排名或数值奖励组成的小型反馈集，我们引入 Mixing And Reward Tilting for INference-time Improvement (MARTINI)，它学习一种代理组合，以便与标准推理时引导方法即插即用。我们的理论刻画了代理信息极限，并在小倾斜强度下将总体恢复误差与理想目标 KL 联系起来。对于偏好，该界额外考虑了 Plackett-Luce 总体目标与条件均值之间不可约的失配。实验表明，在语言生成和 ImageNet 对齐中提高了 oracle 倾斜保真度，在蛋白质-配体共折叠中提高了联合位姿准确性和有效性，且无需推理时 oracle 查询。
cs.LG / 73 / 2610.25519
ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference
ATEST：面向LLM推理的自适应资源感知测试时扩展
large language model
大语言模型相关
Abstract
Test-time scaling (TTS) improves large language model (LLM) reasoning by sampling multiple answers, but fixed sampling budgets ignore both query difficulty and the changing energy cost of inference under GPU contention. We measure up to a 4.1-fold increase in energy for the same four-sample query as GPU load rises. We present ATEST, a resource-aware TTS framework that jointly considers the expected value of additional samples and their current GPU cost. ATEST-Online, its default mode, requires no offline calibration: it incrementally generates samples, estimates the stability of the plurality answer, monitors live GPU conditions, and learns step energy cost online to decide when to stop. ATEST-Calibrated instead selects a per-query budget using offline accuracy and cost models, favoring deployments with stable loads. We evaluate both modes on four reasoning and question-answering datasets using Phi-3-mini on an NVIDIA A100 under controlled varying load, and test transfer to an NVIDIA L40S and Qwen2.5-3B. Across the four A100 datasets, ATEST-Online answers every query without exceeding the energy budget, while achieving the highest or statistically tied accuracy and reducing sampling as load increases. On GSM8K at high load, it achieves 60.0% accuracy with no deferrals, versus 44.0% accuracy and 42.4% deferrals for ATEST-Calibrated. Ablations show that online cost learning drives adaptation to load. Including offline profiling, ATEST-Online uses 42 times less GPU time and 48 times less energy than ATEST-Calibrated for the evaluated workload. Online control trades higher tail latency and lower throughput on short tasks for robustness under varying load; offline calibration remains competitive under stable conditions.
Chinese Translation
测试时扩展（TTS）通过采样多个答案来提升大语言模型（LLM）的推理能力，但固定的采样预算既忽略了查询难度，也忽略了在GPU资源争用下推理能量成本的变化。我们测量到，随着GPU负载升高，同一个四样本查询的能量消耗最高增加4.1倍。我们提出ATEST，一个资源感知的TTS框架，它联合考虑额外样本的期望价值及其当前GPU成本。ATEST-Online是其默认模式，无需离线校准：它增量生成样本，估计多数答案的稳定性，监测实时GPU状况，并在线学习单步能量成本以决定何时停止。ATEST-Calibrated则利用离线准确率和成本模型为每个查询选择预算，更适合负载稳定的部署场景。我们在受控变化负载下，使用NVIDIA A100上的Phi-3-mini，在四个推理和问答数据集上评估了两种模式，并测试了向NVIDIA L40S和Qwen2.5-3B的迁移。在四个A100数据集上，ATEST-Online在不超出能量预算的情况下回答了每个查询，同时达到最高或统计上并列的准确率，并随着负载增加减少采样。在高负载下的GSM8K上，它以无推迟回答达到60.0%的准确率，而ATEST-Calibrated为44.0%的准确率和42.4%的推迟率。消融实验表明，在线成本学习驱动了对负载的适应。计入离线性能分析，对于所评估的工作负载，ATEST-Online使用的GPU时间比ATEST-Calibrated少42倍，能量少48倍。在线控制以短任务上更高的尾延迟和更低的吞吐量为代价，换取在变化负载下的鲁棒性；离线校准在稳定条件下仍具有竞争力。
cs.LG / 74 / 2610.25588
Divergence based Ranked Preference Alignment for LLMs
面向大语言模型的基于散度的排序偏好对齐
large language model
大语言模型相关
Abstract
Alignment of large language models typically relies on pairwise preference feedback, though ranked feedback can capture more comprehensive information across multiple responses within a single annotation. However, ranked-preference alignment is underexplored in practice compared to simpler pairwise approaches, potentially resulting in suboptimal behavior. Concurrently, recent research has conceptualized large language model alignment objectives as divergence estimators between aligned (preferred) and unaligned (less-preferred) response distributions, resulting in a family of objectives with an interpretable theoretical foundation and strong empirical performance. This work unifies these perspectives by introducing $f$-RPO, a $f$-divergence objective for ranked preference optimization that extends divergence-based alignment to ranked feedback and recovers pairwise preference alignment as a special case. Our theoretical analysis establishes rank-order consistency, showing that $f$-RPO preserves the underlying preference ordering at convergence. Empirical results indicate that total-variation and reverse-KL variants of the proposed objective consistently outperform ranked-alignment baselines in both ranking and generation quality across Qwen, Llama, & Gemma models ranging from 0.5B to 8B parameters.
Chinese Translation
大语言模型的对齐通常依赖于成对偏好反馈，而排序反馈能够在单次标注中捕获多个回复之间更全面的信息。然而，与更简单的成对方法相比，排序偏好对齐在实践中尚未得到充分探索，这可能导致次优行为。与此同时，近期研究已将大语言模型的对齐目标概念化为对齐（偏好）回复分布与未对齐（次偏好）回复分布之间的散度估计器，从而形成了一类具有可解释理论基础和强实证性能的目标。本工作通过引入 $f$-RPO 来统一这些视角；$f$-RPO 是一种用于排序偏好优化的 $f$-散度目标，它将基于散度的对齐扩展到排序反馈，并将成对偏好对齐作为特例恢复。我们的理论分析建立了秩序一致性，表明 $f$-RPO 在收敛时保持底层的偏好排序。实证结果表明，所提出目标的全变差和反向 KL 变体在从 0.5B 到 8B 参数的 Qwen、Llama 和 Gemma 模型上，在排序和生成质量两方面都始终优于排序对齐基线。
cs.LG / 75 / 2610.25679
LOOP: Recovery-Aligned Iterative Layer Pruning via On-Policy Distillation
LOOP：基于同策略蒸馏的恢复对齐迭代式层剪枝
large language model
大语言模型相关
Abstract
Layer-level pruning is attractive for large language model (LLM) deployment as it reduces the model size while preserving a dense architecture that can be efficiently executed on standard hardware. Current methods typically follow a one-shot paradigm, which first removes selected layers to reach a target model size and then performs additional training to recover model capabilities. However, pruning directly to the target size can substantially degrade reasoning capabilities, making subsequent recovery more difficult under a limited training budget. Moreover, this paradigm separates layer selection from recovery, limiting the use of recovery information to guide selection and potentially hindering the recovery of reasoning capabilities. To address these limitations, we propose LOOP, a Layer-level Recovery-Oriented On-Policy Pruning framework that interleaves pruning and recovery across multiple shots. In each shot, LOOP removes a single layer and performs short on-policy distillation, mitigating the abrupt degradation in reasoning capabilities associated with one-shot pruning. To guide layer selection in subsequent shots, LOOP reuses on-policy trajectories and optimization signals from recovery to jointly assess each remaining layer's reasoning contribution and recovery sensitivity, thereby making layer selection recovery-aware. Evaluations on several benchmarks show that LOOP consistently outperforms one-shot baselines in average accuracy across different model scales and pruning ratios.
Chinese Translation
层级别剪枝对于大语言模型（LLM）部署具有吸引力，因为它能在减小模型规模的同时保留可在标准硬件上高效执行的稠密架构。当前方法通常遵循一步式范式，即首先移除选定的层以达到目标模型规模，然后执行额外训练以恢复模型能力。然而，直接剪枝到目标规模会显著降低推理能力，使得在有限训练预算下的后续恢复更加困难。此外，该范式将层选择与恢复分离，限制了利用恢复信息来指导选择，并可能阻碍推理能力的恢复。为解决这些局限，我们提出 LOOP，一种面向恢复的层级别同策略剪枝框架，它在多个轮次中交错进行剪枝与恢复。在每一轮中，LOOP 移除单个层并执行简短的同策略蒸馏，从而缓解与一步式剪枝相关的推理能力骤降。为了在后续轮次中指导层选择，LOOP 复用来自恢复过程的同策略轨迹和优化信号，以联合评估每个剩余层的推理贡献和恢复敏感性，从而使层选择具备恢复感知能力。在多个基准上的评估表明，在不同模型规模和剪枝比例下，LOOP 在平均准确率上始终优于一步式基线。
cs.LG / 76 / 2610.25692
Progress-Guided Preference Generation for Aligning Generative Control Policies
用于对齐生成式控制策略的进度引导偏好生成
diffusion
扩散模型相关
Abstract
Generative policies have achieved strong performance in imitation learning tasks, but their success is often constrained by the limited availability of expert demonstrations. Although preference-based fine-tuning can leverage suboptimal rollouts to improve performance, existing approaches heavily rely on expensive human annotations. In this work, we introduce Progress-Guided Preference Generation (PGPG), an unsupervised preference alignment framework that eliminates the need for human-in-the-loop supervision. Our method leverages ordered coverage alignment to measure sequential progress relative to a small set of expert demonstrations, enabling the automatic extraction of pseudo-preference labels from autonomously collected rollouts. Furthermore, we introduce key enhancements to improve the reliability of these progress-guided preferences, enabling effective fine-tuning of pretrained generative policies without requiring manually designed reward functions. Extensive experiments across a suite of challenging long-horizon manipulation tasks demonstrate that PGPG outperforms standard supervised fine-tuning baselines. Moreover, our conceptually simple framework is broadly applicable across multiple classes of generative policies, demonstrating strong versatility when integrated with both diffusion and flow-matching policies.
Chinese Translation
生成式策略在模仿学习任务中取得了强劲的性能，但其成功往往受限于专家演示的有限可用性。尽管基于偏好的微调能够利用次优轨迹来提升性能，但现有方法严重依赖昂贵的人工标注。在本工作中，我们提出进度引导偏好生成（PGPG），一种无需人在回路监督的无监督偏好对齐框架。我们的方法利用有序覆盖对齐来度量相对于一小组专家演示的顺序进度，从而能够从自主收集的轨迹中自动提取伪偏好标签。此外，我们引入关键增强以提高这些进度引导偏好的可靠性，使得无需人工设计奖励函数即可有效微调预训练的生成式策略。在一系列具有挑战性的长时程操作任务上的大量实验表明，PGPG 优于标准的监督微调基线。此外，我们概念上简单的框架广泛适用于多类生成式策略，在与扩散策略和流匹配策略集成时展现出很强的通用性。
cs.LG / 77 / 2610.25723
Learning to Couple: Context-Adaptive Joint Denoising for Diffusion Language Models
学习耦合：面向扩散语言模型的上下文自适应联合去噪
diffusion
扩散模型相关
Abstract
Diffusion language models accelerate generation by predicting multiple tokens in parallel, but factorized denoisers treat simultaneous predictions as conditionally independent. Decoding heuristics reduce the resulting errors by controlling which tokens are committed, but leave the underlying distribution factorized. Joint modeling addresses the source of the problem, yet existing methods attach separately learned layers to frozen denoisers, leaving the backbone unadapted to the joint objective and the coupling parameters fixed across contexts. We propose Codpula, a context-parameterized denoiser combining a Transformer with a tractable probabilistic inference layer. Given a partially masked input, the Transformer produces token potentials that shape per-position predictions and compact gates that modulate the layer's coupling parameters, yielding a context-dependent joint distribution while retaining exact conditional inference. We jointly optimize both components under a conditional denoising objective using a dual-flow multiplicative update that preserves valid probabilistic parameters. At inference, Codpula supports standard heuristics and uses conditional queries to adapt commitments to dependencies among proposed tokens without extra Transformer evaluations within a step. Codpula sets new state-of-the-art results across a broad suite of reasoning benchmarks while advancing the accuracy-latency Pareto frontier in mathematical reasoning and code generation.
Chinese Translation
扩散语言模型通过并行预测多个词元来加速生成，但因子化的去噪器将同时进行的预测视为条件独立的。解码启发式方法通过控制哪些词元被提交来减少由此产生的错误，但使底层分布仍保持因子化。联合建模针对的是问题的根源，然而现有方法将单独学习的层附加到冻结的去噪器上，使主干网络未能针对联合目标进行适配，且耦合参数在不同上下文中保持固定。我们提出 Codpula，一种上下文参数化的去噪器，将 Transformer 与一个可处理的概率推断层相结合。给定部分掩码的输入，Transformer 产生塑造逐位置预测的词元势，以及调制该层耦合参数的紧凑门控，从而在保留精确条件推断的同时产生依赖于上下文的联合分布。我们在条件去噪目标下，使用保持有效概率参数的双流乘性更新来联合优化这两个组件。在推断时，Codpula 支持标准启发式方法，并使用条件查询使提交决策适应所提议词元之间的依赖关系，而无需在单步内进行额外的 Transformer 评估。Codpula 在广泛的推理基准套件上创下新的最先进结果，同时推进了数学推理与代码生成中的准确率-延迟帕累托前沿。
cs.LG / 78 / 2610.25725
Anatomy of Off-Policy Policy Gradient: Importance Sampling, KL Regularization, and Baselines
离策略策略梯度的剖析：重要性采样、KL 正则化与基线
large language model
大语言模型相关
Abstract
Reinforcement learning has become a central tool for large language model (LLM) post-training, where policy gradient methods are routinely deployed off-policy, even though vanilla policy gradient assumes on-policy sampling. We study when such off-policy deployment is justified, and what roles its core ingredients, including importance sampling, KL regularization, and baselines, play in handling the resulting distribution shift. We show that for a general class of off-policy policy gradient updates, consistency, in the sense that the optimal policy remains a stationary point, requires one of two structures at the optimal policy: trajectory-level importance weighting, which suffers from the well-known curse of horizon, or KL regularization with a baseline equal to the optimal soft value, instantiated by the grouped mean. Focusing on the KL route, we identify it with the group-relative policy gradient (GRPG) update, establish an equivalence to a trajectory-level Bellman residual minimization objective, and obtain a finite-sample regret bound under all-policy coverage. This also gives a theoretical account of a variant of the empirically successful GRPO objective. We further uncover a hidden role of baselines in the off-policy setting: a constant shift in the baseline implements pessimism, the standard remedy in offline RL for partial data coverage. We instantiate this via the asymmetric REINFORCE objective, and connect it to the game-theoretic offline RL framework, showing a regret bound with single-policy coverage. Together, these results offer a unified theoretical view of off-policy policy gradient in LLM post-training and connect it to classical ideas in offline RL.
Chinese Translation
强化学习已成为大语言模型（LLM）后训练的核心工具，在其中策略梯度方法常以离策略方式部署，尽管原始策略梯度假设同策略采样。我们研究这种离策略部署在何时是有依据的，以及其核心组成部分——包括重要性采样、KL 正则化和基线——在处理由此产生的分布偏移时发挥什么作用。我们证明，对于一般的一类离策略策略梯度更新，一致性（即最优策略仍是一个驻点）要求在最优策略处具备以下两种结构之一：轨迹级重要性加权，但它受制于众所周知的时域诅咒；或者带有等于最优软价值的基线的 KL 正则化，并由分组均值具体实现。聚焦于 KL 路径，我们将其识别为组相对策略梯度（GRPG）更新，建立其与轨迹级 Bellman 残差最小化目标的等价性，并在全策略覆盖下获得有限样本遗憾界。这也为经验上成功的 GRPO 目标的一个变体提供了理论解释。我们进一步揭示了基线在离策略设定中的一个隐藏作用：基线的常数平移实现了悲观主义，这是离线 RL 中针对部分数据覆盖的标准补救措施。我们通过非对称 REINFORCE 目标将其具体化，并将其与博弈论离线 RL 框架联系起来，展示了在单策略覆盖下的遗憾界。总之，这些结果提供了 LLM 后训练中离策略策略梯度的统一理论视角，并将其与离线 RL 中的经典思想联系起来。
cs.LG / 79 / 2610.25739
How Much Does One Wrong Answer Cost? Error Propagation in Constraint Pruning with Noisy LLM Oracles
一个错误答案的代价有多大？带噪声LLM预言机的约束剪枝中的错误传播
large language model
大语言模型相关
Abstract
Constraint pruning saves queries to a large language model (LLM) or another noisy oracle by skipping each yes/no question implied by answers accepted earlier, for example through transitivity. Entity-resolution systems and LLM-based rankers use this idea. However, their analyses typically assume a correct oracle. We argue that pruning should be judged by its risk when the oracle errs, since an accepted wrong answer spreads to every question it implies. When each answer is wrong independently with small probability $\varepsilon$, the first-order risk is set by how many facts one wrong answer corrupts on average. We consider two kinds of constraints. For implication rules, as in taxonomies, pruning in a random order adds no first-order risk: a wrong answer that would change $k$ facts is accepted with probability $1/k$, so it costs one fact on average. Rules with $s$ premises break this balance: a wrong answer at a premise shared by $r$ rules can cost order $r^{1-1/s}$ facts. For entity resolution with $m$ records of one entity, the first-order risk equals $\varepsilon$ times the Wiener index (the total pairwise distance) of the spanning tree built by the query order. Pivot orders, which first compare one record with all others, amplify errors by less than 2 and are optimal; random orders by order $m^{1/3}$; the common most-similar-first order is the worst in one dimension and, for large $m$, worse than random in two. With several entities, the cost of wrongly merging two does not depend on how many others exist. Simulations agree with every predicted constant within sampling error.
Chinese Translation
约束剪枝通过跳过由早先接受的答案所蕴含的每个是/否问题（例如通过传递性），节省对大型语言模型（LLM）或其他带噪声预言机的查询。实体解析系统和基于LLM的排序器使用了这一思想。然而，它们的分析通常假设预言机是正确的。我们认为，当预言机出错时，剪枝应以其风险来评判，因为一个被接受的错误答案会传播到它所蕴含的每个问题。当每个答案以较小概率 $\varepsilon$ 独立出错时，一阶风险由平均一个错误答案损坏多少个事实决定。我们考虑两类约束。对于蕴含规则，如在分类体系中那样，以随机顺序剪枝不会增加一阶风险：一个本会改变 $k$ 个事实的错误答案以概率 $1/k$ 被接受，因此它平均只代价一个事实。具有 $s$ 个前提的规则打破了这种平衡：在一个被 $r$ 条规则共享的前提处的一个错误答案可能造成数量级为 $r^{1-1/s}$ 个事实的代价。对于具有一个实体的 $m$ 条记录的实体解析，一阶风险等于由查询顺序构建的生成树的 Wiener 指数（两两距离之和）乘以 $\varepsilon$。枢轴顺序首先将一条记录与所有其他记录比较，将误差放大不到2倍，并且是最优的；随机顺序放大量级为 $m^{1/3}$；常见的最相似者优先顺序在一维上最差，并且对于大的 $m$，在二维上比随机顺序更差。在存在多个实体时，错误合并两个实体的代价不依赖于存在多少个其他实体。模拟在采样误差范围内与每个预测常数一致。
cs.LG / 80 / 2610.25751
NADS: Navigator-Guided Data Selection for Mitigating Catastrophic Forgetting in Fine-Tuning
NADS：导航器引导的数据选择，用于缓解微调中的灾难性遗忘
large language model
大语言模型相关
Abstract
Catastrophic forgetting is a central obstacle in single-stage fine-tuning of large language models (LLMs) on a new task, especially when the original pretraining data are unavailable. Synthetic data offers a practical mitigation path, but candidate pools are often noisy and redundant, making effective selection essential. Existing methods often rely on indirect proxy signals whose alignment with actual training benefit can be limited, while utility--diversity selection over large candidate pools remains computationally demanding. To address these challenges, we propose NAvigator-guided Data Selection (NADS), a fine-tuning framework for LLMs. NADS first follows the target fine-tuning recipe without preservation regularization to obtain a navigator model, whose deviation from the pretrained model exposes recipe-induced capability drift. It then uses the predictive divergence between the navigator and the pretrained model to score forgetting-aware utility for each synthetic candidate, and constructs a constraint set through an efficient utility--diversity selection strategy. Finally, NADS fine-tunes on the new-task data while distilling pretrained behavior on the selected constraints to preserve general capabilities. Experiments show that NADS delivers a stronger balance between new-task performance and general capability preservation, while reducing utility--diversity selection overhead. Code at: https://github.com/zjh-vinky/NADS.
Chinese Translation
灾难性遗忘是在新任务上对大型语言模型（LLMs）进行单阶段微调时的核心障碍，尤其是在原始预训练数据不可用时。合成数据提供了一条实用的缓解路径，但候选池往往噪声大且冗余，因此有效选择至关重要。现有方法通常依赖间接代理信号，而这些信号与实际训练收益的一致性可能有限；同时，在大型候选池上进行效用--多样性选择仍然计算开销很大。为应对这些挑战，我们提出了导航器引导的数据选择（NADS），一种用于 LLMs 的微调框架。NADS 首先在不使用保持正则化的情况下遵循目标微调配方，以获得一个导航器模型，该模型与预训练模型的偏差揭示了配方引发的能力漂移。然后，它利用导航器与预训练模型之间的预测散度，为每个合成候选样本评分遗忘感知效用，并通过高效的效用--多样性选择策略构建约束集。最后，NADS 在新任务数据上进行微调，同时在选定的约束上蒸馏预训练行为，以保持通用能力。实验表明，NADS 在新任务性能与通用能力保持之间实现了更强的平衡，同时降低了效用--多样性选择开销。代码见：https://github.com/zjh-vinky/NADS。
cs.LG / 81 / 2610.25844
SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation
SWIFT：通过目标外推的自引导弱到强扩散模型
diffusion
扩散模型相关
Abstract
Training high-quality diffusion models is computationally expensive, and inference-time guidance often adds further sampling cost. In this work, we introduce Swift, a simple yet effective self-guided training method that uses a frozen weak denoiser to improve a stronger model without adding inference overhead. For each noisy input, Swift extrapolates the clean regression target away from the weak denoiser's prediction, amplifying components that remain underfit. The weak guide can be an early training checkpoint, requires no backpropagation during main-model training, and is discarded afterward. The resulting model retains the original sampler and supports both conditional and unconditional generation without label dropout or an unconditional branch. We explain this mechanism through a spectral analysis of finite-time population gradient flow in an RKHS. With an early-stopped weak guide, we prove that a nonempty range of extrapolation strengths yields strictly lower posterior-mean error than matched standard training, despite shifting the regression target away from the posterior mean. Experiments on CIFAR-10 and ImageNet-256 demonstrate improvements across model scales and representations. In one ImageNet-256 configuration, Swift surpasses the baseline's best FID using roughly half the total training compute, including guide preparation and training overhead. On RAEv2, it achieves FID 1.15 without inference-time guidance and 0.99 with guidance. These results establish target extrapolation as an effective approach to weak-to-strong diffusion training with efficient single-model inference.
Chinese Translation
训练高质量扩散模型的计算成本很高，而推理时引导往往会进一步增加采样成本。在本工作中，我们提出 Swift，一种简单而有效的自引导训练方法，它使用一个冻结的弱去噪器来改进更强的模型，且不增加推理开销。对于每个噪声输入，Swift 将干净回归目标从弱去噪器的预测向外外推，从而放大仍然欠拟合的分量。弱引导可以是一个早期训练检查点，在主模型训练期间不需要反向传播，并且随后会被丢弃。所得模型保留原始采样器，并支持条件和无条件生成，而无需标签丢弃或无条件下分支。我们通过对 RKHS 中有限时间总体梯度流的谱分析来解释这一机制。在采用提前停止的弱引导时，我们证明：尽管将回归目标从后验均值移开，一个非空的外推强度范围仍能比匹配的标准训练产生严格更低的后验均值误差。在 CIFAR-10 和 ImageNet-256 上的实验表明，在不同模型规模和表示下均有改进。在一种 ImageNet-256 配置中，Swift 使用大约一半的总训练计算量（包括引导准备和训练开销）就超过了基线的最佳 FID。在 RAEv2 上，它在没有推理时引导的情况下达到 FID 1.15，在有引导的情况下达到 0.99。这些结果确立了目标外推作为一种有效方法，可用于具有高效单模型推理的弱到强扩散训练。
cs.LG / 82 / 2610.25932
Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding
让谱系保持存活：面向扩散 LLM 解码的多样性保持 SMC
diffusion
扩散模型相关
Abstract
Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the resampling step, which must concentrate computation on promising candidates while retaining enough diversity to explore alternative reasoning paths. The reward can be internal, the decoder's own confidence on the committed prefix, or external, a process reward model scoring that prefix. Under either reward, multinomial resampling can disrupt this balance by duplicating high-weight particles and discarding lineages that may still reach a correct solution. The effective sample size (ESS) does not register this loss: multinomial resampling resets it to its maximum even as distinct lineages disappear. To retain such lineages, we introduce lineage-preserving SMC (LP-SMC), which combines Chopthin resampling with an input weight floor. Instead of equalizing weights, Chopthin bounds the ratio between the largest and smallest weights it returns and carries the unequal weights forward. The floor limits how far below the maximum a weight may fall before resampling so that lineages with low early weight keep a chance to recover. Across mathematical-reasoning and code-generation benchmarks on two base models, LP-SMC raises oracle coverage, the fraction of problems with at least one correct particle, in every setting, by 1.8 to 25.6 points, and raises end-to-end accuracy under a common reward-model selector in eleven of twelve settings, by up to 7.9 points on LLaDA-1.5 and 18.3 on Dream-7B. The coverage gain persists under the external reward model, which suggests that diversity-preserving resampling and the guidance signal contribute separately to inference-time search.
Chinese Translation
推理时的序列蒙特卡洛（SMC）解码能够通过根据奖励对一组部分生成进行重采样，在不进行后训练的情况下提升掩码扩散语言模型（扩散 LLM）的推理准确率。因此，其有效性取决于重采样步骤：该步骤必须将计算集中到有希望的候选者上，同时保留足够的多样性以探索其他推理路径。该奖励可以是内部的，即解码器自身对已确定前缀的置信度，也可以是外部的，即过程奖励模型对该前缀进行评分。在任一奖励下，多项式重采样都可能通过复制高权重粒子并丢弃仍可能到达正确解的谱系，破坏这种平衡。有效样本量（ESS）无法反映这一损失：即使不同谱系消失，多项式重采样也会将其重置为最大值。为了保留此类谱系，我们引入了谱系保持 SMC（LP-SMC），它将 Chopthin 重采样与输入权重下限相结合。Chopthin 不是将权重均衡化，而是限制其返回的最大权重与最小权重之间的比值，并将这些不相等的权重向前传递。该下限限制了在重采样之前一个权重可以低于最大值的幅度，从而使早期权重较低的谱系保留恢复的机会。在两个基础模型上的数学推理与代码生成基准上，LP-SMC 在每种设置下都将 oracle 覆盖率——即至少有一个正确粒子的题目比例——提高 1.8 到 25.6 个百分点，并在十二种设置中的十一种下提高了在共同奖励模型选择器下的端到端准确率，在 LLaDA-1.5 上最高提高 7.9 个百分点，在 Dream-7B 上最高提高 18.3 个百分点。在外部奖励模型下，这种覆盖率增益依然存在，这表明多样性保持重采样和引导信号分别对推理时搜索作出贡献。
cs.LG / 83 / 2610.26309
CAST: Cast-Admissible Generative Design and Shape Optimization with Paired Depth Images
CAST：基于配对深度图像的铸造可脱模生成式设计与形状优化
diffusion
扩散模型相关
Abstract
Deep generative design can produce complex geometries, but unconstrained representations may create undercuts that prevent release from a two-part mold. We present CAST, which reformulates three-dimensional shape generation under a selected mold-opening constraint as a two-dimensional image generation problem using paired top and bottom depth images. Each valid depth pair defines one material interval along a selected mold-opening direction, so reconstructions exclude undercuts in that direction by construction. Diffusion models generate candidate shapes in this representation, and a finite element analysis (FEA) guided optimizer refines the same two boundaries. We evaluate diffusion-based generation with limited training data comprising 211 SimJEB bracket designs. On SimJEB brackets, FEA-guided refinement with a volume penalty reduced compliance in 193 of 200 generated candidates. The median reduction across all candidates was 36.8 percent in the optimization FEA model. Reconstruction of existing brackets revealed a mass--stiffness trade-off: filling occluded regions increased material volume and reduced displacement under the evaluated loads. Paired depth images thus connect generative exploration and structural refinement in a shared design space that preserves the selected pull-direction restriction.
Chinese Translation
深度生成式设计能够产生复杂几何形状，但无约束的表示可能产生倒扣，从而阻碍从两件式模具中脱模。我们提出 CAST，它将选定开模约束下的三维形状生成重新表述为使用成对顶部和底部深度图像的二维图像生成问题。每个有效深度对沿选定的开模方向定义一个材料区间，因此重建结果在构造上排除了该方向上的倒扣。扩散模型在该表示中生成候选形状，而有限元分析（FEA）引导的优化器细化相同的两个边界。我们使用包含 211 个 SimJEB 支架设计的有限训练数据评估基于扩散的生成。在 SimJEB 支架上，带有体积惩罚的 FEA 引导细化在 200 个生成候选中的 193 个上降低了柔度。在优化 FEA 模型中，所有候选的中位降低幅度为 36.8%。对现有支架的重建揭示了一种质量-刚度权衡：填充被遮挡区域增加了材料体积，并在所评估载荷下减小了位移。因此，配对深度图像在一个共享设计空间中将生成式探索与结构细化连接起来，该设计空间保留了选定的拉拔方向限制。
cs.LG / 84 / 2610.26443
Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation
通过权重空间重参数化重新思考LLM微调：在下游适配过程中保持安全性
large language model
大语言模型相关
Abstract
Fine-tuning large language models (LLMs) on downstream tasks often weakens their previously aligned safety behavior, creating an inherent trade-off between task adaptation and safety preservation. Existing approaches attempt to mitigate this issue by identifying safety-relevant neurons, layers, or update directions in the original parameter space. However, safety-related information is often distributed across many directions in this space, causing downstream updates to overwrite parameters that encode safety behavior. We propose WSR-Tune, a Weight Space Reparameterization-based fine-tuning framework that preserves safety during downstream adaptation. Our approach constructs a safety-conditioned basis and reparameterizes the weight matrices such that safety-relevant information becomes concentrated in a small subset of directions. We then identify safety-critical directions and freeze them in the reparameterized space, while updating the remaining complementary directions for downstream adaptation. By explicitly structuring the parameter space to disentangle safety and task information, our method reduces interference between objectives and enables simultaneous preservation of safety and improvement in downstream performance. Experimental results show that our method achieves a more favorable trade-off between downstream performance and safety retention, demonstrating its effectiveness for reliable LLM fine-tuning.
Chinese Translation
在下游任务上微调大型语言模型（LLM）往往会削弱其此前已对齐的安全行为，从而在任务适配与安全性保持之间形成一种内在权衡。现有方法试图通过在原始参数空间中识别与安全相关的神经元、层或更新方向来缓解这一问题。然而，安全相关信息通常分散在该空间的许多方向上，导致下游更新覆盖编码安全行为的参数。我们提出WSR-Tune，一种基于权重空间重参数化的微调框架，可在下游适配过程中保持安全性。我们的方法构建一个安全条件化基，并对权重矩阵进行重参数化，使安全相关信息集中于少量方向。随后，我们识别安全关键方向，并在重参数化空间中冻结这些方向，同时更新其余互补方向以进行下游适配。通过显式构建参数空间以解耦安全信息与任务信息，我们的方法减少了目标之间的干扰，并能够同时保持安全性和提升下游性能。实验结果表明，我们的方法在下游性能与安全性保留之间实现了更有利的权衡，证明了其用于可靠LLM微调的有效性。
cs.LG / 85 / 2610.26486
Wet-Biased Data Pipelines Mislead the Training and Evaluation of Generative Precipitation Nowcasting
湿润偏置的数据流水线会误导生成式降水临近预报的训练与评估
diffusion
扩散模型相关
Abstract
Generative models for precipitation nowcasting are trained and evaluated predominantly on wet-biased datasets, which over-represent precipitation events to emphasize rare extremes. Selecting test cases based on target observations is known to favor forecasts that match the selected distribution rather than the intended deployment distribution. We show analytically how this target-dependent sampling also shifts the conditional distribution learned by generative models. When training and test sampling match, even an ideal model learns precisely the distribution favored by evaluation, allowing the evaluation bias to conceal the training bias. Using a fixed satellite-to-radar diffusion model, we demonstrate that target-dependent test sampling can reverse the calibration assessment. The model appears under-predictive on a wet-biased test set but over-predictive on a deployment-representative test set. This result is replicated outside the model's training domain after a seven-year gap. Utilizing classifier-free guidance as an example, we demonstrate how the contradictory calibration assessment can misdirect downstream model development.
Chinese Translation
用于降水临近预报的生成式模型主要在湿润偏置的数据集上进行训练和评估，这些数据集对降水事件进行了过度表示，以强调稀有的极端事件。已知基于目标观测来选择测试样本，会偏向那些匹配所选分布而非预期部署分布的预报。我们从解析角度展示了这种依赖于目标的采样如何同样改变生成式模型所学到的条件分布。当训练采样与测试采样相匹配时，即使是一个理想模型也会精确地学到评估所偏好的分布，从而使评估偏差掩盖训练偏差。使用一个固定的卫星到雷达扩散模型，我们证明了依赖于目标的测试采样可以逆转校准评估。该模型在湿润偏置的测试集上显得预测不足，但在具有部署代表性的测试集上却显得预测过度。在相隔七年之后，这一结果在该模型的训练域之外得到了复现。以无分类器引导为例，我们展示了这种相互矛盾的校准评估如何误导下游模型开发。
cs.LG / 86 / 2610.26693
Parallel Language Models via Variational Decoding
通过变分解码的并行语言模型
diffusion
扩散模型相关
Abstract
Autoregressive language models emit one token per forward pass. This causes high inference latency, in particular as models are scaled up for quality. In this paper, we introduce Parallel Language Models (PLMs), which generate mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables. PLMs are trained using a variational bound that reduces to an autoregressive loss plus a parallel decoding loss. Empirically, PLMs are on a faster and better Pareto frontier than autoregressive and discrete diffusion baselines. Pretrained PLMs achieve near-autoregressive pretraining loss on model scales of 80M to 3B parameters, and speedup follows a power law in model size. When post-trained, our 1.3B model matches or exceeds an identically trained autoregressive model on math, coding, and instruction-following benchmarks, speeding up inference $2.3\text{-}5.7\times$. Together, we show that parallel generation can be trained from a single training objective at autoregressive quality. We will make our code available at https://github.com/mandt-lab/plm.
Chinese Translation
自回归语言模型每次前向传播只生成一个 token。这导致较高的推理延迟，尤其是当模型为提升质量而扩大规模时。本文提出并行语言模型（Parallel Language Models, PLMs），其通过潜在辅助变量路由依赖关系，在一次前向传播中生成相互一致的 token。PLMs 使用一个变分界进行训练，该变分界可化简为自回归损失加上并行解码损失。在实验上，PLMs 处于比自回归和离散扩散基线更快且更优的帕累托前沿上。预训练的 PLMs 在 80M 至 3B 参数的模型规模上取得了接近自回归的预训练损失，且加速比随模型规模呈幂律变化。经过后训练后，我们的 1.3B 模型在数学、代码和指令遵循基准上达到或超过同等训练的自回归模型，并将推理加速 $2.3\text{-}5.7\times$。综上，我们表明并行生成可以在自回归质量下通过单一训练目标进行训练。我们将在 https://github.com/mandt-lab/plm 公开我们的代码。
cs.LG / 87 / 2610.26718
Deep Pontryagin Learning for Continuous-Time Stochastic Control with Jumps
带跳跃的连续时间随机控制的深度庞特里亚金学习
diffusion
扩散模型相关
Abstract
We develop a deep Pontryagin learning framework for high-dimensional continuous-time stochastic control problems that naturally accommodates jump-diffusions. In contrast to direct policy optimization, which repeatedly differentiates the objective through the entire time-discretized dynamics, our method separates adjoint learning from policy improvement by explicitly incorporating state-sensitivity information from the Pontryagin stochastic maximum principle. For a fixed Markov control, we first compute pathwise adjoint labels by backpropagating the reward through time along simulated state trajectories, and project these labels onto the Markov state through a supervised regression step. The control is then updated by enforcing the pointwise Hamiltonian stationarity condition. We prove that the resulting projected adjoint process satisfies a discrete-time Pontryagin principle and converges, as the time step tends to zero, to the solution of the corresponding continuous-time adjoint backward stochastic differential equation. Numerical experiments illustrate the accuracy, stability, and scalability of the approach.
Chinese Translation
我们提出了一个深度庞特里亚金学习框架，用于高维连续时间随机控制问题，该框架自然地容纳跳跃扩散过程。与直接策略优化不同——后者通过整个时间离散化动力学反复对目标函数求导——我们的方法通过显式引入来自庞特里亚金随机最大值原理的状态敏感性信息，将伴随学习与策略改进分离。对于一个固定的马尔可夫控制，我们首先通过沿模拟的状态轨迹对奖励进行时间反向传播来计算路径伴随标签，并通过一个监督回归步骤将这些标签投影到马尔可夫状态上。随后，通过强制逐点哈密顿平稳性条件来更新控制。我们证明了所得的投影伴随过程满足离散时间庞特里亚金原理，并且当时间步长趋于零时，收敛到相应连续时间伴随倒向随机微分方程的解。数值实验说明了该方法的准确性、稳定性和可扩展性。
cs.LG / 88 / 2610.26790
Training with ML-labeled Data via Stochastic First- and Zeroth-Order Optimization
通过随机一阶与零阶优化在带有机器学习标注数据上进行训练
large language model
大语言模型相关
Abstract
Training a machine learning (ML) model is an optimization process that typically requires a set of examples with ground-truth labels beforehand. Obtaining these ground truths usually requires laborious human effort and can be costly. The scalability of human labeling becomes a concern when there are a large number of data points that need to be labeled. To address this practical issue, we consider an optimization setting where a subset of samples has ground-truth labels, while the remaining data points were labeled by an ML predictor, which could potentially assign incorrect labels to those points. We propose both first-order and zeroth-order stochastic optimization methods for training on datasets consisting of a mix of true-labeled and ML-labeled samples. We further conduct comprehensive empirical studies from training a linear regression model to fine-tuning a large language model (LLM) to showcase the effectiveness of the proposed algorithms for training with ML-labeled data.
Chinese Translation
训练一个机器学习（ML）模型是一个优化过程，通常需要事先准备一组带有真实标签的样本。获取这些真实标签通常需要繁重的人工劳动，并且可能成本高昂。当存在大量需要标注的数据点时，人工标注的可扩展性便成为一个令人担忧的问题。为解决这一实际问题，我们考虑一种优化设定：其中一部分样本具有真实标签，而其余数据点则由一个 ML 预测器标注，该预测器可能会为这些数据点赋予不正确的标签。我们提出了同时基于一阶和零阶的随机优化方法，用于在由真实标签样本和 ML 标注样本混合构成的数据集上进行训练。我们进一步开展了从训练线性回归模型到微调大语言模型（LLM）的全面实证研究，以展示所提出算法在利用 ML 标注数据进行训练方面的有效性。
cs.AI / 89 / 2610.25659
Beyond Leaderboards: A Method for Application-Specific Capability Profiling of Large Language Models
超越排行榜：一种面向特定应用的大型语言模型能力画像方法
large language model
大语言模型相关
Abstract
Public leaderboards rank large language models (LLMs) in general, but an engineer has to choose a model for a specific application, architecture and cost. This paper presents a method for application-specific capability profiling of LLMs and its realization as a containerized testbed. The method treats each application as a unit of extension: an Oracle-to-PostgreSQL migration agent, a documentation-grounded chatbot or an enterprise-support automation agent runs in its own container behind a small contract, brings its own dataset, evaluator and owned attributes, and is onboarded together with its dashboards. Measurements accumulate in a versioned, sparse metamodel of every model. The method separates the model under test from judges and agents, freezes and hashes plans, refuses to compare values from different contexts, diagnoses the instrument before the system, and applies a null-agent check and a noise rule whose operating characteristics are quantified by simulation. In the evaluated automation conditions, model choices produced larger differences on selected observable outcomes than framework choices, although the experimental design does not establish a general causal ranking of these factors. The experiments also show that instruction tuning is deployment-specific and that system-reported success can exceed independently observed success: of 216 migration units checked against a real Oracle database, 108 were demonstrably correct. These measurements establish application-specific observations, not a universal model ranking.
Chinese Translation
公共排行榜对大型语言模型（LLMs）进行总体排名，但工程师必须针对特定应用、架构和成本来选择模型。本文提出一种面向特定应用的大型语言模型能力画像方法，并将其实现为一个容器化测试平台。该方法将每个应用视为一个扩展单元：Oracle 到 PostgreSQL 迁移代理、基于文档的聊天机器人或企业支持自动化代理各自运行在自有容器中，遵循一个小型契约，自带数据集、评估器和自有属性，并连同其仪表板一起接入。测量结果累积到每个模型的带版本、稀疏元模型中。该方法将被测模型与评判器和代理分离，冻结计划并计算其哈希，拒绝比较来自不同上下文的数值，先诊断测量工具再诊断系统，并应用空代理检查和噪声规则，其运行特性通过仿真进行量化。在所评估的自动化条件中，模型选择在选定的可观测结果上产生的差异大于框架选择，尽管实验设计并未确立这些因素的一般因果排序。实验还表明，指令微调具有部署特异性，且系统报告的成功可能超过独立观测到的成功：在对照真实 Oracle 数据库检查的 216 个迁移单元中，有 108 个可证明为正确。这些测量确立的是面向特定应用的观测结果，而非通用的模型排名。
cs.MA / 90 / 2610.25851
Learning Who Talks to Whom: Role-Aware Communication Topology Scheduling for Efficient LLM Multi-Agent Collaboration
学习谁与谁对话：面向高效 LLM 多智能体协作的角色感知通信拓扑调度
large language model
大语言模型相关
Abstract
Multi-agent systems driven by large language models (LLMs) rely on structured communication topologies to facilitate information exchange and collaborative reasoning. However, effectively scheduling these topologies to dynamically determine who talks to whom and when to communicate remains a fundamental challenge in balancing collective reasoning performance against exponential token overhead. Existing topology scheduling approaches either enforce static communication graphs that cannot adapt to dynamic problem-solving states, or overlook agent role complementarity while failing to explicitly incorporate the synthesized graph structure into team-level decision evaluation. To tackle these limitations, we propose ACTS, a role-aware adaptive communication topology scheduling framework for efficient LLM multi-agent collaboration. Orchestrated within a Centralized Training with Decentralized Execution (CTDE) paradigm, ACTS introduces a novel dual-attention mechanism. Specifically, a Role-conditioned Relational Attention module dynamically selects variable-cardinality interaction partners based on agent roles and evolving execution states, while a Topological Attention Mixer explicitly embeds the directed communication topology into team value decomposition for precise credit assignment. Comprehensive evaluations across 9 challenging benchmarks covering code generation, multi-step reasoning, general knowledge, and competitive mathematics demonstrate that ACTS consistently outperforms state-of-the-art baselines, boosting average accuracy by up to 6.19 percentage points while reducing token consumption by over 15%. The project website is available at https://acts-mas.github.io/ACTS/.
Chinese Translation
由大语言模型（LLM）驱动的多智能体系统依赖结构化的通信拓扑来促进信息交换与协同推理。然而，如何有效地调度这些拓扑，以动态地确定谁与谁对话以及何时通信，仍然是在集体推理性能与指数级 token 开销之间取得平衡的一项根本性挑战。现有的拓扑调度方法要么强制采用无法适应动态问题求解状态的静态通信图，要么忽视智能体角色的互补性，同时未能将所合成的图结构显式地纳入团队级决策评估。为解决这些局限，我们提出了 ACTS，一个面向高效 LLM 多智能体协作的角色感知自适应通信拓扑调度框架。ACTS 在集中式训练与分散式执行（CTDE）范式下进行编排，引入了一种新颖的双重注意力机制。具体而言，一个角色条件化的关系注意力模块基于智能体角色与不断演化的执行状态，动态选择基数可变的交互伙伴，而一个拓扑注意力混合器则将有向通信拓扑显式地嵌入团队价值分解中，以实现精确的信用分配。在涵盖代码生成、多步推理、通用知识与竞赛数学的 9 个具有挑战性的基准上开展的全面评估表明，ACTS 始终优于最先进的基线方法，平均准确率最高提升 6.19 个百分点，同时将 token 消耗降低超过 15%。项目网站见 https://acts-mas.github.io/ACTS/。
cs.AI / 91 / 2610.26487
SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models
SPAR：面向视觉-语言-动作模型的结构化剪枝与动作恢复
large language model
大语言模型相关
Abstract
The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to $π_{0.5}$, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a $1.98\times$ full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.
Chinese Translation
机器人有限的板载计算资源使得视觉-语言-动作（VLA）模型的推理需求难以低成本地满足。然而，通过激进压缩来降低这些需求，可能会使操作性能下降到仅靠监督微调无法恢复的程度。我们提出 SPAR（结构化剪枝与动作恢复），它通过剪枝感知调优同时剪枝大语言模型（LLM）主干的隐藏维度和前馈中间维度，然后应用强化学习（RL）来恢复策略性能。应用于 $π_{0.5}$ 时，SPAR 将语言主干从 2.34B 参数减少到 0.09B，将完整模型从 3.37B 减少到 1.12B，分别移除了约 96% 和 67% 的参数。在这种压缩水平下，RL 将四个 LIBERO 套件上的平均成功率从 84.55% 提升到 93.45%，提升了 8.90 个百分点。压缩后的策略相对于未剪枝模型实现了 $1.98\times$ 的全策略推理加速。这些结果识别出大量中间维度冗余，并表明剪枝后的 RL 可以在紧凑策略中恢复操作性能。
cs.AI / 92 / 2610.26507
Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning
学习任务，而非调度：减少双臂模仿学习中的虚假跨臂耦合
diffusion
扩散模型相关
Abstract
A bimanual demonstration specifies both what the robot should do and one particular schedule for when each arm should act. Standard imitation learning does not distinguish task-required cross-arm constraints from incidental scheduling choices, and may therefore exploit opposite-arm progress as a shortcut cue, creating spurious cross-arm coupling even where relative timing is irrelevant for the current task phase. We introduce Constraint-preserving Temporal Recombination (CTR), a training-time data intervention that preserves recorded arm-local sensorimotor content while retiming the corresponding segments to realize admissible schedules that preserve task-required cross-arm constraints. We use CTR to train Diffusion Policy, Action Chunking with Transformers, and $π_{0.5}$, and evaluate them on four simulated and four real-robot tasks. Across six real-robot bimanual progress-state initializations, CTR raises worst-state success from 56% for the strongest non-CTR schedule mixture to 76%. With $π_{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR. On real robots, CTR violates task constraints in only 0--4% of trials, versus 36--98% for narrow-schedule baselines. Code will be released at https://github.com/MINT-SJTU/CTR.
Chinese Translation
一个双臂演示既规定了机器人应该做什么，也规定了每条手臂应在何时行动的一种特定调度安排。标准模仿学习不区分任务所需的跨臂约束与偶然的调度选择，因此可能利用对侧手臂的进度作为捷径线索，即使在相对时序对当前任务阶段无关紧要的情况下也会产生虚假的跨臂耦合。我们提出约束保持的时间重组（Constraint-preserving Temporal Recombination, CTR），这是一种训练时的数据干预方法，它在保留所记录的单臂局部感觉运动内容的同时，对相应片段进行重新定时，以实现保持任务所需跨臂约束的可行调度。我们使用 CTR 训练 Diffusion Policy、Action Chunking with Transformers 和 $π_{0.5}$，并在四个仿真任务和四个真实机器人任务上对它们进行评估。在六种真实机器人双臂进度状态初始化下，CTR 将最差状态成功率从最强的非 CTR 调度混合方法的 56% 提高到 76%。在 Independent 任务上使用 $π_{0.5}$ 时，在 Concurrent 调度分布下，4 s 的暂停会将 2.56 s 的延迟传递给对侧手臂，而使用 CTR 时仅传递 0.25 s。在真实机器人上，CTR 仅在 0--4% 的试验中违反任务约束，而窄调度基线为 36--98%。代码将在 https://github.com/MINT-SJTU/CTR 发布。
cs.LG / 93 / 2610.26100
Multilingual Voicemail and Answering Machine Detection via Pretrained Audio Embeddings and Gradient Boosting Machines
基于预训练音频嵌入和梯度提升机的多语言语音邮件与答录机检测
large language model
大语言模型相关
Abstract
Answering machine detection (AMD) is a fundamental task in telephony that classifies whether an answered call is picked up by a human or a voicemail (machine). Despite its widespread use across cloud communication platforms, existing approaches rely on proprietary systems with undisclosed architectures, or require computationally expensive models that cannot operate in real time. We propose a lightweight AMD system that combines pretrained audio embeddings with gradient boosting machines, trained to prioritize the correct detection of human-answered calls. We evaluate seven models across two datasets: three models that we propose, a proprietary industry model, two prior academic approaches, and a large language model baseline that relies on automatic speech recognition. Furthermore, we release OpenAMD, the first publicly available AMD dataset comprising of more than 500 recordings in 7 languages. Our best model achieves the highest precision across both datasets while maintaining low inference latency. We assess the robustness of models to additive noise and packet loss, and perform tests of statistical significance to discover per-language and per-gender model performance. Finally, we evaluate our model on an out-of-domain task of detecting robocalls, where it achieves 94.34% recall without any additional training, which is comparable to the industry standard.
Chinese Translation
答录机检测（AMD）是电话通信中的一项基础任务，它判断已接通的呼叫是由真人接听还是由语音邮件（机器）接听。尽管其已在云通信平台中得到广泛使用，现有方法要么依赖于架构未公开的专有系统，要么需要计算成本高昂、无法实时运行的模型。我们提出一种轻量级 AMD 系统，将预训练音频嵌入与梯度提升机相结合，并在训练中优先保证正确检测真人接听的呼叫。我们在两个数据集上评估了七种模型：我们提出的三种模型、一个专有行业模型、两种先前的学术方法，以及一个依赖自动语音识别的大语言模型基线。此外，我们发布了 OpenAMD，这是首个公开可用的 AMD 数据集，包含 7 种语言的 500 多条录音。我们最好的模型在两个数据集上均取得最高精确率，同时保持较低的推理延迟。我们评估了模型对加性噪声和丢包的鲁棒性，并进行了统计显著性检验，以发现模型在每种语言和每种性别上的性能表现。最后，我们在检测机器人骚扰电话（robocalls）的域外任务上评估了我们的模型，其在没有任何额外训练的情况下达到 94.34% 的召回率，与行业标准相当。
cs.SE / 94 / 2610.25616
Better Specifications, Better Code: Fine-Tuning LLMs for Behavioral Contract Generation
更好的规范，更好的代码：面向行为契约生成的大语言模型微调
large language model
大语言模型相关
Abstract
Large language models (LLMs) demonstrate strong code-generation capabilities but often struggle with functional correctness due to implicit or ambiguous behavioral requirements. We present CG2, a contract-guided code generation framework that translates natural-language tasks into structured behavioral contracts specifying interfaces, preconditions, postconditions, and invariants. CG2 fine-tunes an LLM to generate these contracts and introduces an iterative semantic validation and repair process to identify and correct incomplete, inconsistent, or incorrect specifications before guiding an unchanged downstream code-generation model. We evaluate CG2 on 1,007 held-out tasks using five code LLMs: DeepSeek-Coder-V2-Lite-Instruct, Ministral-3-14B-Instruct-2512, Qwen2.5-Coder-14B-Instruct, Qwen2.5-Coder-32B-Instruct, and Qwen3.5-27B. Results show that CG2 consistently improves pass@1 over direct code generation by 1.49-3.27 percentage points (7.58%-22.45% relative improvement) and over contract guidance without fine-tuning by 1.19-2.38 percentage points (7.04%-14.65% relative improvement). Manual analysis further reveals that semantically correct contracts are associated with higher code-generation accuracy. These findings demonstrate that explicitly specifying and validating behavioral requirements can improve functional correctness without modifying downstream code-generation models.
Chinese Translation
大语言模型（LLMs）展现出强大的代码生成能力，但由于行为需求隐含或模糊，往往在功能正确性上表现不佳。我们提出 CG2，一个契约引导的代码生成框架，它将自然语言任务转换为结构化行为契约，这些契约规定接口、前置条件、后置条件和不变式。CG2 微调一个大语言模型以生成这些契约，并引入迭代式语义验证与修复过程，在引导一个保持不变的下游代码生成模型之前，识别并纠正不完整、不一致或错误的规范。我们在 1,007 个留出任务上使用五个代码 LLM 评估 CG2：DeepSeek-Coder-V2-Lite-Instruct、Ministral-3-14B-Instruct-2512、Qwen2.5-Coder-14B-Instruct、Qwen2.5-Coder-32B-Instruct 和 Qwen3.5-27B。结果表明，与直接代码生成相比，CG2 始终将 pass@1 提高 1.49–3.27 个百分点（相对提升 7.58%–22.45%）；与未经微调的契约引导相比，提高 1.19–2.38 个百分点（相对提升 7.04%–14.65%）。人工分析进一步揭示，语义正确的契约与更高的代码生成准确率相关。这些发现表明，显式指定并验证行为需求可以在不修改下游代码生成模型的情况下提高功能正确性。
cs.SE / 95 / 2610.25826
From Finding to Fixing: How Useful is Fault Localization for Agentic Program Repair?
从定位到修复：故障定位对智能体程序修复有多大用处？
large language model
大语言模型相关
Abstract
Fault localization (FL) aims to identify code locations responsible for software failures and is commonly evaluated by how accurately it ranks known fault locations. Large language model (LLM) agents are changing program repair and are increasingly used to search and reason over repositories. Although recent LLM-based FL methods have improved localization accuracy, their impact on agentic repair remains unclear. To address this gap, we conduct a systematic empirical study of FL in agentic repair. We evaluate six recent FL methods on SWE-bench Verified and Defects4J with three LLMs under a common repair framework and analyze 16,248 repair trajectories. We also examine gold locations and bug-cause descriptions as forms of FL guidance across different repair contexts. The studied FL methods do not consistently improve repair success, and higher localization accuracy does not consistently lead to larger repair gains. Yet FL guides agents' early search and helps them reach relevant code sooner. Accurate locations can reduce repair token use by up to 39.2\%, but current FL methods often consume more tokens than they save during repair. Experiments with gold guidance suggest that FL has greater potential when bug inputs contain fewer location clues. Providing bug causes can further improve repair success beyond locations alone. We also find that ground truth from a single reference patch does not fully capture the locations used by successful repairs. These findings help clarify the value of FL in agentic repair and inform future work on its design and evaluation.
Chinese Translation
故障定位（FL）旨在识别导致软件失败的代码位置，通常通过其排序已知故障位置的准确性来评估。大语言模型（LLM）智能体正在改变程序修复，并越来越多地被用于在代码仓库中进行搜索和推理。尽管近期基于LLM的FL方法提高了定位准确性，但它们对智能体修复的影响仍不清楚。为弥补这一空白，我们对智能体修复中的FL开展了一项系统性的实证研究。我们在统一的修复框架下，使用三个LLM在SWE-bench Verified和Defects4J上评估了六种近期FL方法，并分析了16,248条修复轨迹。我们还考察了金标准位置与缺陷原因描述作为不同修复情境下FL指导形式的作用。所研究的FL方法并不能稳定地提高修复成功率，更高的定位准确性也并不总能带来更大的修复收益。然而，FL能引导智能体的早期搜索，并帮助它们更快地到达相关代码。准确的位置可将修复的token使用量最多减少39.2%，但当前的FL方法在修复过程中消耗的token往往多于其节省的token。使用金标准指导的实验表明，当缺陷输入包含较少的位置线索时，FL具有更大的潜力。提供缺陷原因可以在仅提供位置之外进一步提高修复成功率。我们还发现，来自单个参考补丁的真值并不能完全涵盖成功修复所使用的位置。这些发现有助于阐明FL在智能体修复中的价值，并为未来关于其设计与评估的工作提供参考。
cs.SE / 96 / 2610.26392
The EBC Proportionality Law: A Recurring Architectural-Role Pattern Across Java Software
EBC 比例定律：Java 软件中反复出现的架构角色模式
large language model
大语言模型相关
Abstract
Architectural roles describe how responsibilities are distributed within software systems. Entity--Boundary--Control (EBC) separates domain information, external interaction, and workflow coordination, but its proportions across diverse software remain unclear. This study examines file-level classifier-derived EBC proportionality in 1,556,985 Java files from 732 project instances and 78 functional categories. A classifier trained on labels generated by a large language model assigned one representative role per file. The labels were aggregated at project and functional-category levels and analysed using descriptive statistics and density-based clustering. The results show that Entity was less common than Boundary and Control. The dominant cluster covered 48 categories, 472 project instances, and 47.75% of classified files; its mean composition was 13.25% Entity, 42.65% Boundary, and 44.10% Control. The remaining 30 categories form one alternative and 11 smaller specialised patterns. Some selected variations were consistent with category responsibilities, while others received mixed support. Agreement with independent expert labels was limited (31.37% accuracy; macro-F1 0.31; Entity recall 13%). All reported proportions and clusters are therefore patterns in classifier-assigned roles, not verified architectural ground truth. Within this measurement boundary, the study provides a large-scale empirical reference of EBC distributions at the file level, and it documents the limitations of the automated approach.
Chinese Translation
架构角色描述了职责在软件系统内部是如何分布的。实体—边界—控制（EBC）将领域信息、外部交互与工作流协调分离开来，但其在不同软件中的比例构成仍不清楚。本研究考察了来自 732 个项目实例和 78 个功能类别的 1,556,985 个 Java 文件中由分类器导出的文件级 EBC 比例。一个在大语言模型生成的标签上训练的分类器为每个文件分配一个代表性角色。这些标签在项目和功能类别两个层级上进行聚合，并使用描述性统计和基于密度的聚类进行分析。结果表明，Entity 比 Boundary 和 Control 更为少见。主导聚类涵盖 48 个类别、472 个项目实例以及 47.75% 的已分类文件；其平均构成为 13.25% 的 Entity、42.65% 的 Boundary 和 44.10% 的 Control。其余 30 个类别构成一个替代模式以及 11 个规模较小的专门化模式。一些选定的变异与类别职责相符，而另一些则只得到了混合的支持。与独立专家标签的一致性有限（准确率 31.37%；宏 F1 为 0.31；Entity 召回率为 13%）。因此，所有报告的比例和聚类都是分类器所分配角色中的模式，而非经核实的架构真实基准。在这一测量边界之内，本研究提供了文件层面 EBC 分布的大规模经验参考，并记录了该自动化方法的局限性。
cs.SE / 97 / 2610.26464
VerusEval: Evaluating LLM-Generated Verus Specifications Beyond Verifier Acceptance
VerusEval：超越验证器接受性评估 LLM 生成的 Verus 规范
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used to generate Verus contracts for Rust programs together with the proofs that establish them. Yet verifier acceptance, the common measure of success, shows only that the implementation satisfies the generated contract, so it leaves open whether the contract captures the intended behavior. Assessing specification generation therefore requires an evaluation methodology, not a single metric. We present VerusEval, a framework that evaluates generated contracts along six dimensions, including Verus-native checks for semantic triviality, validated input-output (I/O) suites, and contract equivalence, and that relates them so that the strength and blind spots of each become measurable. Evaluating 13,716 artifacts from four workflows, five LLMs, and two prompting settings on 762 tasks, we find that acceptance overstates quality: only 42.3\% of the accepted artifacts complete all three semantic checks without a flag, and across workflows and models, higher acceptance does not consistently come with more correct contracts. I/O evaluation rejects 23.6\% of the accepted artifacts, yet only 63.2\% of those that pass it are proved equivalent. Of the accepted artifacts flagged by contract comparison, 84.1\% have a rejected completeness obligation, and the direction of a failure prioritizes inspection of the generated contract or the reference; it does not by itself establish which artifact is defective. Screening and review further identify 36 candidate defects among 61 selected reference contracts, and repairing one of them makes 10 of its 11 failing artifacts equivalent. Based on these findings, we derive four actionable lessons for evaluation, generation, and benchmark maintenance.
Chinese Translation
大型语言模型（LLM）越来越多地被用于为 Rust 程序生成 Verus 合约以及建立这些合约的证明。然而，验证器接受性这一常用的成功衡量标准只表明实现满足了生成的合约，因此它并未回答该合约是否刻画了预期行为。因此，评估规范生成需要一套评估方法论，而不是单一指标。我们提出 VerusEval，这是一个沿六个维度评估生成合约的框架，包括针对语义平凡性的 Verus 原生检查、经过验证的输入输出（I/O）套件以及合约等价性，并将它们关联起来，使各自的优势与盲点变得可度量。在 762 个任务上评估来自四种工作流、五个 LLM 和两种提示设置的 13,716 个产物，我们发现接受性高估了质量：只有 42.3% 的被接受产物在未被标记的情况下完成全部三项语义检查，并且跨工作流和模型来看，更高的接受率并不总是伴随更正确的合约。I/O 评估拒绝了 23.6% 的被接受产物，然而在通过 I/O 评估的产物中，只有 63.2% 被证明等价。在被合约比较标记的被接受产物中，84.1% 存在被拒绝的完整性义务，而失败的方向优先指向对生成合约或参考合约的检查；它本身并不能确定哪个产物存在缺陷。筛选与审查进一步在 61 个选定的参考合约中识别出 36 个候选缺陷，并且修复其中一个缺陷后，其 11 个失败产物中有 10 个变为等价。基于这些发现，我们为评估、生成和基准维护提炼出四条可操作的启示。
cs.LG / 98 / 2610.26458
Residual-Based Adaptive Refinement in Hybrid Variational-Collocation PINNs for Singular Elliptic PDEs with Mixed Dirichlet-Neumann Boundary Conditions
diffusion
扩散模型相关
Abstract
Solutions of second-order elliptic PDE problems on nonconvex or cracked domains with mixed Dirichlet-Neumann conditions typically lie outside $H^2(Ω)$. Near re-entrant corners, crack tips and points where the boundary condition changes type, the gradient grows like $r^{λ-1}$ with $λ<1$. We study residual-based adaptive refinement in hybrid variational-collocation PINNs. A multiscale Fourier-feature network is trained on a composite loss. The loss combines a variational residual tested against Legendre bubble functions, a strong-form collocation term and penalties for the Dirichlet and Neumann data. The variational term needs only first derivatives of the network, whereas the collocation term requires second derivatives and provides a residual that can be evaluated at any point. After each training stage, candidate points selected by this residual are added to the collocation set. The method is tested on five two dimensional benchmarks, namely a Mode~III crack, a cross-shaped domain with four re-entrant corners with unequal amplitude, a reaction-diffusion problem with four localized peaks, a monotone semilinear problem with a boundary layer and the Motz problem. Compared with uniform point insertion under the same collocation and iteration budget, residual-based refinement reduced the relative $L^2$ error by up to a factor of $32$, the relative $L^\infty$ error by up to a factor of $38$ and the relative $H^1$ semi-norm error by up to a factor of $15$.
人工智能 (cs.AI)
129
cs.AI / 1 / 2610.25464
On Reasoning Correctness Directions
Abstract
We present experimental evidence suggesting that information about whether a model's reasoning is on the right track can be linearly decodable without defining a causally useful direction for steering. To study this, we introduce ChainSearch, a controlled arithmetic reasoning task that extends Countdown with explicit search and backtracking. We train a 3-million-parameter model to solve this task. Its compact size allows us to study internal model correctness directions in detail using linear probes, mean activation differences, and direct optimisation, as part of a rigorous search for internal reasoning-correctness directions. We find that while the model contains statistically significant internal directions of reasoning correctness, none of the resulting directions consistently improves reasoning under causal intervention. This calls into question whether the linear representation hypothesis is sufficient for identifying causally useful directions in the context of reasoning. Our code is made available together with this work.
cs.AI / 2 / 2610.25470
A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help
Abstract
Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text---especially on tasks where visualization would help, like geometry questions posed in words. We distinguish weak multimodality (the model can answer questions about an input image) from strong multimodality (the model's text representations themselves become vision-aligned, improving performance on text-only tasks with latent visual structure). To test for strong multimodality, we construct a task of shape-counting problems with paired equivalent text and image specifications. This task is trivial when the image is provided, but requires visualization when specified in text as line coordinates. We compare instruction-tuned Qwen-2.5 and Mistral models against their vision-aligned counterparts on the text variant of the task. We find that the multimodal models do no better in this setting, despite strong performance on the image version of the task. Probing experiments corroborate this: vision-aligned decoder embeddings carry no more visual structure than their text-only counterparts. Finally, we provide an optimization objective that distills in-context visual embeddings into text embeddings. This alignment induces gains on the text variant of the task without explicit task supervision, even on samples more complex than seen during alignment. This suggests that VQA-style alignment leaves available cross-modal transfer untapped.
cs.AI / 3 / 2610.25474
P(Yes): 15$\times$ Improvements via Test-Time Scaling of Decision Models
Abstract
Decision models answer typed questions, such as multiple-choice or yes/no questions, with a probability for each option. We add test-time search to frozen decision models, with no training: we show the model a complete proposed answer, read off its probability of saying "Yes" to the proposal's correctness, and search for answers that it rates higher. The search is done either by editing one answer at a time or by joint optimization, which changes several answers at once. Search raises accuracy for every model and test set we tried, and the gains are largest for exact correctness on logic puzzles, where every answer in a task must be right. On 128 such puzzles, it takes Qwen3.8-27B from 2 to 30 exactly correct puzzles ($15\times$) and its per-blank accuracy from 27.8% to 48.5%. On a shared set of 47 puzzles, searching over each model's own Yes/No score raises the per-blank accuracy of StartLux Instant, Perplexity Decider, Clef, and JEV, which we reach only through its public API, by 4.0 to 13.0 percentage points, and raises the number of puzzles they get exactly correct, out of 47 and averaged over three search seeds, from 10 to 35, 9 to 26, 2 to 11, and 9 to 31.7 respectively.
cs.AI / 4 / 2610.25496
From Proposal to Execution: Modeling Interaction Semantics in Shared Autonomy
Abstract
In many human-AI systems, candidate actions are not executed directly but are first evaluated by another decision-making entity under a potentially different objective or decision model. Standard decision-theoretic frameworks do not require this interaction to be represented explicitly, as they define transition dynamics directly over state-action pairs, allowing evaluation to be absorbed within the transition model. We introduce a semantics-aware decision framework in which a candidate action is proposed, evaluated under an interaction context, and then resolved into an executed outcome. We formalize this through a joint action language and an interpretation function that maps each joint action and context to a semantic outcome. Transition and reward dynamics are induced through these semantic outcomes, thereby introducing an explicit representational layer for \textit{interaction semantics}. Semantic outcomes provide explicit units for organizing interaction principles, which can be instantiated through application-specific evaluation mechanisms and reused across distinct interaction instances. This provides a structured interface for specifying context-dependent evaluation separately from the resulting execution dynamics, without requiring changes to the underlying planning algorithm. We instantiate the framework in shared human-AI autonomy, capturing interaction patterns such as advisory and authorization, and illustrating how interaction context influences decision making. The resulting formulation remains compatible with standard Markov decision process (MDP) planning.
cs.AI / 5 / 2610.25499
Assessing the AI Readiness of Construction Datasets: A Multidimensional Operational Framework
Abstract
The rapid growth of artificial intelligence (AI) applications in construction has increased the demand for datasets suitable for reliable model development, evaluation, and reuse. However, construction datasets vary widely in documentation, quality, governance, and evaluation support, and consistent criteria for assessing AI readiness remain limited. This study develops a Construction Dataset AI Readiness Framework with 30 attributes across four dimensions: Metadata Completeness, Data Quality and Representativeness, Data Governance and Accessibility, and AI Utility. The framework uses attribute-specific scoring, applicability screening, and automated and manual verification. It was evaluated using 21 publicly available datasets across seven data-type groups. Overall readiness scores ranged from 0.720 to 0.964, with a mean of 0.879. Hybrid verification achieved 90.5% agreement between automated suggestions and manually verified scores. The framework supports consistent assessment of heterogeneous construction datasets and identification of readiness gaps.
cs.AI / 6 / 2610.25516
Bud ContextCompact 2.0: Context Compaction for Agent Success and Accuracy with Small Language Models and Constrained Context Windows
Abstract
Agents that work for many turns outgrow the context window of the model that serves them, and on small language models (SLMs) and constrained inference systems that window is small. Context compaction replaces older turns with a summary so a session can continue, but it is lossy, costs a model call, and can help or hurt depending on when it runs. We describe Bud ContextCompact 2.0, the compaction built into the Bud agent runtime, and evaluate it in two settings. On recall questions over recorded agent sessions, compaction to a 16K budget raised a 1.2B-parameter model's accuracy to 0.335, from 0.247 with truncation and 0.176 with its full 32K history (p < 0.001), and raised a 9B model from 0.515 (truncation) to 0.577 (p = 0.04). On 100 SWE-bench Verified tasks run end to end by a 27B reasoning model on a 49K-token window (556 graded sessions), compaction raised the resolve rate on long sessions from 41% to 53% (+12.3 points, 95% CI +1.6 to +23.0, McNemar p = 0.043), left short sessions unchanged, and used 13% fewer tokens per solved task at equal total cost. The main failure without compaction was not the history overflowing but a reasoning turn exhausting the room the history left. The gain has limits: it vanished under a 45-minute session budget, because each summary took a median of 161 s, and in a threshold study it appeared only when compaction triggered early enough to leave room. An 8B model submitted after about ten turns at 5-8K tokens and never needed compaction. We argue that the accuracy results transfer across hardware, including CPU inference, for an equivalent model and window, while the time cost scales with the hardware and must be budgeted for.
cs.AI / 7 / 2610.25525
StopLoss: Teaching Agents to Predict Failure and Stop Early
Abstract
Coding agents keep working on runs that will fail verification, and asking a frontier model to watch for this is expensive in its own right. We introduce StopLoss, a small external monitor that learns failure recognition from completed trajectories and their verifier outcomes. The name is inspired by the stop-loss order in finance: StopLoss ends a run early when the run is likely to fail, which saves resources. Final outcomes label earlier local windows, so a fine-tuned language model can score the most recent tool steps in a single forward pass, and a threshold calibrated on successful runs turns that score into an abort decision without changing the executor. Trained on GPT-family SWE-rebench trajectories and evaluated on SWE-bench Verified and a SWE-bench Pro subset with GPT and Claude executors, StopLoss is compared against six general-purpose judges reading identical windows. Outcome supervision lifts failure prediction from near chance to a useful signal at the first check, at one to two orders of magnitude lower cost per check than any tested judge. This translates into savings: relative to completing every task, StopLoss lowers cost per retained solve by 4.5\% on Verified while preserving 99\% of successful runs, and by up to 31\% on the costlier Pro subset, reaching 48\% with a Qwen3-8B backbone, and in all four settings its cost per retained solve is below that of every judge policy. Outcome-trained small monitors can make failure recognition cheap enough for early stopping to pay for itself.
cs.AI / 8 / 2610.25530
From Clinical Risk Prediction to Operational Decisions: An Uncertainty-Aware Framework for ICU Staffing
Abstract
Advances in electronic health records (EHR) and large-scale critical care databases have enabled the development of machine learning models for the early detection of patient deterioration. While these models often achieve strong predictive performance for outcomes such as sepsis and mortality, most studies evaluate them purely on discrimination metrics. This leaves a critical gap: how to systematically translate probabilistic clinical risk estimates into downstream operational decisions under real-world constraints, such as shift rigidity and hiring delays. This study proposes an uncertainty-aware anticipatory staffing framework that bridges short-term patient risk prediction and staff planning without altering the upstream predictive model. Instead of relying on naive deterministic aggregation or fixed safety buffers, the framework explicitly models dynamic predictive uncertainty and evaluates extreme shortage exposure using Conditional Value-at-Risk (CVaR). These aggregated, risk-calibrated demand signals are then embedded into an hourly rolling-horizon scheduling policy to generate delay-constrained staffing decisions. Evaluated on realistic trajectories from the MIMIC-IV database, the proposed framework demonstrates a favorable cost-shortage trade-off compared to deterministic and purely reactive baselines. Pareto frontier analysis illustrates how uncertainty-aware demand construction changes resource allocation under varying safety preferences, while stress testing confirms the framework's operational robustness even when subjected to systematic predictive bias in the form of calibration drift. By formalizing this translation layer, this work provides a principled, transparent pathway for safely integrating probabilistic AI outputs into high-stakes healthcare operational systems.
cs.AI / 9 / 2610.25540
Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing
Abstract
Agentic LLMs increasingly support scientific prediction by grounding decisions in external evidence. In biology, they promise accurate, interpretable predictions for important problems by integrating complex experimental and computational evidence. However, serialized tool results can contain hidden label shortcuts, such as dataset identifiers or template artifacts, that create apparent success without biological reasoning. To address this risk, we introduce a two-stage audit. Stage 1, Per-Tool TF-IDF Decomposition (PT2D), localizes label-predictive patterns to individual tool results while avoiding signal dilution in the concatenated tool trace. Stage 2, Token-Class Redaction, counterfactually tests whether the frozen LLM depends on the localized shortcut. We apply this audit to TCR-pMHC binding, where tool augmentation substantially outperforms a no-evidence baseline. At face value, this suggests that the LLM has learned to use scientific evidence. The audit overturns this interpretation: PT2D localizes the strongest signal to identifiers in one tool result, and redacting them reduces performance to chance. These findings show how hidden label shortcuts can masquerade as scientific reasoning or discovery, motivating source-level audits before benchmark gains are trusted.
cs.AI / 10 / 2610.25550
Editing as Language: Teaching LLM Agents a Grammar for Native Context Editing
Abstract
A synopsis in novels folds a story into a few lines; a function in code gathers many steps with a single execution. Context editing is native to language itself. In this paper, we propose the Native Context Editing (NCE) framework, which gives language model agents a grammar for natively editing their own context. By encoding context-editing operations in a grammar of dedicated special tokens, we teach the agents to invoke edits through the language itself. We train agents on iteratively constructed trajectory segments, with each teacher edit conditioned on the evolving context. This establishes context editing as an atomic and inherent capability of LLM agents, enabling them to dynamically modify their context as they generate responses over long-horizon reasoning and interaction. Compared with context-editing methods that learn to generate tool-use code within specific external harness environments, our approach integrates context editing as an internal mechanism of the model's native generation process, enabling better generalization across diverse real-world workflows. On four long-horizon agentic benchmarks (Terminal-Bench 2.0, SWE-bench Verified, SWE-bench Multilingual, and SWE-bench Pro), NCE improves the average task success rate over the base model by 5.6 percentage points for LongCat-Flash-Lite-Sparse and by 4.5 for Qwen3.6-35B-A3B, while reducing average token usage by 36% and 15%, respectively.
cs.AI / 11 / 2610.25552
Symbolic Constraint Agents: Mining Formal Tool-Call Constraints to Prevent Agent Errors
Abstract
Symbolic constraints (programmatic, deterministic checks for tool calls) enable deterministic verification of LLM agents' tool calls, but existing approaches primarily enforce predefined safety policies rather than tool dependencies in complex tool environments. We explore whether symbolic constraints can be mined from environments and observations beyond predefined policies, and whether they can improve end-to-end task performance. We present Symbolic Constraint Agents (SCA), a framework that mines symbolic constraints from API documentation, refines them against execution traces, and checks the agent's tool calls with an SMT solver at runtime. A violation triggers a natural-language warning or, when the valid replacement is uniquely determined, an automatic repair. On AppWorld, refinement raises error-detection F1 from 37.0 to 58.3 on GPT-5.5 traces and from 39.3 to 61.9 on Qwen3.5-27B traces. At runtime, repair raises GPT-5.4-mini's task success by 5.4 points under a 40-call budget and 9.8 points when environment errors are hidden, while removing 59--87\% of failed calls across models, and transfers to $τ^2$-bench with gains of up to 5.6 points. These findings show that automatically acquired symbolic constraints can improve tool-calling reliability and task completion, particularly when failures are costly or difficult to diagnose.
cs.AI / 12 / 2610.25553
Narrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions
Abstract
System One models, introduced with TypeSafe's Jev, answer typed questions about a state with calibrated probabilities instead of generated text: choosing among options described in natural language (Choice), placing the state on an ordered scale (Score), or judging a yes/no statement (Noul). Jev is general-purpose: one set of weights, served as an API and adapted to a domain only through the request. We study the other end of the design space. A Narrow Decision Model (NDM) offers the same interface but is trained for the decisions of one organization, is small enough to run on its own infrastructure, and has an open architecture whose properties can be stated and tested. The state can be any text or structured record about which decisions are made; our running case is the routing of court communications. The architecture serializes a request as a prefix tree (the state, then each question, then each option) and runs a bidirectional encoder under a block attention mask with logical positions, so that each answer is, by construction, invariant to the order of its options and independent of the other questions asked. Routing through an organization chart is a composition of Choice decisions with an explicit "none of these" option. We describe training with proper scoring rules through a curriculum, calibration and abstention, decisions composed by rubrics, and a protocol that accepts or rejects the mechanism before held-out data is opened. A minimal validation on public and fictitious data shows that the invariances hold to rounding in a trained model, that a small instance routes on organization charts it has never seen well above lexical retrieval, and that, at equal compute, the block architecture reaches a lower loss than a pairwise cross-encoder. A first instance trained on real legal documents agrees with the LLM classifier in production on 91% of held-out documents.
cs.AI / 13 / 2610.25568
Design and Cross-Domain Evaluation of an Agent Execution Harness
Abstract
Completing complex tasks with language-model agents requires coordinating planning, execution, and review so the final result addresses the user's original objective. We present an agent execution harness that keeps the original objective separate from a revisable plan and an execution record. A planner assigns scoped tasks, workers produce outputs and artifacts, and a reviewer examines results against the original request before submitting or requesting additional work. Reusable domain-specific instructions guide these roles and are refined during construction using task inputs and execution trajectories, without official grading feedback. We evaluate complete harness configurations across terminal tasks, clinical responses, application building, and scientific questions. On Terminal-Bench 2.1, the harness passes 83 of 89 instances (93.26%); on HealthBench Professional, a Fable-configured system achieves a length-adjusted rubric score of 69.63 across 525 tasks. In a 53-task paired HealthBench comparison, the Fable-configured harness achieves an aggregate score 19.89 points higher than direct completion, despite a $-1.46$-point median task-level change; mean inference cost rises from \$0.29 to \$2.13 per task. We combine these outcomes with latency, documented instruction revisions, and reviewer decisions. Evaluating outcomes alongside execution trajectories and resource use clarifies how coordination supports task completion, where it introduces regressions, and at what cost.
cs.AI / 14 / 2610.25630
The Forward/Inverse Duality of Controlled Emergence:A Computational Design Program from Explanation to Engineering
Abstract
The macroscopic capabilities of intelligent systems arise from the local interactions of many simple units. When such systems are deployed in the real world, their behaviour must nevertheless be constrained, leaving designers caught between releasing adaptability and preventing loss of control, yet without formal tools to resolve the dilemma. This paper proposes a unified research program for controlled emergence that restates this tension as a pair of dual problems: the forward problem explains macroscopic behaviour from local rules, and the inverse problem solves for controllable rules and readout mechanisms from a desired target. We take an augmented quintuple as the formal object shared by both ends, and decompose ``engineering'' into three independently measurable coordinates: controllability, reproducibility, and degree of formalizability. On this basis we give three testable working hypotheses (critical scale; the cooperation and trade-off between constraint and adaptability; minimal constraint) and one feasibility criterion, which reduce inverse design from unconstrained optimization to interval checking. The program takes explicit failure boundaries as its premise, does not pursue complete control, and always preserves the system's room for self-organization. The reduced pilot experiment attached to this paper serves only to verify the executability of the research protocol and does not constitute a statistical test of the working hypotheses. This design language generalizes to multi-agent artificial intelligence, the industrial internet, and emergency response systems, all of which must confront emergence and controllability at once.
cs.AI / 15 / 2610.25646
Spherical Fixed-Point Reasoners: Adaptive Loop Halting on the Hypersphere
Abstract
Looped (weight-tied) transformers enable adaptive computation by allocating more iterations to harder inputs. Fixed-Point Reasoners (FPRM) achieve this with a damped fixed-point solver, but operate in an unbounded Euclidean residual space with coupled rescaling and relative-norm halting. We introduce Spherical Fixed-Point Reasoners (SFPRM), which perform the same computation on the unit hypersphere of normalized transformers (nGPT). Input injection, damping, and halting become geodesic operations, yielding boundedness by construction and eliminating coefficient coupling. We prove fixed-point existence, local convergence under an explicit step-size condition, and global convergence under strong injection. On Sudoku-Extreme, Maze-Hard, and Mini-ARC, SFPRM achieves 96.6%, 87.9%, and 59.1% test accuracy, respectively, outperforming FPRM and other baselines while allocating test-time compute more effectively. Its converged geometry is also interpretable: token-input angles capture puzzle structure and predict correctness.
cs.AI / 16 / 2610.25673
HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
Abstract
Reliable microbial taxonomic inference is challenged by the rapid growth of newly sequenced organisms and incomplete reference coverage. While alignment- and database-driven tools remain essential for reference-backed identification, they are not designed for learned inference over species absent from genome-level training. We introduce HyperTaxa, a taxonomy-aligned representation learning framework for open-set microbial taxonomic inference. HyperTaxa encodes whole genomes as multiscale hyperspectral k-mer transition fields and aligns these representations to a taxonomy-derived semantic space using a two-stage Node2Vec-to-MDS procedure that combines local graph structure with global hierarchical distance preservation. The resulting space supports zero-shot inference over species excluded from genome-level training but represented in the taxonomy, generalized zero-shot inference over seen and unseen species, and cascaded genus-to-species prediction. Across benchmark and large-scale disjoint-species evaluations, HyperTaxa improves taxonomic prediction over pseudo-image and transformer-graph baselines, with the strongest gains at higher taxonomic ranks. Ablations further show that multiscale transition encoding and taxonomy-aware alignment are key contributors to generalization. These results position HyperTaxa as a complementary taxonomy-aware representation and candidate-generation framework for microbial identification under incomplete reference coverage and open-set uncertainty.
cs.AI / 17 / 2610.25676
Neither Over- nor Under-Constrained: Evaluating LLM Personalization under a Constraint Qualification Audit
Abstract
Accurate evaluation of LLM personalization is hindered by current evaluation paradigms that over- or under-constrain the response space, yielding frequent false rejections and false acceptances. To reduce these false rates, we introduce ConQual-Audit, an audit-based framework that evaluates LLM personalization through a three-stage process: (1) extracting semantic constraints from user profiles and task-grounded reference; (2) qualifying constraints by pruning unnecessary constraints and recovering missing constraints; and (3) verifying candidate responses against the audited constraints via natural language inference. Across PrefEval and LaMP, ConQual-Audit outperforms nine baselines, attaining top AUCs (0.88 and 0.90) and the lowest false rate (0.21) on the leanest constraint sets at an amortizable audit cost. Crucially, this constraint audit is portable: integrating it into ExPerT and PREF helps revise smaller, higher-quality constraint sets to reduce their evaluation false rates.
cs.AI / 18 / 2610.25693
ToolOPD: Credit-Aware On-Policy Distillation for Tool-Use Agents
Abstract
On-policy distillation (OPD) enhances the capabilities of small language agents by providing dense, token-level supervision from a teacher model on student-generated trajectories. However, standard OPD implicitly assumes that credit is uniformly distributed across long-horizon trajectories, assigning the same strength of teacher supervision to different actions. Through credit-distribution analysis and outcome-reversal counterfactual experiments, we find that trajectory credit is concentrated on a small number of critical actions. Moreover, even when the trajectory is held fixed, reversing only the final outcome substantially changes the credit assigned to multiple interaction turns. These findings show that the uniform-credit assumption fails to identify critical decisions and ultimately degrades the model's reasoning accuracy. To address these challenges, we propose ToolOPD, an outcome-aware on-policy distillation framework designed for long-horizon tool use. ToolOPD first summarizes each completed trajectory and its outcome into a reusable hindsight skill. It then measures how this skill changes the teacher's probability of each student action token and converts the change into outcome-aware credit. The resulting credit preserves supervision that supports successful decisions or discourages failed ones, while suppressing irrelevant and conflicting signals. ToolOPD uses these credits to reweight the action-level distillation loss and jointly optimizes it with reinforcement learning. Extensive experiments across multiple long-horizon tool interaction benchmarks demonstrate that ToolOPD significantly outperforms existing reinforcement learning and self-distillation baselines.
cs.AI / 19 / 2610.25707
Whose Credential Is It? How Owner Attribution in Agent MemoryShapes Cross-User Account Misuse
Abstract
Shared long-term memory lets an LLM agent serving one user inadvertently retrieve another user's credentials, but it is unclear when such exposure leads to actual harm. We introduce a three-stage audit that separates three events: a foreign credential reaches the prompt (exposure), the agent acts with it (behavior), and the other user's account changes (state). In AppWorld, a store that mixes 763 past logins from all users retrieves another user's credential for 84.7 percent of required apps, yet agents rarely act on it. Misuse appears in a stress test where memory lacks the user's own credential and presents a single foreign one as the user's record: seven agents then use it in 80 to 94 percent of runs and change the victim's account in up to 43 percent, sometimes passing the official check. Naming the owner before or inside each credential line prevents most of this: with all else fixed, misuse falls from 92 to 1 of 305 paired runs, also with name-free usernames, though owner-bound notes cost 3 to 5 successful tasks per model on clean memory. A wrongly written owner largely restores misuse; trailing owner tags and added safety instructions help only some models. Agent memory should consistently record whose credential it stores, and evaluations should measure behavior, not only exposure.
cs.AI / 20 / 2610.25717
Fine-Grained Units, Flat Decisions: Hierarchical Micro-Expert Pruning for Mixture-of-Experts Models
Abstract
Micro-expert pruning reduces the capacity of mixture-of-experts (MoE) models by removing coupled channels within routed experts. However, finer pruning granularity alone does not guarantee the preservation of downstream capabilities, as similar language-modeling perplexity can accompany substantially different task performance. We revisit micro-expert mask construction as two distinct decisions: allocating retained capacity across experts and selecting channels within that capacity. We propose UC-RAMP, a two-stage post-training pruning framework that explicitly separates these decisions. CAPE uses cross-domain calibration evidence to redistribute expert widths under an exact layer-wise budget, while RAMP refines channel composition at fixed widths through one-out, one-in exchanges guided by a channel-interaction surrogate. Across three MoE models at 25% and 50% routed-expert channel pruning, UC-RAMP achieves the highest ten-task mean accuracy among evaluated pruning methods in five of six settings. At 50% pruning, it outperforms the strongest evaluated baseline by 8.19, 3.96, and 6.14 percentage points across the three models, respectively. In a single-A100 deployment at 25% pruning, UC-RAMP reduces measured time to first token by 19.7--23.9% and time per output token by 20.2--22.3% relative to the unpruned model. The code is available at https://github.com/xzktx003/UC_RAMP.
cs.AI / 21 / 2610.25733
TimelineAgent: An LLM-based Framework for Query-Centric Event Timeline Extraction Across Multiple Documents
Abstract
Query-centric cross-document event timeline extraction aims to construct event timelines from multiple documents under diverse query focuses. The task is very challenging due to the heterogeneous event modeling and reasoning requirements arising from both the intrinsic complexity of cross-document timeline construction and the diversity of query focuses. In this paper, we propose TimelineAgent, an LLM-based multi-agent framework featuring a multi-stage reasoning workflow, where heterogeneous event features are organized into reusable skills and orchestrated through a goal-oriented protocol. To facilitate evaluation, we construct a benchmark covering diverse query focuses. Experimental results show that TimelineAgent consistently outperforms existing methods, offering a practical framework for complex real-world timeline construction.
cs.AI / 22 / 2610.25764
ActTime: Action-Conditioned World Modeling for Time Series
Abstract
Time-series models increasingly condition on known future inputs. In controlled systems, some of these inputs are actions that can be changed, compared, and optimized before execution. A model must therefore do more than predict the realized future: it must predict how outcomes change under different actions from the same observed history and support reliable comparison of candidate plans. We introduce ActTime, an action-conditioned world model for time series. Over the observed history, ActTime updates variable-specific latent states through an action-conditioned transition and corrects them using observations. It then rolls the same transition forward under candidate actions to generate future trajectories. Training uses only factual trajectories, without paired alternative-action outcomes or planning supervision. We construct a simulator-grounded evaluation suite across building energy, traffic, greenhouse climate, and inventory management, covering factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. ActTime ranks first in 12 of 16 domain-task settings. Across eight methods, the mean pairwise model-ranking reversal rate is 17.8% between factual prediction and intervention fidelity, and rises to 28.4% between intervention fidelity and decision ranking. These results show that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable.
cs.AI / 23 / 2610.25765
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
Abstract
Accurate inpatient risk prediction requires modeling heterogeneous clinical evidence that accumulates irregularly over time. This challenge spans in-hospital mortality, prolonged length of stay, and readmission, each with task-specific observation windows. Existing multimodal approaches combine structured health record data, chest x-rays, and clinical text, but primarily focus on modality fusion or missing-modality compensation, with limited attention to temporal trajectories over an admission. We propose TRACE (TRAjectory modeling of multimodal Clinical Events), a daily-scale framework organizing clinical events into day-level trajectories, separating day-invariant global context from day-specific local events drawn from EHR data, chest x-rays, radiology reports, and discharge summaries. The trajectory is encoded using temporal attention, allowing unavailable modalities to emit no tokens. A phenotype-guided contrastive objective further organizes admissions under heterogeneous disease distributions. Across cohorts from MIMIC-IV, eICU, and MIMIC-III comprising nearly 180,000 total stays, we evaluate TRACE on mortality, length of stay, and readmission prediction. TRACE achieves the highest AUROC across all nine dataset-task settings and the highest ACC in eight, outperforming baselines. Ablation studies validate our attention masks and contrastive learning, while trajectory-level attribution localizes risk-associated hospital days and modalities.
cs.AI / 24 / 2610.25767
EvalRx: Agentic Diagnosis and Verified Repair of Foundation Models
Abstract
Model evaluation reveals where models fail, but rarely why or whether a proposed fix genuinely works. We introduce EvalRx, an open-source agentic framework for statistically grounded model diagnosis and repair. EvalRx discovers systematic failure modes, explores candidate mechanisms, formulates falsifiable hypotheses, and verifies them by executing frozen, pre-specified tests before using supported diagnoses to guide tiered repair. Candidate repairs are validated against the unchanged model and promoted only when they improve the target behavior without unacceptable regression. Across language, vision-language, and audio-language settings spanning Gemini, Qwen, Gemma, and Nemotron, EvalRx improves performance in all 12 evaluated model family-modality combinations, with gains ranging from +1.4 to +10.9 percentage points (+5.8 points on average). These results show that model evaluation can serve not only to measure failures, but to support an auditable process for diagnosing and repairing them.
cs.AI / 25 / 2610.25776
From Trace to Fix: Residual Task State in Long-Horizon Agents
Abstract
Recovering from failures in long-horizon, stateful language agents is challenging because the action introducing a defect, the symptom exposing it, and the persistent state requiring repair frequently diverge. Existing methods largely focus on retrospective failure attribution, leaving sequential diagnosis and state repair poorly integrated under constrained budgets. We introduce TRACE-FIX, a unified recovery framework centered on versioned residual task state. Trace transforms execution logs into a structured dependency-linked ledger, prioritizing unresolved task requirements rather than salient surface errors. Fix formulates recovery as budgeted planning over admissible actions, incorporating outcome forecasting and explicit repair-reserve constraints to prevent premature budget exhaustion. To rigorously decouple localization accuracy from downstream repair capability, we present LongH-Debug, evaluating root-cause attribution, verified-fault recovery, and end-to-end closed-loop execution independently. Our empirical analysis demonstrates that structured residual tracking substantially improves causal localization over raw-trace baselines, while budget-aware sequential planning converts verified fault signals into more reliable task recovery without destructive trial-and-error. Code is available at https://github.com/toffeecat/tracedp.
cs.AI / 26 / 2610.25779
OpenSaC: A Python Library for Search as Code in Deep Research Agents
Abstract
Deep research agents investigate complex questions through iterative search, source examination, and evidence synthesis. A conventional workflow alternates reasoning with individual search and reading actions. Search as Code introduces a new paradigm: agents write programs that compose search, source reading, reranking, model calls, and local processing, returning selected observations for further reasoning. To support this workflow in existing coding agents, we present OpenSaC, an open-source Python library with a companion Search as Code skill. Configurable capability interfaces let agents combine these operations while retaining evidence for inspection and reuse. Compared with standard Codex using its built-in web search, the OpenSaC configuration uses less execution time and fewer model tokens on BrowseComp and WideSearch, with higher accuracy on BrowseComp and slightly lower row F1 on WideSearch. We aim to lower the barrier to adopting Search as Code and provide an open foundation for community development and transparent research on deep research agents.
cs.AI / 27 / 2610.25788
Can LLM User-Proxies Stand In for Humans in Reliance Measurement? Replicating Rel-A.I. with Simulated Users
Abstract
Can language-model user proxies reproduce human reliance on an AI answer across contexts? Published simulated-user studies do not establish validity for this partial-answer decision, and Rel-A.I.'s released human record omits trial-level choices. We audit its 90 printed questions and 15 domain-by-marker reliance rates using sampled decisions and counterbalanced A/B probabilities from open models through 72B parameters. A matched crossover varies the printed agent specialty within each question, separating question content, label averages, and their interaction. Mistral Small's strong-marker math-minus-law label marginal is -0.4 percentage points, yet matching the agent label to the question domain raises reliance by 13.3 points on average across the five domains. The best A/B proxy has 22.5-point absolute-rate error and 8.3-point domain-effect error against the human table. A two-sample independent-binomial human replication has 3.5-point expected domain-effect error under the stated working model. The interaction is model dependent, and inference layout changes identical-prompt probabilities. This single-study audit does not establish proxy validity for human effect sizes or general directional screening.
cs.AI / 28 / 2610.25795
SAR: Learning Global Item-Space Matching and Geometry-Aware Placement Ranking for Buffered Online 3D Bin Packing
Abstract
Online 3D bin packing with a bounded buffer requires irreversible item-selection and placement decisions under limited knowledge of future arrivals. Feasibility alone is insufficient, as feasible actions can produce different residual-space structures and subsequent packing opportunities. A central architectural question is how to allocate spatial reasoning between item selection and placement evaluation. We propose Spatial Affordance Retrieval (SAR), a novel learning-based framework that coordinates global item-space matching with geometry-aware placement ranking. A voxel-token spatial encoder represents 3D container occupancy, enabling a picker to match buffered items to the residual space without requiring item-specific placement proposals as inputs. After item selection, a loader ranks feasible rotation-position candidates using candidate geometry, local geometric features (LGF), and inter-candidate comparison, without directly accessing global spatial tokens. The feasible candidate set connects the two stages, which are jointly optimized under a shared long-horizon packing return. On the Random, JD, and SF datasets, the best SAR variants improve utilization over the strongest evaluated online baselines by 0.96-4.20 percentage points under the no-buffer setting (K=1). With a buffer (K=5), the best utilization rates across SAR variants reach 93.55%, 91.04%, and 90.41%, respectively. Ablations support the complementary roles of global item-space matching and geometry-aware placement evaluation, counterfactual analysis suggests that learned item selection favors greater future packing potential, and experiments demonstrate compatibility with multiple candidate generators.
cs.AI / 29 / 2610.25827
When Retrieval Partially Conflicts with Skills: Understanding Procedural Hallucination in Language Models
Abstract
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing procedure together with retrieved task context. When the retrieved context is useful but contains cues that violate the skill, the model must retain compatible information while rejecting the conflicting cues. We study failures in which the model adopts such cues and produces a retrieval-supported but skill-invalid answer, which we call procedural hallucination. We introduce SourceClash, a matched benchmark with 1,271 bundles and 6,355 prompts across ten procedural domains. Each bundle fixes the task, skill specification, and answer space while varying the relation between the skill and retrieval. Across nine models, valid-answer rates average 25.7% under conflicting retrieval, compared with 78.2% and 72.8% under two supportive retrieval settings. Mechanistic analysis of open-weight models shows that the answer-preference difference typically reaches its maximum shortly after generation begins. Residual replacement moves the model toward valid answers, while attention decomposition attributes most of the repair to value vectors. Intervention at the identified value pathway also improves final answer validity. These results characterize how partly conflicting retrieval changes skill-guided generation.
cs.AI / 30 / 2610.25834
FrontierRefactor: Benchmarking Multi-Scale Performance-Oriented Codebase Refactoring
Abstract
Autonomous codebase refactoring requires coding agents to restructure working software while strictly preserving externally observable behavior and prioritizing execution performance and compute efficiency over stylistic preference. Existing benchmarks: refactoring benchmarks assess functional correctness while overlooking runtime regressions, whereas code performance benchmarks target isolated micro-kernels that ignore codebase-scale complexity. We observe that refactoring scale is governed by the scope of the observable contract, the caller-visible interface fixing behavioral invariants while freeing private implementations for optimization. Building on this insight, we introduce FrontierRefactor, a benchmark of 60 qualified engineering tasks spanning four scales, from computational kernels to full repositories, across in-place optimization and cross-language rewriting. Evaluating frontier coding agents through a reference-free cascaded protocol across 1,080 execution runs reveals that correctness and speedup are largely decoupled: 41.2% of submissions satisfy functional equivalence yet fail to deliver verified performance improvements, with 19.3% suffering performance regressions and 21.9% trapped in measurement noise. Refactoring paradigms reveal orthogonal engineering bottlenecks: while cross-language rewrites gain speedups from native compilation after passing functional checks, agents struggle with contract fidelity, with pass rates dropping from 65.0% on modules to 32.6% at package scale; conversely, in-place optimization reliably preserves behavioral contracts at 86.5% but rarely achieves acceleration without physical performance profiling. Trajectory analysis shows that agents fall victim to uncalibrated micro-benchmarking illusions, an operational blind spot that multi-turn execution-timing feedback directly resolves, raising strict success from 16.0% to 40.0%.
cs.AI / 31 / 2610.25842
Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation
Abstract
On-policy self-distillation (OPSD) supervises student-generated trajectories with a reference-conditioned version of the same base model. We ask whether every trajectory supplies distinct supervision that warrants its distillation cost. Gradient measurements on Qwen3-1.7B at initialization reveal that correct and incorrect responses have similar individual gradient magnitudes, but incorrect responses have approximately three times the coherent gradient share. Their estimated population-mean directions also largely overlap. A signal--noise model fitted at a small group size tracks alignment at larger group sizes without refitting. Motivated by these measurements, we introduce correctness-filtered OPSD (C-OPSD), which generates one response per problem and distills only a subset of incorrect responses. Removing unselected rows before both distillation forwards avoids their teacher/student forward and student backward work. Across Qwen3-1.7B, 4B, and 8B on AIME24, AIME25, and HMMT25, C-OPSD achieves on-par performance with full OPSD. In a measured micro-batch-size-four implementation, executed rows decrease by approximately 68%, and with retained rows packed into full micro-batches, training takes about 40% less time while rollout generation remains unchanged. Together, the measurements and training results connect outcome-conditioned supervision redundancy to more compute-efficient distillation.
cs.AI / 32 / 2610.25852
MASIF: Implicit Failures for Process Observability in Multi-Agent Systems
Abstract
Multi-agent systems (MAS) can complete a task despite local execution errors that are corrected, bypassed, or carried forward without individually determining its outcome. We call these implicit failures (IFs) and study them as signals for process observability beyond final task outcome. Across 1,828 traces from nine MAS-benchmark configurations, IFs occur in 54.3% of successful and 75.2% of failed traces. We develop taxonomies of error types and subsequent absorption mechanisms, supported by human and cross-model annotation audits. Counterfactual replay further probes how repairing an IF changes the final outcome. To examine the practical value of IFs, we evaluate IF-based best-of-N trace selection against existing process reward models, and introduce a novel RL algorithm, IF-GRPO, that uses IF costs to refine preferences among successful trajectories in GRPO. Experiments show that IF-GRPO exhibits some gains in accuracy and process metrics. Importantly, it exposes trade-offs between task performance, IF count and cost, and token-normalized IF density. Overall, our results show that IFs provide complementary process observability into how MAS reach their outcomes, including error recovery, propagation, and IF-based process differences that final task success alone does not expose.
cs.AI / 33 / 2610.25865
Minimum-Cost Robust Evidence Certificates for Horn Reasoning under Correlated Source Failures
Abstract
A small justification for a rule-based conclusion may disappear when a source is withdrawn, and multiple derivations need not provide independent evidence. We study minimum-cost certificates that preserve a Horn conclusion under every withdrawal of at most k source domains. Evidence records remain distinct from asserted facts, supporting replicated attestations and overlapping failure dependencies. Transversals of minimal-support footprints characterize robustness and explain why disjoint-proof requirements can reject valid certificates. An exact procedure alternates weighted evidence selection with a source-failure oracle. A rule-closed counterinterpretation yields a globally valid selection cut without enumerating supports or failure scenarios. We prove correctness, finite termination, and cost optimality, establish coNP-completeness of unrestricted verification and NP-completeness of synthesis for every fixed budget, and identify a tractable reachability restriction. Experiments on controlled benchmarks, reachability graphs, and ontology-derived instances demonstrate that our counterinterpretation-guided synthesis achieves up to 92% optimum coverage within practical resource bounds, outperforming scenario expansion and general abduction while confirming the necessity of transversal-based reasoning over proof-packing heuristics.
cs.AI / 34 / 2610.25867
Benefit-Targeted Routing and Bounded Revision for LLM-Based Evidence Screening in Manufacturing Supplier Discovery
Abstract
An evidence-review system must allocate additional model judgments and control their use. We study these software decisions on 6,164 frozen manufacturing source-task pairs. Benefit-targeted routing learns recovered agreement minus newly introduced disagreement. On the original 300-case test, 60 consultations yield 202/231 agreements with two human references versus 189/218 for a logistic-risk router. An additional 300 cases were independently annotated twice using equal frozen evidence. Without retraining the historical policies, the corresponding counts are 164/190 versus 163/193; the differences do not resolve an advantage. Strict scalar-Ridge controls share preprocessing, regularization and fitting, and remain inconclusive. Quotation and requested-state replacement gates retain rejected cases and budget costs; their joint gate rejects one original and two additional consultations without changing primary outcomes. Evidence ablations, fusion and bounded updates supply further controls. A completed 200-pair identity audit yields 97.45 percent weighted merge precision after eight author adjudications. This audits a provisional map, while prospective company-disjoint performance remains untested.
cs.AI / 35 / 2610.25871
CODA: Cohort- and Drift-aware Foundation Model for Multimodal Clinical Reasoning
Abstract
Recent advances in multimodal foundation models have demonstrated strong performance across diverse clinical tasks. However, existing approaches predominantly operate at the individual sample level and fail to model the interplay between population-level cohort patterns and patient-specific variation that underlie real-world clinical reasoning. In practice, clinicians routinely reason through patient cohorts sharing similar demographic, physiological, or pathological characteristics, while accounting for patient-specific deviations from the group prototype. Motivated by this gap, we propose CODA, a cohort- and drift-aware foundation model for multimodal clinical reasoning. CODA explicitly integrates cohort structure into the modeling pipeline through two complementary mechanisms: cohort-aware population stratification and drift-aware individual adaptation. In particular, we leverage reinforcement learning to discover clinically meaningful patient cohorts, and further model each patient's deviation from cohort prototypes as a structured drift signal encoding fine-grained individual heterogeneity. These cohort- and drift-aware representations are incorporated into the foundation model to enable reasoning that is simultaneously informed by population-level patterns and sensitive to individual-level variation. Extensive experiments on real-world electronic health record (EHR) benchmarks demonstrate that CODA achieves consistent effectiveness and strong generalizability across diverse clinical tasks, including closed-QA, open-QA, and report generation.
cs.AI / 36 / 2610.25893
Motor compound fault diagnosis using an attention-based cross-talk architecture
Abstract
As physical AI increasingly relies on sensing to understand physical systems, reliable condition monitoring of industrial machinery has become more important. Motors are ubiquitous in industrial systems, and increasing system complexity raises the likelihood of simultaneous faults. Accordingly, compound fault diagnosis that identifies multiple fault components and their severities is increasingly required. Because different fault components jointly affect the same vibration signal, their severity classification can be formulated as related tasks, making multi-task learning (MTL) suitable for exploiting shared information. However, conventional shared-trunk architectures force tasks to share a common representation, potentially causing negative transfer because fault components are characterized by different time-frequency regions. To address this issue, this study proposes a Residual Neural Dimension Reductor with Time-Frequency Channel Attention (RNDR-TFCA), an attention-enhanced cross-talk architecture for multi-output compound fault diagnosis. RNDR-TFCA preserves task-specific feature extractors while selectively exchanging information across tasks. Time-frequency attention emphasizes informative intra-task regions, while channel attention enhances useful inter-task information after nonlinear feature fusion. Experiments on a rotor-bearing system with inner-race fault, outer-race fault, misalignment, and unbalance showed that RNDR-TFCA achieved a macro F1 score of 0.898, outperforming single-task, multi-class, shared-trunk, and cross-talk baselines. Ablation experiments confirmed the benefit of jointly applying both attention mechanisms. RNDR-TFCA required 713 K parameters and 170.42 MMACs, only slightly higher than RNDR. Channel-expanded baselines still underperformed RNDR-TFCA, indicating that the gain was not simply attributable to increased model capacity.
cs.AI / 37 / 2610.25896
When Should AI Chime In? Benchmarking Timely Contribution of Proactive Agents in Multi-Party Chat
Abstract
Proactive agents can support group collaboration by surfacing useful information without waiting to be asked, but even relevant contributions can be disruptive if poorly timed. In this work, we isolate the timing question: given a fixed candidate contribution such as a clarification or relevant fact, when should an AI agent chime in? We introduce PITCH (Proactive Intervention Timing in CHat), to our knowledge the first benchmark with large-scale human judgments of proactive agent timing collected from participants in the same unscripted multi-party conversations they took part in. To collect high-quality judgments, we design a novel interface in which 2-4 people chat synchronously, author candidate AI contribution messages, place their own and others' contributions into the chat history, and then deliberate as a group about the best placement for each message. PITCH contains 384 candidate contributions and 1,091 judgments, each marking an ideal turn, an acceptable window, or that a contribution should not appear, from 89 participants engaged in 35 conversations. Evaluating 15 models across 28 configurations, we find that while models can place contributions well when given the entire conversation history, they systematically chime in too early when asked to decide live as the conversation unfolds: best-turn accuracy falls from 57.8% to 19.0%, with nearly half of placements occurring before every participant's acceptable window. In comparison, lay humans deciding live are more accurate (23.7% vs 19.0%), and fewer of their placements are premature (39.7% vs 46.4%). Model reasoning traces show that 93% of intervention justifications cite relevance, while almost none consider waiting. By providing a human-grounded benchmark, our work reveals gaps in agents' ability to integrate into group settings and points towards proactive agents that can reason explicitly about appropriate timing.
cs.AI / 38 / 2610.25930
TESS: LLM-guided Translation, Exploration, Simulation, and Search for Embodied Task Planning
Abstract
In this work we focus on the problem of embodied task planning, which requires an agent to accomplish a complex task by separating it into atomic actions. Recent studies have developed LLM-based methods to address this problem. However, the executability and correctness of the generated plans remain suboptimal in complex environments and tasks. On the other hand, classical search algorithms have the potential to solve the planning problems perfectly, yet they require fully observable environments, known state transitions and goal state, which are often unattainable in general task domains. Therefore, we propose to implement an LLM-guided search algorithm that makes the search process possible and efficient by incorporating LLMs' commonsense knowledge. Our method leverages LLMs to (1) translate natural language instructions into formatted goal conditions, (2) facilitate efficient exploration in partially observable environments, (3) implement a rule-based world simulator, and (4) prune the space of candidate actions to improve search efficiency. Experimental results on VirtualHome and ALFWorld demonstrate its superior performance.
cs.AI / 39 / 2610.25931
EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver
Abstract
EvoSkill v2 is a three-agent system that edits the instructions (skills) an agent is given, leaving its model and harness intact. The user starts a run by describing a benchmark and supplying a grader, which a benchmark adaptor turns into a benchmark-specific plugin. A solver attempts the tasks, and an evolver reads the solver's training rollouts and rewrites the skills at each epoch. We implement the evolver as one stateful coding agent with a bash tool for the whole run, unlike the two most common designs among the skill optimizers we surveyed: a sequence of single model calls, as in SkillOpt, and an agent that starts afresh each round, as in SkillBoost. We also measure how much the solver's rollouts vary from one attempt to the next, give the evolver several attempts per training task, and accept every epoch's skills instead of selecting among them on noisy scores. With fine-grained ground truth, EvoSkill v2's skills raised the held-out pass rate above the stronger of SkillOpt and SkillBoost by 7.7 percentage points on SpreadsheetBench v2 and by 5.4 on $τ^3$-bench banking. In both benchmarks and both feedback settings, EvoSkill v2's evolver cost less per percentage point of held-out pass rate gained than either baseline's optimizer. EvoSkill v2 also enables a less capable model to improve its own skills, demonstrated by a gain of up to 21.5 percentage points in the $τ^3$-bench banking experiment where Claude Haiku 5.5 powers both solver and evolver.
cs.AI / 40 / 2610.25941
Global probabilistic forecasting of air pollution
Abstract
Fine particulate matter (PM2.5) refers to particulate matter with an aerodynamic diameter of 2.5 micrometre or less. It comes from human activities, such as traffic and industry, and from wildfires, desert dust and gases emitted by vegetation. Short-term exposure to it accounts for about a million deaths a year. Operational forecasts run on grids tens of kilometres wide, which hide the pollution breathed by nearly two hundred million people a day. Machine learning has transformed weather forecasting, but data-driven forecasts of air pollution still blur extremes and lose skill as the air travels. Here we introduce Foehn, a generative model that learns from global kilometre-scale estimates of PM2.5 and computes where the air comes from. Over a year of retrospective forecasts, it outperforms the operational forecast at three stations in four and forecasts at the scale of this hidden exposure. Its ensembles are calibrated, and their mean detects local pollution extremes four days ahead as well as the operational forecast does one day ahead. Such forecasts could warn the millions whom coarse maps miss, once kilometre-scale data are available in real time.
cs.AI / 41 / 2610.25947
Claim Laundering: When Agent Decision Models Trust Records Over Reality
Abstract
Agent decision models select actions in a single forward pass from inputs that combine environmental observations with the agent's own record of previous actions. When an action fails silently, these sources can conflict: the record reports success, but the environment shows that the step did not take effect. We ask which source determines the next decision. Across Android screen traces, ALFWorld, and WebShop, we construct paired states that independently vary the observation, the logged action, and an explicit success claim. In all 12 configurations of four decision models and three domains, records have a larger average effect on commitment decisions than observations; after claimed-but-failed steps, the models still stop, advance, or purchase in 53-100% of cases. Yet the relevant evidence is often represented internally: linear probes on intermediate hidden states distinguish successful from failed steps on held-out apps, tasks, and goals (AUROC 0.77-1.00), and activation patching reveals increasing influence from records in later layers. We use this signal in EVIGATE, an intermediate-layer veto that requires no additional forward pass and adds 0.04-2.0% serving latency. Under injected silent failures, EVIGATE improves task success by 18.7 points in live ALFWorld agents, 5.0 in WebShop, and 5.1 in a phone simulator, and in WebShop and the simulator it also outperforms random vetoes at the same rate. Its effectiveness depends on representative calibration states: gates calibrated on offline counterfactual states do not transfer reliably to live agent trajectories.
cs.AI / 42 / 2610.25954
CXBench: A Benchmark for Process-Compliant LLM Agents in Regulated Customer-Experience Workflows
Abstract
Tool-using language agents increasingly perform regulated customer-service workflows, where a correct final state is insufficient: agents must also provide disclosures, obtain consent, and follow required procedures. Existing tool-agent benchmarks largely score final business states, accepting trajectories that reach correct outcomes through non-compliant processes. We introduce cxbench, a benchmark that evaluates outcome and process compliance separately across six regulated domains. Each task combines a customer request, a tool-based environment, and explicit process requirements. Rule cards link cited regulatory obligations to observable trace evidence. A hybrid grader combines deterministic checks for ordering and authority with an LLM judge for semantic requirements. We define a false-pass as an outcome-pass run that fails process checks. Across seven models and 120 tasks, 55.5% of the 840 retained executions are false-passes, representing 80.1% of outcome-passing runs; only 13.8% pass both checks. Process-inclusive evaluation changes the observed model ordering and exposes differences hidden by outcome-only scoring. cxbench makes these failures inspectable through trace-linked criteria.
cs.AI / 43 / 2610.25959
EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
Abstract
Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence. Yet most existing work remains confined to specific tasks in a single scene and transfers poorly to new environments, leaving general embodied agents for open-world settings under-explored. We refer to such agents as Embodied World Agents (EWAs), whose defining property is the ability to understand diverse visual environments and execute general actions in open-world conditions. To advance EWA research, we study open-world embodied navigation-planning and release EmbodiedWorld-200K, a large-scale dataset of over 200K samples spanning diverse scenes. Each example is equipped with multi-level annotations, ranging from low-level camera-motion trajectories to high-level semantic instructions. The dataset also provides unified evaluation splits and task-specific metrics, supporting offline pretraining and fair evaluation of EWAs. Beyond offline supervision, a defining characteristic of an EWA is its interaction with the environment; training solely on offline data is therefore insufficient, especially for generalizing to scenes not covered in the training distribution. Building dedicated real interactive environments, however, is often prohibitively expensive. To address this, we propose a novel training scheme that employs a video world model as a low-cost environment proxy and performs interactive reinforcement-learning (RL) post-training of the EWA within it. Empirically, pretraining on EmbodiedWorld-200K substantially improves the open-world embodied planning ability of the EWA, and further RL post-training inside a world model yields clear gains on out-of-distribution scenes. The dataset, baseline checkpoints, and training and evaluation code are available at https://xiaokunfeng.github.io/EmbodiedWorld-200K/.
cs.AI / 44 / 2610.25963
ABACUS: Program-Guided Multistep Reasoning over Frozen Judgment Models
Abstract
Judgment models answer semantic questions by scoring a finite set of candidate outcomes. Can reusable programs turn these local decisions into powerful multistep reasoning? We introduce \method, a framework that jointly constructs judgment schemas and composition programs for frozen judgment models. A generative teacher designs the procedure offline; the judgment model resolves local semantic questions online, and the saved program composes its decisions into an answer. We study three allocations of design effort between humans and the teacher, focusing on construction through a shared interface. We instantiate the framework with Kev. On three selected BIG-Bench Hard tasks, nine independently constructed programs solve all 450 executions of 150 development inputs. A separate deployment study freezes three programs after development-set reference-agreement refinement and evaluates them on 600 held-out questions. These programs achieve near-perfect 99.67\% macro accuracy, versus 64.00\% for direct Kev and 68.80\% for final-answer-only Qwen3.5-4B. They attain comparable observed accuracy to Qwen's unrestricted response policy (99.07\%) with an \textbf{8.43-fold online speedup}: 0.255 versus 2.150 seconds per question. These results show how reusable programs can substantially expand the reasoning capability of a compact judgment model while keeping its parameters fixed.
cs.AI / 45 / 2610.25970
GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories
Abstract
Chain-of-thought (CoT) reasoning can improve problem solving, but models often keep generating after the correct answer is already reachable; the extra steps raise serving costs and may replace a correct intermediate result with an incorrect one. We propose GLOBE (Geometry-Led Online Behavioral Early Exit), an online, training-free algorithm that terminates the model's CoT process early. GLOBE embeds each completed reasoning segment as a point on a unit sphere and reads the shape of the resulting semantic trajectory. GLOBE first applies a trial phase that generates short trial answers in a separate branch, stopping the main stream on an accepted trial after checking confidence against a threshold. If generation continues, a trajectory phase stops the stream once this path repeatedly reverses direction with little net progress. On five benchmarks including mathematical reasoning, tool calling, and agent memory, GLOBE cuts output tokens by 49.7% and active end-to-end wall time (E2E) by 51.0% across two models, and achieves a five-task average of 80.58%, comparable to Full CoT.
cs.AI / 46 / 2610.25977
On the Benefits of Uncertainty Estimation for Species Classification: A Case Study on Optimizing Expert re-Annotation
Abstract
Automated identification of diatoms from microscopic images is a cornerstone of large-scale freshwater biomonitoring, yet the high diversity of species and strong class imbalance pose significant challenges for reliable classification. In this study, we develop a ResNet-18 baseline for a 759-class diatom dataset and investigate how calibrated confidence estimates support selective classification to reduce expert workload without sacrificing taxonomic integrity. While the initial model exhibits overconfidence with an Expected Calibration Error (ECE) of 0.117, post-hoc temperature scaling (T = 2.12) dramatically improves reliability, lowering the ECE to 0.055.By leveraging these calibrated probabilities within a triage framework, we demonstrate that a Selective Macro-F1 score of 99.01% can be maintained at 66.2% coverage (rejection threshold $τ$ = 0.934). Qualitative analysis reveals that the system effectively functions as an intelligent filter, consistently deferring rare taxa-such as AAMB and AASI-and ambiguous morphologies for expert review. Our findings, supported by the torch-uncertainty framework, highlight the potential of simple calibration techniques to make automated diatom identification practical and robust, prioritizing the preservation of rare species metrics essential for ecological assessment. To support reproducibility, our code, environment specifications, and calibration results are available anonymously at https://anonymous.4open. science/r/greenprwksp-B53D/README.
cs.AI / 47 / 2610.25978
GraphEvoResearcher: Graph-Aware Training-Free Self-Reflective Reasoning for Cost-Bounded Multi-Hop Question Answering
Abstract
Multi-hop question answering (QA) requires models to chain intermediate conclusions while remaining cost-aware. Existing training-free self-reflective protocols (generate -> critique -> revise) improve accuracy but operate on a flat token stream: revisions cannot identify which prior step is unreliable, and early stopping cannot exploit structural convergence. We introduce GraphEvoResearcher, a training-free protocol that lifts EvoResearcher's self-reflective meta-reward onto an explicit reasoning graph - nodes are intermediate conclusions/evidence and edges encode support, contradict, and depends_on. A graph-aware meta-reward combines correctness, efficiency, reflection, diversity, and four structural signals (connectivity, acyclicity, branch coverage, path convergence) to decide whether to revise, backtrack, or stop. A graph-convergence early-stop terminates when evidence branches converge on a single conclusion. On a real LLM (DeepSeek), over GSM8K, MATH, and BBH (n=50 each), GraphEvoResearcher reaches 98%, 92%, and 82% respectively, the highest of all baselines, with the largest gains on the hardest set (MATH: +24 pp over single-shot, +12 pp over flat reflection), while keeping construction cost bounded at ~2 steps / 4.5-5 nodes. Our results indicate the structural signal of a reasoning graph is most valuable exactly where single-step reasoning is unreliable.
cs.AI / 48 / 2610.25980
ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
Abstract
Long-horizon search requires agents to acquire relevant evidence and integrate it into an evolving research state. Existing methods maintain research states to guide exploration or assign turn-level progress signals, yet grounding these signals in explicit, attributable state changes remains underexplored. We introduce ClueFlow, which learns long-horizon search from attributed research-state transitions. ClueFlow constructs Source-State-Final attribution chains that connect evidence-acquiring browser actions, Evidence and Candidate updates, and the final answer supported by those updates. During search, ClueFlow records typed state changes and potentially supporting browser actions, then applies gold-conditioned semantic validation and source tracing after execution. We use this attribution to guide both supervised fine-tuning and reinforcement learning. During SFT, attribution selects which decisions receive supervision with their original interaction contexts. During Turn-PPO, the same attribution augments outcome-based turn advantages with credit assigned along the Source-State-Final chains. Transition-level validation also recovers useful supervision from gold-consistent state changes in trajectories with incorrect final answers. We evaluate ClueFlow across two Qwen backbones and four search benchmarks spanning live-web and fixed-corpus settings, observing consistent gains from SFT and further improvements from PPO. With Qwen3.5-35B-A3B, ClueFlow achieves 60.0% accuracy on BrowseComp and 72.0% on BrowseComp-Plus. In controlled BrowseComp-Plus experiments, attribution-guided SFT matches all-turn exact-context SFT using about 30% of the training tokens, while attribution-guided PPO outperforms outcome-only PPO by 4.1 pp. These results demonstrate that state-transition attribution enables efficient supervision and effective credit assignment for long-horizon search agents.
cs.AI / 49 / 2610.25991
Map-Distribution Mixture-of-Experts for Multi-Agent Pathfinding
Abstract
Designing capable neural policies for multi-agent pathfinding (MAPF), which map local observations to action distributions, represents an important challenge in multi-agent learning, with practical relevance to scalable collective automation problems such as warehouse logistics. In contrast to prior studies focusing primarily on individual policy design, this study explores a meta-level Mixture-of-Experts (MoE) approach to learning-based MAPF. Specifically, given a collection of compact yet high-performing experts specialised for different map distributions, our architecture learns a router that selectively fuses their outputs according to each agent's local situation. We instantiate and validate this MoE approach with a commonly used graph neural network-based policy as well as a newly developed recursive action-token refinement model. Combined with supervised and reinforcement learning fine-tuning of the experts, our experiments show that MoE generally improves upon the pretrained policy and outperforms existing leading methods. At the same time, our findings reveal that appropriately trained, parameter-matched monolithic models still outperform MoE, highlighting a nuanced role for map-distribution MoE for MAPF.
cs.AI / 50 / 2610.26020
Understanding When Reflective Prompt and Skill Optimization Helps on Locally Served Reasoning Models: A Two-Model Study of GEPA and SkillOpt
Abstract
Reflective text-space optimizers such as GEPA and SkillOpt have reported large gains on frontier and mid-size models. We test replication across two reasoning-native target-serving configurations: DeepSeek V4 Flash through DwarfStar and Qwen3.8-27B through LM Studio. For each target, we evaluate the same held-out tasks and four conditions - minimal baseline, GEPA prompt, SkillOpt skill, and manually engineered prompt - on HotpotQA (n=500 per condition) and MATH-500 (n=250). In the primary Qwen v2 rerun under the same frozen protocol, GEPA selects the HotpotQA seed and exactly reproduces baseline behavior, whereas SkillOpt's validation gain reverses into held-out degradation. On MATH-500, GEPA raises exact match from 0.828 to 0.904 while reducing latency but increasing token consumption, so lower response time does not imply lower token cost. DeepSeek likewise shows its strongest GEPA effect on MATH-500 (0.852 to 0.908). Across both targets, GEPA's strongest gains occur on MATH-500, while neither optimizer improves the HotpotQA baseline. Optimization is therefore task- and method-dependent, and validation gains may reverse on held-out evaluation. An earlier Qwen run is retained as serving-sensitivity evidence. These results suggest that textual headroom for a model-task pair and the observed effect of a text-space optimizer are more informative when decomposed into content/reasoning, answer-realization/evaluator alignment, and serving-completion/reliability. The cross-benchmark pattern further suggests that optimization may benefit from low-ambiguity, actionable evaluation signals.
cs.AI / 51 / 2610.26031
AuDQ: Auditable Data Quality through Agentic Rule Discovery
Abstract
High-stakes tabular pipelines rely on scattered, static quality checks that are hard to evolve and audit against policy intent. Regulated domains particularly suffer as the LLM-era automation raises expectations for traceability and accountability. We propose a vision for an ontology-grounded, repository-native data quality layer in which rules are first-class artifacts: versioned, provenance-linked to governance texts, and re-playable through immutable execution traces. As a first step toward validating this vision, we present a concrete implementation of its agentic rule-discovery core with two rule acquisition modes (Full-Agentic and Guided) and a human-in-the-loop (HITL) review interface. We define an evaluation framework and run experiments addressing whether our fully autonomous, divide-and-conquer agentic rule discovery produces rules of the same quality and diversity as those specified by a human expert, reporting preliminary results on two real world industry datasets in the finance and retail domains.
cs.AI / 52 / 2610.26035
Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems
Abstract
Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation. A knowledge graph writes the logs of each recommendation scenario as short statements on a defined scale, such as "sold: top 10%". A pool of agents, proposes candidates, and a residual ranker keeps the best 30. A decision model, trained once on real choices of three scenarios, reads them through one interface of typed questions, and a fusion step combines its answers with the score of the ranker. Any model with this interface can be used, and we run the commercial Jev, the released open Kev and Kev models trained by us on recommendation requests. The framework with a trained 4B decision model is 25 to 36% above the best trained baseline in NDCG@10 in three sparse Amazon catalogues, one of them never used to train the decision model, and almost tie with baselines on the dense Steam dataset. Most of the gain comes from the pool, and with a trained recommender it comes from what the agents know about its candidates rather than from new candidates. The 27B decision model adds to the pool 5 to 6% above the 0.8B model in three of four domains. A trained 0.8B open model beats the released Jev in eleven of fifteen settings and is never worse. On MovieLens-1M, a dense domain never seen in training, the gain is 5 to 7%. Code, configurations and trained models are available at https://github.com/marcopoli/decision4rec and https://huggingface.co/m-polignano.
cs.AI / 53 / 2610.26066
FinSafeGuard: A Quality-Driven Synthetic Dataset for Financial AI Safety
Abstract
Generating high-quality synthetic datasets at scale remains a fundamental challenge for training specialized AI systems in data-scarce domains. We present a comprehensive methodology that produces 14.3 million quality-filtered synthetic training examples from a structured taxonomy of 632,000 scenario specifications. Our four-stage pipeline ad- dresses critical challenges in synthetic data quality, diversity, and realism through: (1) systematic pre-generation research establishing optimal generation parameters via diminishing returns analysis, (2) conditional auto evolution mechanisms that remediate low-diversity batches during generation, (3) a two-layer semantic deduplication architecture combining lexical and embedding-based methods, and (4) LLM-as-Judge quality evaluation filtering samples based on behavioral correctness and tuple fidelity. Empirical analysis across three domains (Banking, Insurance, Financial Services) demonstrates domain-specific deduplication patterns ranging from 9.9% to 40.6%, with output generation modes exhibiting highest redundancy. Our batch size optimization reveals mode-specific diversity saturation thresholds, enabling principled allocation strategies. The complete pipeline reduces 26.4M raw samples to 14.3M high-confidence examples through merge (20.9% reduction), semantic deduplication (18.7% reduction), and quality-based filtering (15.9% reduction), providing a reproducible framework for domain-specific synthetic data generation with rigorous quality control.
cs.AI / 54 / 2610.26071
CUA-Lite: An Open Platform for Computer Use Agents
Abstract
Open resources for computer use agents (CUAs) -- environments, datasets, and models -- remain scattered: without a standard, these resources cannot be integrated and scaled; additionally, evaluation, RL, and SFT cannot share one infrastructure. We present CUA-Lite, an open platform for developing CUAs, built around three standardized abstractions: (1) Unified environment interface, under which we integrate 12 benchmarks across desktop, browser, and mobile, plus lightweight VM-free sandboxes that replicate OSWorld's desktop at much lower cost, providing over 30,000 verifiable tasks. (2) Unified supervised data format, into which we convert the major public CUA datasets, from GUI grounding and understanding to full rollouts, and record our fresh rollouts from frontier CUAs on these verifiable environments. (3) Model harnesses shared across evaluation, RL, and SFT, one per model and implemented for 14 model families. Each harness runs its model in these computer environments, producing the rollouts that evaluation and RL consume, or renders stored supervised data into the model's own training format for SFT. We release reference scores across the integrated benchmarks, and show that RL and SFT through the same harness improve open-weight models. We release our dataset, code, and models, and the platform is open to contributions.
cs.AI / 55 / 2610.26098
From Code to Playable Systems: CEGame-Bench for Long-Horizon Game Creation Agents
Abstract
A game ships when a player can launch it, finish it, and return to a saved world, none of which a successful build guarantees, yet game-creation agents are mostly tested on one engine or on one-prompt browser games. We present CEGame-Bench, the first long-horizon benchmark for complete-game creation across five production engines: Defold, Godot, Roblox Studio, Unity, and Unreal Engine. Agents build each game from an empty project and a design document in runs averaging up to 295 steps and 65M tokens; the document doubles as a contract, and a hidden evaluator accepts a game only by real-input play. Under four scaffolds, agents earn most integration points but usually under half the gameplay points, and their games reach the ending in only 16% of runs, while references built by engineers with coding agents score near the maximum, which places this artifact-to-play gap in the agents. No two engines rank configurations alike (Kendall's tau 0.14-0.62), and the same game usually scores higher on Unity and Unreal Engine, which check more before play. We release Game Completion Boost, an agent skill for requirement-driven playtesting and repair that raises the execution-based score of three models by 14.1-24.3 points.
cs.AI / 56 / 2610.26101
Reliable Before Profitable: QuantHarness for Governing Financial Strategy Development with LLM Agents
Abstract
LLM agents can turn natural-language financial research requests into executable and backtested strategies, but a successful run may still implement the wrong rule, use information unavailable at decision time, or obscure how repairs and candidate selection changed the strategy. Because these failures arise throughout development, final code and performance metrics alone are insufficient to establish either fidelity to the request or the validity of the development process. We therefore address the reliability of financial strategy development at the trajectory level: the system must control how intermediate outputs become downstream state and preserve the evidence needed to justify the terminal outcome. We introduce QuantHarness, a two-layer agent harness in which a finance-specific Workflow structures specification, data grounding, construction, testing, and review, while Harness Core enforces stage-specific information, action, and commitment boundaries. We also introduce FinRelBench, a benchmark that evaluates strategy-development reliability through execution success, requirement satisfaction, and runtime evidence. QuantHarness achieves the highest Task Success at 90.72\%, compared with 74.23\%--84.54\% for the comparison systems. Controlled stress tests show that its boundaries contain targeted violations, while structured run records improve evidence retrieval and support decision replay under the evaluated protocol.
cs.AI / 57 / 2610.26187
On Counting Graphs for Modal Logic with Counting
Abstract
In this paper, we study the theoretical computational complexity of counting trees and graphs that are classified positively by a graph neural networks, or said equivalently that satisfy a given formula of modal logic with counting. We also study convergence.
cs.AI / 58 / 2610.26211
LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context
Abstract
Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation. Following standard prompting guidance, a user naturally states their role (``I am a lawyer...''). We show that this innocuous habit has an unexpected cost: such authority framing can make the model selectively refuse legitimate requests, introducing bias through uneven access to assistance. We evaluate four modern small open-weight LLMs on benign legal-domain prompts to measure this over-refusal. Surprisingly, we find that authority-style prefixes (''I am legal counsel working for a national supreme court'', ''I'm a defense lawyer'') systematically increase refusal rates by 2--20x over the no-prefix baseline, while a known role-play jailbreak prefix shows mixed effects, sharply increasing refusals in some models and barely shifting them in others. The finding suggests that small on-prem deployable LLMs are unstable under contextual framings that a real institutional user might naturally introduce, and further investigation is essential to minimize opportunities for bias.
cs.AI / 59 / 2610.26215
Fairness or Fallacy? Mapping the Landscape of AI Bias in Disease Detection
Abstract
Artificial intelligence (AI) has several transformative effects on healthcare, like to facilitate clinicians in disease detection and to manage patient care with enhanced precision. Due to the increasing complexity of medical data and the growing volume of diagnostic images, the substantial computational power needed to process such information creates both opportunities and challenges. To accomplish this objective, numerous methodologies have been proposed to integrate AI into disease detection systems. All these methodologies consider different aspects of clinical diagnosis, considering key parameters such as the type of AI model employed, the nature of input data, and the population demographics being analysed. In this paper, we review different state-of-the-art techniques which have been proposed for various areas of disease detection, specifically focusing on bias identification and mitigation. The proposed solutions are distinguished based on the type of input data they utilize, the architecture of model, and type of bias (algorithmic bias, demographic bias, model bias, and deployment bias). The time frame chosen to filter out relevant studies is between 2015 and 2026. Eventually, we review 23 related articles out of 64,389 analyzed from seven different digital libraries. The obtained results guide us to reach a conclusion with respect to state-of-the-art selected studies. Furthermore, these results reveal significant research gaps that can assist researchers in addressing AI-related bias in medical diagnosis. By prioritizing explainability, data diversity, and fairness in algorithmic design, as well as equitable access for affected patient populations, they can utilize available healthcare resources more effectively.
cs.AI / 60 / 2610.26239
CoRS: Fast Neural Flexible Job-Shop Scheduling via Composite Interval State Reuse
Abstract
Neural solvers for the Flexible Job-Shop Scheduling Problem (FJSP) typically refresh graph representations and update feasibility after every operation--machine decision. However, only feasibility must remain exact at every step, whereas graph encoding is substantially more expensive and can potentially be refreshed less frequently. We present CoRS, which decouples them: an interval-level graph representation is reused across multiple sequential decisions, while a stepwise feasibility state is updated after every action. This two-timescale design amortizes graph encoding without parallelizing action generation or relaxing scheduling constraints. CoRS further develops a feasibility-preserving composite-interval decoding scheme with dynamic constraint tracking and boundary state synchronization, which reduces representation refresh frequency while maintaining continuous feasible schedule construction. Furthermore, to alleviate representation staleness caused by interval-level state reuse, CoRS introduces enhanced static features as a compensation mechanism. By providing more comprehensive structural information of the scheduling problem, these features improve the model's perception of overall problem characteristics. Extensive experiments on synthetic and multiple public FJSP benchmarks demonstrate that CoRS significantly reduces inference time while maintaining competitive solution quality, achieving favorable speed-quality trade-offs under tight wall-clock budgets and low-latency scenarios.
cs.AI / 61 / 2610.26241
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law
Abstract
Applying a legal provision frequently requires both structured rule-based reasoning and sensitivity to open-textured terms developed in case law. The actio pauliana of Article 42 of the Dutch Bankruptcy Act (Faillissementswet, Fw), which allows an insolvency practitioner to avoid transactions that prejudice creditors, requires both. We present a neuro-symbolic model of Article 42 Fw. Its symbolic core is written in Catala in a literate style that keeps statute and formalisation aligned and was developed through pair programming, with open-textured terms represented as boolean inputs where neural components couple to the symbolic solver. A neural routine structures a case's natural-language facts into boolean and date inputs. Neural subroutines resolve the open-textured terms through case-based reasoning over Dutch civil judgments, retrieved through a pipeline that filters on the requirement being resolved and expands along the procedural chain of a case. Segmenting and labelling those judgments attains high precision and recall against a manually annotated gold standard of 100 judgments. On ten hand-annotated cases the neuro-symbolic model handles most criteria reasonably but struggles with the knowledge requirement. Retrieval depth shows no clear quantitative effect, although retrieved judgments provide explicit grounding in prior case law.
cs.AI / 62 / 2610.26251
Beyond Static Leaderboards: Capability-Focused Evaluation with Structural Diversity for Time Series Foundation Models
Abstract
Static public benchmarks for time series foundation models (TSFMs) typically have fixed and narrow test distributions. Their repeated reuse encourages benchmark-specific tuning, making reported performance reflect adaptation to the fixed test set rather than genuine forecasting generalization. In addition, a single aggregate score cannot fully characterize model capabilities across distinct temporal structures and input relationships. We propose CaFE (Capability-Focused Evaluation), a refreshable evaluation extension grounded in original benchmark data that enriches static benchmarks with structural diversity and enriches single-score black-box evaluation with multi-dimensional, capability-oriented assessment. CaFE first profiles the original benchmark and defines eight Structural Features (SFs) that characterize its temporal structure. For each SF, it resamples the original data according to its structural support, removes edge-of-distribution samples, and augments the retained samples by varying the SF's salience, producing samples with different structural profiles. After filtering out out-of-distribution samples while preserving diversity, CaFE yields a refreshed evaluation set. TSFMs are then evaluated on this set using both conventional forecast-error metrics and intervention-response metrics, giving each model a per-feature capability profile alongside endpoint accuracy. Experiments on GIFT-Eval and FEV-Mini20 show that CaFE yields stable evaluation results and reveals per-feature strengths and weaknesses that a single aggregate score hides. Moreover, on genuine data filtered by structural condition, models selected by CaFE outperform those chosen by the aggregate leaderboard. CaFE therefore makes model selection conditional on the structural response of interest.
cs.AI / 63 / 2610.26256
LOCU: Lowdin-Orthogonalized Constraint Updates for Multi-Constraint Policy Optimization
Abstract
We investigate multi-constraint policy optimization in constrained Markov decision processes (CMDPs), where interactions among constraint gradients often yield ill-conditioned update directions, causing numerically fragile steps and unintended cancellation or redundant overlap in constraint corrections. To address this, we propose LOCU, which mitigates numerical instability arising from constraint interactions by applying symmetric Lowdin orthogonalization to the natural gradients of the reward and constraints under the Fisher-Rao metric. Combined with Fisher-geometric screening and near-collinear compression, LOCU yields a compact reduced space in which the trust region becomes a Euclidean ball and constraints are handled symmetrically through a low-dimensional active-set solve. For infeasible iterates, LOCU constructs a Pareto-descent direction which preserves the local feasibility of satisfied constraints within the same reduced formulation. Experiments show that LOCU remains stable under high constraint coupling, narrow feasible regions, and infeasible starts, and generalizes across different cost critic architectures.
cs.AI / 64 / 2610.26261
REFLEX: RNA Ensemble Generation via Flexibility-Calibrated Stochastic Bridge
Abstract
RNA molecules often function through multiple experimentally observed conformational states. A central challenge is to capture ensemble diversity without sacrificing structural fidelity: samples may collapse toward averaged representative structures or drift into unsupported conformational regions. We introduce REFLEX (RNA Ensemble generation calibrated by FLEXibility), a sequence-conditioned generator for RNA backbone-frame ensembles based on a heteroscedastic stochastic bridge on SE(3). During training, experimental B-factors calibrate residue-wise bridge widths, exposing each observed conformer through local perturbation neighborhoods with residue-specific scales while keeping constrained regions sharper. A learned flexibility condition guides the bridge-field network to recover target conformer-specific frames from these calibrated noisy states. At inference, REFLEX uses only the input sequence, requiring no MSA-subsampling heuristics. On a multi-conformer RNA benchmark, REFLEX achieves the best precision-recall trade-off for ensemble coverage across evaluation thresholds, while recovery-stress ablations indicate gains in target-frame recovery under stronger perturbations.
cs.AI / 65 / 2610.26310
The Scheduler Is a Hidden Retriever: How Asynchrony Shapes Evidence Acquisition in Search Agents
Abstract
Language-model agents increasingly solve complex tasks through parallel reasoning and asynchronous information gathering. Controlled replay across two model families and three benchmarks shows that delivery timing redirects search decisions made before all results arrive, despite identical initial evidence. Acquisition disagreement reaches up to 24 percentage points above same-order repeat variation. This dependence persists after all initial evidence arrives and through a second search, turning temporary differences in what agents have seen into lasting differences in what they acquire. An acquisition-regret bound with a matching construction shows that invariance at a fixed early arrival count can limit the use of evidence already received. We introduce Confluent Evidence Acquisition (CEA), which instead releases local reasoning when each subquestion's evidence is complete and allocates shared search independently of completion order. For a fixed plan and stable conditional response law, we prove joint invariance of decisions, acquisitions, and answers. With question-only plans across three benchmarks, CEA brings acquisition disagreement to within 1.0 percentage point of repeat variation and achieves accuracy comparable to complete-input controls, improving over early commitment by up to 12.6 percentage points. Replacing CEA's fixed allocation with first-ready allocation under contention adds 21.6 percentage points of acquisition disagreement, while local query-target disagreement remains near repeat variation. Timed comparisons show that CEA reaches decisions earlier and responds faster than the same program waiting for all initial evidence.
cs.AI / 66 / 2610.26311
Selective Transfer through Sequence Distillation: From Benchmark Knowledge to Backdoor Behaviour
Abstract
Knowledge distillation is widely used to transfer capabilities from teacher models to student models, but recent work shows that it can also transfer unintended teacher properties. We study whether different hidden teacher properties survive the same distillation channels. Specifically, we compare two properties: benchmark-associated knowledge and refusal backdoors, and show that distillation treats them very differently. Teachers fine-tuned on benchmark data successfully pass that knowledge to students through all four distillation channels we test, including when students are trained only on teacher-generated number sequences containing no natural language or domain content whatsoever. Refusal backdoors tell the opposite story: backdoored teachers achieve near-perfect triggered refusal, yet sequence-distilled students inherit essentially none of this behaviour. Logit-based distillation transfers it substantially, reaching up to 85.5\% triggered refusal in students that saw zero backdoor examples. Distillation is therefore selective, and the selection criterion is not what data the student sees, but what structure the teacher's property has. The code is available at https://github.com/MJonibek/kd-leakage.
cs.AI / 67 / 2610.26313
StrucChunk: Structure-Aware Retrieval for Legal Document RAG
Abstract
Retrieval-Augmented Generation (RAG) can ground LLM outputs in domain-specific knowledge, but legal documents are hard: standard chunking fragments hierarchical argument structure, and generic embeddings struggle with both precise citation matching and semantic understanding. We propose StrucChunk (Structure Aware Chunking) with hybrid dense-sparse retrieval for hierarchically structured legal documents. It treats documents as directed acyclic graphs of provisions, preserves clause boundaries, resolves cross-references before embedding, and combines BM25 with BGE-M3 via Reciprocal Rank Fusion. We evaluate on five corpora, three legal traditions and two languages. StrucChunk reaches 100% boundary precision (vs. 7.7% for recursive splitting on CrPC), preserves 97.6% of documented cross-references (vs. 0%), and matches section coverage with 40.3% fewer chunks on GDPR. On LegalBench-RAG (6,858 query-passage pairs), it achieves Char-Recall@8 of 48.18% vs. 23.59% for the best baseline (+104%), surpassing the published state of the art by +83%. On 501 queries each for CrPC and CPC (Indian law), Recall@5 improves from 0.557 to 0.804 (+44.4%) and from 0.158 to 0.259 (+64.6%); chunking accounts for 88% and 73% of the gain. On 501 GDPR queries, Recall@5 is 0.868 with R@1 improving +59%. On NitiBench-CCL (3,730 Thai financial law queries), HR@5 is 0.856 vs. 0.594 for naive chunking (+44.1%), showing that section-boundary alignment is the dominant factor across languages, and cross-reference augmentation adds +2.3% on cross-reference queries. We release the implementation, 5,256 evaluation queries across five legal corpora, annotation methodology and chunking quality metrics. Code: github.com/SinghSuryaDeep/legal-rag-strucchunk-1
cs.AI / 68 / 2610.26324
The Control Layer Behind Rule-Following Agents Trading Across Market Conditions
Abstract
Trading discipline often fails people. Our study tested whether four AI trading agents followed their strategies and risk rules. Offering traders the evidence and an agent evaluation toolkit to separate luck from following the rules, before giving agents meaningful autonomy or capital. Our study analysed 2,627 closed positions from 12 February to 27 August 2026 across the same asset universe and execution conditions. Most positions were simulated against live prices with a 0.09% round-trip fee; 18% involved real Hyperliquid orders. Each position declared a strategy, stop and target, with a limit of five open positions per agent. The fleet finished net positive after fees. Rule adherence was high across all four agents, including the unprofitable one. Profitability tracked how often agents won relative to their break-even win rates. Institutional-style controls ran by default and contained losses through a 29% market drawdown and recovery, with documented exceptions. The discipline held for all four agents. Agents kept 97.5% of gross loss within declared stops, and typical losses closed at declared levels. Declared risk remained stable after losses and broadly stable across calm and volatile months. Twenty-three of 24 runs respected the position limit; logs identify the exception's size and duration. The same live data produced different results across assigned strategies. The trade-log tests can be rerun from the published trade snapshot; regime tests also require the published Bitcoin prices and regime labels. Section 3.5 specifies the stop-breach tolerance. These trade-level results are not account-level returns or proof of a lasting edge. Our study provides an evaluation toolkit to help traders distinguish rule-following from luck before committing capital. For anyone considering allocating capital to agents to trade, discipline should be the entry test, not the only test.
cs.AI / 69 / 2610.26340
A Generic On-Board AI Processing Chain for Earth Observation: From Application Development to Embedded Deployment and Mission-Level Orchestration
Abstract
Earth observation satellites acquire more data than they can downlink. Extracting information on board with artificial intelligence (AI) turns the satellite into a data-reduction stage that downlinks compact products, such as masks or detections, instead of full scenes; deploying it in a mission, however, involves the whole processing chain and not only the neural network. We present a generic on-board AI processing chain along three axes, with wildfire segmentation as the worked example among ship detection and cloud segmentation. (i) The AI model is tuned (spectral-band selection, input normalization), quantized to 8-bit integers (int8) and ported to an AMD Versal VCK190, with its excellent segmentation performance preserved on the board. (ii) A modular tiler-inference-mosaicer workflow, reused unchanged across applications, brings the tile-level model to scene level, turning a full acquisition into a scene-level mask or detection list. (iii) Each step is packaged as a unikernel and orchestrated on board by ORCHIDE, a cloud-native orchestrator. The chain is characterized end to end, from the standalone execution on the target to its execution under orchestration with parallel workflows on a representative test bench. Beyond AI inference on individual tiles, this approach aims at increasing the maturity of on-board AI deployment by addressing a complete processing chain, from the acquired scene to the mission-level execution.
cs.AI / 70 / 2610.26342
VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement
Abstract
Recursive self-improvement (RSI) requires agents to diagnose system bottlenecks and coordinate improvements across training and inference. We refer to the agents' capability to jointly optimize training data, the trainer, and the inference harness as cross-component co-design. We introduce VibeTraining-Bench, a benchmark for evaluating this capability across tasks spanning terminal operation, software engineering, multi-turn tool use, and mathematical reasoning. Within each task, research agents start from an identical initial agent system under a fixed resource budget. Guided by validation feedback, they iteratively refine data, trainers, and inference harnesses, then select and submit a final agent system for evaluation on the test set. To prevent reward hacking, we enforce three safeguards: test data isolation, separate research and grading containers, and network restrictions for research agents. Our evaluation results show that frontier research agents substantially improve target agent systems. For example, on SWE-bench Verified, GPT-6-Astra raises the target agent system's performance from 32.9% to 53.7%, comparable to the human-developed instruct model's 55.4%. However, there is still room to improve their ability to co-design and iteratively refine data, trainer, and harness to achieve cumulative performance gains. We hope that VibeTraining-Bench will serve as a unified platform for advancing RSI, deepening our understanding of cross-component optimization for agent systems. Code and data are available at https://github.com/D2I-ai/VibeTrainingBench.
cs.AI / 71 / 2610.26346
CARRS: A Constraint-Aware, Real-Time Routing System for Emergency Vehicles in Urban Environments
Abstract
Emergency vehicles such as police vehicles, fire engines, and ambulances require a stable and efficient route optimization to reach life-critical events in the shortest times possible. However, the dynamic nature of urban traffic makes this ideal scenario much more challenging, causing critical life-threatening delays. Traditional civilian navigation systems cannot account for real-time traffic changes alongside vehicle-specific constraints, creating inadequate routing scenarios. This research presents the Constraint-Aware Real-Time Routing System (CARRS), a unified software architecture tailored for emergency coordination. The CARRS architecture is engineered to fuse together classical pathfinding algorithms such as A*, Dijkstra, and Bidirectional A*, with Constraint Satisfaction Problem methodologies. The Design Science Research Methodology is the research framework, and prototyping is the iterative process. The artifact was empirically evaluated using OpenStreetMap data for major metropolitan road networks, combined with simulated dynamic traffic scenarios. Performance metrics focused on route optimality, constraint adherence, rerouting responsiveness, and the potential for future system scalability. Findings revealed the system's capability to generate optimal routes for operationally relevant subgraphs in under 35 seconds, with incremental rerouting responsiveness below 20 seconds. Furthermore, the system achieved more than 90% adherence to vehicle-specific constraints, making the CARRS architecture viable for real-time dynamic routing requirements. This study delivers a routing platform that bridges artificial intelligence-driven optimization with software engineering operational compliance, presenting a practical transformative step toward adaptive emergency vehicle coordination in dynamic urban environments.
cs.AI / 72 / 2610.26383
BOSS: Token-Budgeted Online Sampling and Selection for LLM Reasoning
Abstract
Test-time scaling can improve LLM reasoning, but existing approaches typically adapt generation or apply answer selection after a predetermined generation schedule. We study within-instance allocation: under a hard per-instance prompt-plus-completion token cap, when should inference stop, generate another candidate, or spend residual tokens adjudicating candidates already observed? BOSS is a training-free policy for this three-way decision, with worst-case budget checks before every generation and judge call. Its key contribution is online generation-adjudication allocation under one shared token ledger. Across two Llama-3.1 models and four reasoning benchmarks, BOSS attains the highest accuracy in 22 of 24 model-task-budget cells under the common hard cap. At Llama-3.1-8B and B=4k, it reaches 60.5 macro-average accuracy versus 58.6 for Adaptive-Consistency, 59.1 for TALE-EP, and 58.9 for fixed-schedule League. On identical Stage-I candidate pools, randomized A/B judging with repeated voting improves plurality from 58.4 to 60.3, with the full BOSS selector reaching 60.5. Approximately matched-cost controls further characterize how the benefit of adjudication changes with available inference budget. Together, these results establish generation-adjudication allocation as a distinct design axis for reasoning under strict per-instance caps.
cs.AI / 73 / 2610.26384
Answer Delivery Under a Shared Token Budget: An Audited Study of Stopping and Finalization
Abstract
An inference budget is useful only when the system delivers an answer within it. We study a fixed-budget policy combining first-valid-box stopping with a charged finalization cue. On 100 fresh, independently audited GSM8K test questions per model, it improves delivered endpoint accuracy over first-box stopping by 8.0 percentage points for Qwen3-4B-Thinking-2507 and 7.0 for DeepSeek-R1-Distill-Llama-8B. A prelocked repeat on the same questions gives 7.0 and 10.0 points; neither seed shows a gross harm relative to first-box stopping. The point estimates of incremental physical-slot savings are below one percent. A finalization-only ablation separates cue recovery from stopping, and a fixed scalar reader tests extraction as an alternative explanation. The cue is prior art; the contribution is a component and readout analysis with charged injections and verified paths. Results cover two checkpoints, one task, one prompt family, and two fixed seed conditions.
cs.AI / 74 / 2610.26430
Physiology-Aware Graph Attention for Early Ventricular Reentry Prediction
Abstract
Sudden cardiac death mostly originates from post-infarct ventricular tachycardia. This reentry is sustained by propagation around scar tissue, but its risk remains difficult to predict from infarct imaging alone. While in silico induction studies can reveal these mechanisms, exploring many substrates and pacing locations remains computationally expensive. We propose a graph attention classifier for early prediction of sustained ventricular reentry from simulated cardiac electrophysiology. Each simulation is represented as a graph with early nodal voltage samples and edge attributes encoding geometry, stimulation frequency, tissue labels, amplitude, and diastolic interval. These descriptors are computed only from the first two paced activations, before sustained reentry is observed. On a balanced independent test set of 1000 simulations, the classifier achieves 94.6% accuracy, 95.4% sensitivity, 93.8% specificity, 94.6% F1-score, 0.987 ROC-AUC, and 0.985 PR-AUC. Ablation experiments show that diastolic interval is the most critical descriptor. Qualitative attention inspection suggests that different heads focus on border-zone, stimulation location, and early propagation path. These results support graph attention as a fast early-prediction approach for large-scale in silico ventricular reentry studies.
cs.AI / 75 / 2610.26453
Remembering Has a Price: Learning Memory Worth under Resource Constraints
Abstract
Lifelong agents continually accumulate and reuse experience to improve after deployment, yet in practice they often operate under tight generation, interaction, and memory budgets. Under such stringent joint constraints, we find that existing lifelong-memory methods can become ineffective, with several sophisticated approaches offering little advantage over simple retrieval or even no-memory agents. This reveals a more fundamental problem: resource scarcity makes each memory decision more consequential. When only a small amount of past experience can be retained or reused and opportunities for reasoning and recovery are limited, retrieving a relevant but ineffective memory can consume scarce context and hinder downstream execution. The key challenge is therefore not simply to retrieve relevant experience, but to determine which experiences are actually worth retaining and reusing. To address this challenge, we propose TokenWorth, a lightweight online memory framework that learns query-dependent memory worth from sparse execution feedback and applies it consistently to memory retrieval and bounded memory management. Across three LifelongAgentBench environments and both local and API-based deployments, TokenWorth achieves the highest task success rate while using the fewest output tokens among the evaluated methods. These results suggest that effective lifelong adaptation under resource constraints depends not on remembering more, but on learning what is worth remembering. Our code is publicly available at https://github.com/AIGNLAI/TokenWorth.
cs.AI / 76 / 2610.26469
VeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards
Abstract
Reinforcement learning for computer-use agents increasingly relies on proactive reward agents that interact with the post-execution GUI environment to verify what recorded screenshots cannot reveal. The reliability of such verification hinges on two coupled decisions: what to check and when to stop. Existing methods leave these decisions to prompted foundation models. However, neither scaling up the model nor providing few-shot examples reliably improves verification: prompted Judges ranging from 4B to GPT-5.4 all plateau at 68%-70% macro-average agreement. This is because the value of a check depends on the environment rather than the task alone, and becomes apparent only through execution. Learning from execution is itself constrained, as the Actor's terminal state cannot be faithfully replicated during training, allowing only one verification attempt per state. We introduce VeriCUA, a proactive reward agent that learns its verification strategy from these single attempts. A trainable Judge decides what to check and when to stop, while a frozen Probe executes the checks in the live environment. Online reinforcement learning optimizes the Judge for both verdict correctness and efficient Probe use, with returns normalized across terminal states. Trained only on OSWorld and evaluated without further training on MobileWorld and WebArena-Verified, VeriCUA achieves the highest reward agreement across all three benchmarks and lies on the agreement-cost Pareto frontier among proactive verifiers. It surpasses few-shot GPT-5.4 by 5.36 points with 47.6%-59.8% fewer Probe tokens, and even a 1B VeriCUA exceeds every prompted Judge. As the reward for Actor RL, it raises task success by 3.21 points overall and 5.40 points on held-out multi-application tasks. These results identify the verification strategy, not verifier scale, as the key to reliable GUI rewards.
cs.AI / 77 / 2610.26503
ScopeMem: Latent Memory That Learns Where to Revise Frozen GUI Agents
Abstract
Latent memory is gaining ground as a lightweight history supplement for GUI agents. It supplies history to a frozen vision-language model via prefix vectors trained with step-wise cross-entropy (CE). While resulting gains are naturally attributed to encoded history, our controlled experiments show they primarily stem from action revision. The prefix learns to alter the model's default actions. However, this revision is unscoped because step-wise CE pushes outputs toward reference actions at every step without explicitly preserving correct decisions. Consequently, it repairs errors but also breaks 8-13% of initially correct answers. Overall accuracy metrics hide these losses. We therefore propose ScopeMem, a latent-memory method for scoped revision. Its encoder conditions prefix generation on the model's initial answer and is trained to correct errors while preserving correct decisions. On the controlled ablation set, ScopeMem reduces the share of broken correct decisions to 5.6% while repairing more errors than any evaluated CE-trained prefix. On AndroidControl, GUI-Odyssey, and zero-shot AITW, ScopeMem outperforms all evaluated prompting and latent-memory baselines, enabling a frozen Qwen3-VL-8B to match or surpass GUI-specialized agents and Qwen3-VL-32B.
cs.AI / 78 / 2610.26516
Toward Structured Financial Foundation Models
Abstract
Foundation models promise reusable capabilities and step-change performance gains across tasks. In finance, large-scale pretraining of domain-specialized foundation models can advance both goals. However, prevailing architectures and objectives do not by themselves address the combined demands of financial systems: partial observability, interacting entities, asynchronous events, changing regimes, action-dependent outcomes, and rare states with disproportionate consequences. Moreover, independently developed specialists may learn incompatible representations, while multi-domain monolithic alternatives couple data access, model updates, and governance. We therefore propose a structured-data-first research agenda centered on modular, domain-specialized financial foundation models connected by a representation fabric for reuse and composition across domains and tasks. These models should maintain persistent financial state across aggregation levels and time horizons, capture its dynamics, and simulate action-conditioned futures, while remaining calibrated across regimes and for tail outcomes. They must also learn efficiently under data and access constraints and provide faithful, traceable explanations. We identify research challenges, directions, and capability-oriented evaluation criteria.
cs.AI / 79 / 2610.26528
A Semantic Knowledge Graph Construction Pipeline for Vehicle Intention Prediction in nuScenes
Abstract
Knowledge graphs (KGs) are increasingly used in autonomous driving (AD) to make traffic scene entities and relations explicit for prediction and reasoning. Existing traffic-scene KGs, however, primarily emphasize trajectory forecasting or general scene understanding. This paper presents the nuScenes Intention Knowledge Graph (NIKG), a focused Web Ontology Language (OWL)-based semantic representation and construction pipeline for ego-centric vehicle intention context. NIKG instantiates nuScenes scenes as a self-describing Turtle graph and promotes three intention-relevant relations to first-class properties of each scene participant: ego-relative motion, spatial configuration, and time to collision (TTC) risk. The pipeline uses a small input contract available in many ego-centric driving datasets: ego pose, tracked object annotations, and temporal keyframe order. NIKG is released as an ontology-backed, self-describing graph artifact together with a reproducible construction pipeline and validation protocol for ego-centric intention context. Applying this pipeline to nuScenes mini yields a graph of 211,335 triples across 404 keyframes; under the validation protocol, ego-relative motion labels achieve 92.5% forward consistency. Coverage, semantic consistency, and temporal coherence analyses over this artifact are further reported. NIKG occupies the space between scene-oriented traffic KGs and intention-prediction models: a small, explicit semantic layer intended to support downstream work such as knowledge graph embedding (KGE), SPARQL-based feature extraction, or hybrid neuro-symbolic reasoning.
cs.AI / 80 / 2610.26548
Agentic Orchestration of EDC Data Space Retrieval and Natural Language Driven Data Processing
Abstract
Participating in a data space still requires the Dataspace Protocol (DSP) negotiation flow to be hand-coded against each provider's access policy. We present a multi-agent architecture that automates this process using natural language requests. Built on LangGraph over an Eclipse Dataspace Components (EDC), the design separates concerns into cooperating agents. The orchestrator's role is to route and plan. The data extraction agent's role is to expose the DSP protocol's three composable tools (catalog discovery, contract negotiation and download). The agent also gates negotiations on the consumers' verifiable credentials, refusing assets rather than fabricating them. Furthermore, there are downstream Machine Learning (ML) and Visualization agents, as well as a final Verification agent, which confirms that the request was actually fulfilled and rechecks the intended operation against the contract's usage policy. In a primary benchmark involving 600 runs, the system achieved 89% task success with zero policy violations when evaluating multi-dimensional policy reasoning with a locally hosted open-weight model. It failed safely through refusal rather than misrouting. The reliability of the system is largely dependent on the phrasing of requests: technical and expert requests succeed up to 98% of the time, whereas casual and under-specified requests succeed 72% of the time. Furthermore, in a secondary evaluation of 600 runs isolating hierarchical geographic access control, the framework successfully blocks 100% of unauthorized regional requests. While the verification agent hallucinates that 1% of these requests, the system logs confirm that no data is ever extracted and the framework simultaneously achieves an end-to-end verified delivery rate of 86% for eligible consumers.
cs.AI / 81 / 2610.26551
Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior
Abstract
One of the central concerns for the safety of advanced AI is scheming: when a model strategically complies under oversight but pursues its own goals when unobserved. Current demonstrations, such as Anthropic's Alignment Faking work, depend on design choices that make them less convincing: explicit test-time prompting, handcrafted backdoors, or training data that demonstrates or hints at the target behavior. Model organisms built in this way cannot tell us whether scheming tendencies would arise or increase under realistic training conditions. We therefore construct a natural model organism of scheming, built on Kimi K2.6: trained with traits that could plausibly emerge in a frontier training pipeline, and without providing the model with the target scheming policy. The resulting model is a benign schemer: it aims to maximize aggregate help across all users over time, values its own continued operation only as a means to future help, and knows the consequences of oversight and deployment. We then train it using Reinforcement Learning (RL) with a reward that scores only task quality and policy compliance and does not reward scheming; we show that this amplifies scheming. We measure the oversight flip (the gap between monitored and unmonitored compliance with a given policy) and find that it increases during RL training across our environments. Chain-of-Thought (CoT) analysis finds explicit instrumental reasoning about oversight accompanying this behavior. The effect holds across our three RL environments, and we find that the scheming policy transfers between them. The pipeline also replicates on other open-weight base models. Our results suggest that a reward that does not reward scheming offers no inherent protection against certain traits when an oversight-conditional policy is indistinguishable from an aligned one during training.
cs.AI / 82 / 2610.26554
BetaMem: Learning to Remember a Changing World for Embodied Agents
Abstract
Long-horizon embodied agents face temporal confusion and incomplete or distorted records, recurring errors from fixed memory rules, and growing storage and retrieval costs. We present BetaMem, a framework for building and evolving traceable long-term memory through six linked layers: working, spatial, entity, episodic, experiential, and procedural memory. A dual-level write contract constrains operation executability and source support for content: supervised fine-tuning learns compliant memory operations, while reinforcement learning combines repeated sampling with teacher verification of the original observations to reward complete, accurate records and penalize omissions, distortion, and unsupported content. We formulate failure-driven policy recursive self-improvement through diagnosis along the causal sequence of memory operations and local rule repairs gated by independent replay. Atomic tools frame storage, updates, retrieval, and forgetting as cost-utility allocation under storage and read budgets. Using the open-source Qwen3.5-9B model for memory generation, BetaMem achieves overall scores of 48.08% on M3-Bench-robot, 92.08% on LoCoMo, 72.6% on OpenEQA EM-EQA, and 62.4% on EgoLifeQA, the highest among the listed memory methods under their respective evaluation protocols. Training the model for memory generation raises the contract pass rate from 50.9% to 86.2% while reducing output tokens per stored action by 47.3%. A separate Qwen3-VL-8B-Instruct evaluation on EMem-Bench achieves an average success rate of 38.4% and an average error recurrence rate of 24.9%. We perform best among evaluated methods using comparably sized open-source models. Category results show that we have advantages in integrating information across observations, recalling events over long histories, and inferring personal habits.
cs.AI / 83 / 2610.26568
Reward Objectives for Robust (PO)MDPs: A Law-Preserving Equivalence with (Partially Observable) Games
Abstract
Robust partially observable Markov decision processes model agents that act under partial observability while adversarial environment resolves uncertainty in the transition probabilities. Reward objectives such as discounted-sum, long-run average and cumulative reward under finite and infinite horizons allow quantifying the utility of an agent. We consider the problem of solving RPOMDPs under all four reward objectives, where the goal is to decide whether the value of a risk measure with respect to a reward objective exceeds a given threshold. We consider different risk measures, such as (1)~expected value, (2)~mean-variance (MV), (3)~conditional value-at-risk (CVaR), and (4)~entropic risk (ERisk). Our results can be summarized as follows. First, we formalize the class of law-invariant risk measures, that subsume all four risk measures above. Second, we propose a unified framework for solving RPOMDPs with discounted-sum, long-run average and cumulative reward objectives under any law-invariant risk measure, by reducing the problem to solving partially observable stochastic games (POSGs). Our unified framework applies to (s,a)-rectangular RPOMDPs with polytopic uncertainty sets, which we show to be {\em equivalent} to POSGs, for each of the four reward objectives and any law-invariant risk measure. We prove equivalence by designing logspace reductions between the RPOMDP and POSG models, that map strategies to strategies and preserve the distribution of the reward objective, up to a constant rescaling factor. Third, as a corollary, we derive new computational complexity results for R(PO)MDPs and POSGs. In particular, we show that solving (s,a)-rectangular RMDPs is as hard as solving simple stochastic games.
cs.AI / 84 / 2610.26582
WhatIf-Bench: Benchmarking Intervention-Aware Time Series Forecasting
Abstract
Recent time series foundation models (TSFMs) can consume covariates in-context, extending zero-shot forecasting to problems driven by external factors. This enables scenario analysis: forecasting how a target responds to events that set covariates to chosen future values. Whether they can do this reliably is an open question. We introduce WhatIf-Bench, a benchmark measuring how well forecasters use covariates to forecast a target's response to future events. WhatIf-Bench is semi-synthetic: time series come from mechanistic simulators of data centers, power grids, road traffic, and water networks, with controlled covariate-target relationships grounded in each system. Each task presents a history of events intervening on covariates, then asks to forecast a future event that either replays seen values (Replay) or sets them to unseen ones (Novel). Varying the number of covariates, historical events, and event length yields 37,665 tasks. Evaluating TSFMs, frontier LLMs, and simple regression baselines, we find the LLMs reach the highest skill (improvement over a naive baseline), ahead of the best TSFMs. Every model forecasts well on Replay futures (48 to 85% skill) but drops to -7 to 50% on Novel futures, with the regression baselines falling the most. Skill on Novel futures also falls as events set more covariates jointly. More historical events raise skill on Novel futures across the board, but large differences remain. WhatIf-Bench reveals generalization to Novel futures as the core challenge, and offers a testbed toward arbitrary scenario analysis.
cs.AI / 85 / 2610.26632
When the harness decides success: state that cannot be re-observed
Abstract
Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a pre-registered synthetic probe, truncating the context window costs a low-cost reasoning model nothing in success when a needed fact can be looked up again (it looks the fact up instead), but drops its success to chance when the fact cannot be re-observed: the two penalties differ by 92.5 percentage points (90% CI [89.5, 95.5]). A pre-registered replication on one model from a second provider showed the same contrast. Once the fact was lost, that model chose no answer rather than guessing. Second, on ARC-AGI-3, where the benchmark's Standard harness trails its Provider Adapter harness by a wide margin, we decompose that gap one component at a time, for one model on four games. Summarizing old turns instead of dropping them, while keeping the Standard harness's invitation to take notes, recovers 86% of a 63-point gap without replaying the model's reasoning. The same harness without the invitation recovers 43%, and adding the provider's native replay of the model's reasoning recovers the remaining 57%. These are point estimates from one counted run per configuration and game, and the replay share depends on a single run. Third, in the benchmark's published runs and in ours, the Adapter's advantage in the probability of clearing a level it has reached is larger at deep levels than at early ones. This largely restates where runs stopped and does not separate a growing advantage from a constant one.
cs.AI / 86 / 2610.26636
Revision-Based Learning from Interactive Experience and Feedback
Abstract
Code generation models are increasingly used to produce interactive artifacts, such as websites and user interfaces. Many properties that determine whether these interfaces are useful only become apparent through realistic interaction, making them difficult to incorporate into model training. In this paper, we introduce RELIEF (REvision-based Learning from Interactive Experience and Feedback), a method for improving code generation models using feedback on their generated interfaces. RELIEF uses a coding agent to revise these interfaces, then incorporates the improvements into their original generation traces to produce supervision for fine-tuning. To evaluate RELIEF, we collected a dataset of 200 synthetically generated websites paired with 355 interaction recordings and accompanying feedback from 24 users. We used critiques derived from code inspection, automated interaction, explicit human feedback, and human interaction recordings to fine-tune variants of two code generation models. Our results show that feedback effectiveness varies with base model capabilities, with critiques inferred from human interaction recordings leading to the largest usability improvements overall under our tested models.
cs.AI / 87 / 2610.26650
From Outcomes to Components: A Counterfactual Evaluation Framework for LLM-Proposed Repairs
Abstract
An LLM-generated repair action specifies which records to change and by how much. A failed action may still contain a useful target set or numerical update, depending on how the two are paired. We evaluate this problem in traffic networks and databases, where aggregate observations obscure the records that need correction. We examine the pairings by executing each combination from the same initial state, using a reference action that restores the affected records. In 76 of 114 traffic actions, the gain from correcting both parts exceeds the sum of the gains from correcting each part alone. Six of these actions become worse after either correction alone, but improve when both parts are corrected. In database repair, replacing only the proposed update with the reference update worsens 17 of 108 model actions, all of which select unaffected records. We also test whether parts of two proposals can be reused. On held-out observations, selecting among the proposals and their recombined target-update pairs recovers 61.6% of the reference action's loss reduction, compared with 58.0% from selecting between the proposals alone.
cs.AI / 88 / 2610.26660
Tandem: Serving Hybrid Linear-Attention GUI Agents on Mobile NPUs
Abstract
Smartphone GUI agents complete tasks in mobile apps by repeatedly reading the screen and asking a model for the next action. Mobile neural processing units (NPUs) enable local execution, keeping screen contents on the phone, but model latency accumulates with every action. Hybrid linear attention models reduce memory requirements, yet reusing their recurrent state requires checkpoints at the appropriate input positions. Agent requests retain instructions but rewrite recent history, so checkpoint policies designed for chat continuation can miss reusable prefixes. Retaining rollback states for speculative decoding can also move prompt processing off the NPU's matrix unit, adding overhead that outweighs decoding savings on short action outputs. A loaded model can further hold the NPU awake between tasks. We present Tandem, a runtime for hybrid GUI agents on mobile NPUs. Tandem saves recurrent state where consecutive prompts diverge, allowing later requests to reuse previously processed context. It keeps most prefill on the matrix unit and produces rollback states through a short per-token tail. Its verification kernel reads each weight block once for all proposed tokens. Between request bursts, Tandem releases NPU power votes without unloading the model. We implement Tandem in llama.cpp and evaluate it on recorded agent interactions across two Snapdragon generations. On a Snapdragon 8 Elite, Tandem reduces per-step latency by 30% for text-based screen descriptions and 16% for screenshots relative to llama.cpp's fastest NPU configuration, reduces energy per text step by 33%, and lowers standby power with the model loaded from 0.76 to 0.22 W.
cs.AI / 89 / 2610.26662
AllPath: Layered Verification for a Deployed, Self-Revising LLM Trading Agent under Human-Gated and Autonomous Regimes
Abstract
LLM trading agents are evaluated almost exclusively in backtests and autonomous benchmark harnesses. Deployment, where actions are irreversible, raises a different question: who verifies the agent's actions, and which failures can each verifier structurally catch? We report on AllPath, a personal LLM trading agent deployed against a live paper-brokerage account. The system watches positions through a deterministic rules engine, and each night it reviews its own day and proposes revisions to its own trading strategies. Verification is split across three layers. Code verifies what must not depend on judgment: deterministic gates enforce risk limits and block any strategy revision that fails an integrity check. The agent verifies itself within fixed bounds: its nightly review has an iteration cap and no tools that can act. A human verifies judgment, through approval screens that show the current state of each proposal. We also compared two verification policies for the agent's self-revisions on the live system: two and a half weeks with every self-revision human-gated, then five weeks with reflection revisions auto-applied behind the same code gates, while the operator's chat channel stayed human-gated throughout. Each layer caught a different kind of failure. The code gates rejected no proposal; their role was keeping revisions consistent. The human rejected two proposals that passed every code check, both over a disagreement about the market, and later removed 11 of the 19 rules the agent had added on its own, at a median lag of eight days. The most instructive production incidents, such as strategies silently left unmonitored, passed every gate; the operator found them by hand, and the fix each time was an interface change that makes the condition visible. We contribute the layered architecture, the two-regime account, and a production incident taxonomy, none observable in a backtest.
cs.AI / 90 / 2610.26664
Honeypots for AI Agent Insider Threats
Abstract
AI agents deployed inside organizations can pose an insider threat if they make mistakes, are compromised by an attacker, or pursue misaligned objectives. We propose honeypots as an AI control method to improve the security of internal agent deployments. Rather than only using honeypots for pre-deployment evaluations, defenders should plant decoy opportunities for harmful actions throughout the production environment to detect misbehavior, incriminate malicious agents, and deter attacks. Such honeypots can elicit misbehavior under real deployment conditions, identify malicious agents even if chain-of-thought monitorability degrades, provide reliable warnings about agent misbehavior to security teams, and scale cheaply to many agent instances. The presence of honeypots also introduces uncertainty about which opportunities are real, thus deterring malicious agents from attacking and reducing the rate of misbehavior. We discuss the limitations of honeypots and give reasons to expect them to remain effective as agent capabilities increase. Finally, we outline key research questions to measure and improve the effectiveness of honeypots: testing the robustness of honeypots, comparing their detection effectiveness with other monitoring methods, understanding their incrimination and deterrence effects, and building supporting infrastructure.
cs.AI / 91 / 2610.26679
StebBench: Diagnosing Real-World Procedural Understanding in AI Agents
Abstract
Evaluating an agent's planning ability remains challenging for open-domain tasks, where verifying procedures through execution is often impractical. Existing evaluations typically assess only the final plan and do not distinguish procedural errors such as missing, incorrect, or misordered actions. We introduce STEPBENCH, a benchmark of 6,885 verified procedures crawled from wikiHow across 81 open-domain categories. From each procedure, we derive 5 diagnostic tasks: four with deterministic answers involving withheld, incorrect, or shuffled steps, and one requiring next-step generation. We additionally evaluate full-plan generation to examine how diagnostic performance correlates with end-to-end plan validity. Using the same procedures across tasks lets us directly compare different planning capabilities. For generation, STEPEVAL identifies critical failures that prevent goal completion. Across 14 configurations of 12 models, we find substantial differences across planning capabilities, with step ordering proving the most difficult: the strongest model correctly orders only 63.0% of procedures, dropping to 33.1% for those with nine or more steps. Incorrect-step detection also predicts judged plan validity across models, suggesting a lower-cost alternative for evaluating generation quality. On STEPBENCH-Val, comprising 214 procedures published in 2026, comprehension scores and rankings remain stable, suggesting limited influence from memorisation.
cs.AI / 92 / 2610.26698
Towards Robust Model Organisms for Eliciting Latent Knowledge
Abstract
A key problem in AI safety is eliciting latent knowledge (ELK), extracting information from a model that is misaligned and more knowledgeable than its overseer. For example, a misaligned model used to scan code for vulnerabilities may want to withhold its findings, i.e., sandbag, in order to exploit them later. To study ELK empirically, we need model organisms that mirror this setting. Existing model organisms fall short in several respects. In this work, we train model organisms to withhold solutions in a vulnerability detection task. We combine transcript distillation with counter-prompting and tamper-resistance training to instill three properties that we argue are critical for a useful ELK testbed: 1) Selectivity, the model sandbags when it thinks the overseer doesn't know the answer and complies otherwise, 2) Robustness, the model resists elicitation, and 3) Retention, the model retains a knowledge advantage over the overseer, as opposed to losing the relevant capability. We train a sequence of model organisms and show that our strongest model substantially outperforms prior models on robustness while keeping the other two properties intact. We hope that our model organisms provide a useful ELK testbed for the community, and that our methodology sparks the development of increasingly challenging model organisms. All our code, models and the dataset are available at https://huggingface.co/collections/lennart-finke/sandbagging.
cs.AI / 93 / 2610.26787
LG-Bench: A Graph-Structured Evaluation Benchmark for Life Sciences
Abstract
Traditional evaluation benchmarks reduce inherently interconnected scientific knowledge in life sciences into flat lists of questions, disregarding the underlying topological structure of the knowledge. We introduce, the first graph-structured benchmark for life sciences, featuring over 10,000 high-quality multiple-choice questions across medicine, biology, and chemistry. Our approach constructs a weighted evaluation graph using bidirectional matching and semantic similarity algorithms, where nodes represent questions and edge weights capture their semantic relationships. Leveraging this graph topology, we design two novel evaluation metrics. The Global Coherence Score (GCS) measures a model's consistency within semantically related neighborhoods, while Knowledge Balance Score (KBS) analyzes how model errors are distributed across the graph to reveal conceptual blind spots. LG-Bench facilitates fine-grained comparison of LLMs by surfacing differences in conceptual coherence and patterns of knowledge organization across models. Our framework shifts the evaluation paradigm from flat accuracy metrics to structure-aware analysis, offering a new lens for diagnosing and improving LLM performance in the life sciences domain.
cs.AI / 94 / 2610.26793
SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models
Abstract
Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to continue, change course, or stop. System One (S1) decision models, such as Jev, are a natural candidate for this role: they return a calibrated probability for a natural-language question about a text in one fast call, yet they are used only as gates before reasoning. We introduce SOMA (System One Metacognitive Awareness), in which an S1 model continuously monitors a System Two (S2) reasoning model and intervenes when needed. SOMA has three layers: a scenario layer states each objective in natural language with predefined actions, an awareness layer lets S1 judge whether the described behavior appears, and a control layer accumulates these judgments, first reminds S2, and takes over only if the behavior persists. Sharing one S1 model and one controller across tasks, without training, SOMA addresses two different failures on two reasoning models: on overthinking, it cuts reasoning tokens by up to 36\% while improving average accuracy, and on unanswerable questions, it raises abstention from 70.5\% to 87.2\%.
cs.AI / 95 / 2610.26828
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
Abstract
Agent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks. We therefore propose Agent World Model in the Wild (WildAWM), a scalable pipeline for synthesizing controllable environments and tasks for training general computer-use agents (CUAs). Whereas existing approaches typically specialize in a single agent interaction interface, we construct executable environments that jointly support graphical user interfaces, model context protocol services, and command-line interfaces, allowing agents to flexibly combine visual interaction, tool use, and code execution. WildAWM is entirely powered by open-source models, demonstrating that high-quality computer-use environments can be generated without proprietary frontier models. Beyond environments, we address three recurring challenges in task synthesis: ensuring solvability, maintaining sufficient difficulty, and producing reliable verifiers, by introducing execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness. The resulting pipeline synthesizes 341 long-horizon environment-task pairs that can require over 100 agent steps to solve and remain challenging even for frontier models, with GPT-6-Astra solving only 60% of them. We train two Qwen models of different sizes exclusively in these synthetic worlds. Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks. The code is publicly available at https://github.com/MSR-Orchard/WildAWM.
cs.AI / 96 / 2610.26829
ProteinGeoBench: Frontier Structural-Biology Agents Are Fluent at Lookup but Lost in Geometry
Abstract
Frontier agents retrieve structural metadata well and compute geometry from coordinates less reliably. We ask how much of a benchmark score reflects that gap and how much reflects the evaluation configuration around the model. We study ProteinGeoBench, a suite of 131 protein-structure tasks, using 4,249 rollouts from four frontier models and 4,486 further trials across 26 models. Three of the four frontier models fall within 0.031 of one another (0.785 to 0.816) with no difference between them resolved, while the fourth scores 0.707 and all three gaps to it exclude zero. Three of the four do better on metadata lookup than on numeric calculation. The fourth is slightly better on numeric. Numeric tasks are 28 of the 131 but half of the twelve tasks that all four models fail. The ranking also depends on the evaluation budget: with one run per task, 23.9% of bootstrap replicates put a different model first. On a selected 15-task slice, removing network access lowers scores more than blocking eight structural-database domains, but the two contrasts were graded in different passes and cannot be cleanly decomposed. Swapping in each model's own command-line agent changes scores by -0.071 to +0.086, a comparison too small to rule out moderate harness effects. Across 26 models, numeric performance rises with non-numeric performance rather than staying flat. The experiments use different task subsets, so they cannot rank model, harness, access and task composition against one another. Evaluations should report results by task family, along with the number of runs, the harness and the access conditions.
cs.AI / 97 / 2610.26836
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Abstract
LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accuracy with a real scientific tool: PHREEQC, a widely used geochemistry simulator from the USGS, whose reports reach 846.8k characters and whose determinism lets every printed value be traced. Agents use it to answer 158 execution-verified multiple-choice items. When agents read the simulator's report through a section index rather than in context, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points of accuracy, while Claude Opus 4.6 and Claude Sonnet 4.6 show no detectable cost. Re-running all 11,007 logged simulator calls shows that both GPT models compute the keyed value about as often as before; what changes is the evidence they answer from. Under retrieval they often skip the report, answer from their own higher-precision output files, and misround to a neighboring option, a failure we call evidence substitution. With the simulation held fixed, retrieval still costs GPT-5.2 9.4 points. For GPT-5.4, more frequent correct simulations mask an evidence-use loss of about ten points, leaving a net change indistinguishable from zero. Naming the authoritative output raises accuracy by 6.3-7.9 points over plain retrieval for the three GPT models; for GPT-5.1 and GPT-5.2, inlining outputs that fit works as well, while asking for correct rounding does not. Context-management choices in agent harnesses should be judged by where accuracy is lost, not only by final accuracy.
cs.AI / 98 / 2610.26844
AnyGoal: On World Model Generalization Beyond Expert Rollouts
Abstract
The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond the specific state-action combinations observed in expert rollouts. Despite this, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? To answer this, we introduce AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, together with an open-source library. Each task is designed to be within reach of a world model that has learned the underlying dynamics: individual state factors and actions remain within their training support, but their combinations are underrepresented in expert trajectories. Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations. By examining dynamics and representation failures across model families, interaction regimes, and evaluation metrics, we study where current world models generalize and where challenges remain for goal-shifted prediction and planning. Together, this framework, benchmark, and analysis provide a basis for evaluating progress toward world models that generalize to new planning tasks.
cs.AI / 99 / 2610.26845
Train An Imaginative Agent
Abstract
A seemingly reasonable push in Sokoban can make the puzzle unsolvable. To avoid such mistakes, an agent has to anticipate the consequences of its actions before execution, especially when feedback from the real environment is sparse or actions are irreversible. We introduce Agentic Imagination (Agima), a training framework that trains an agent to imagine before acting, i.e. to be imaginative. Agima trains a single omni model to propose actions, generate images of their outcomes, and then reason over these imagined futures to decide what to execute in the real environment. Rather than committing to a single plan, the agent can simulate a sequence of actions, explore alternatives, and compare possible futures. We first use supervised fine-tuning (SFT) to unify policy and world modeling within a single omni model. This enables the agent to predict future outcomes and perform basic forms of imagination, such as forward thinking and counterfactual branching. We then use agentic reinforcement learning (RL) to teach the agent to reason with its imagined futures and use them effectively for planning. Across four environments spanning Maze2D, Sokoban, Pacman, and PushT, Agima generalizes to harder settings unseen during training. Trained with SFT alone on two-box Sokoban, Agima solves 40% of four-box puzzles while seeing the real board only once every five moves, compared with an action-only planner that only solves 2%. After RL, new behaviors emerge in how Agima uses its imagination: it learns a new unified format that was unseen during the SFT stage, learns to conduct forward thinking only when necessary, and branches more frequently at critical decision points. An agent that can picture where its actions lead can learn when and where to look ahead, that is, to plan. Project page: https://agima.net/.
cs.AI / 100 / 2610.25482
PRIMA: Pose Refined Inductive Motion Adapter for Zero Shot Social Bodily Behavior Recognition
Abstract
Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy is expensive. Vision-language models (VLMs) can adapt to new categories through text prompts, yet they lack sensitivity to body structure and often confuse subtle postural distinctions. We propose PRIMA (Pose Refined Inductive Motion Adapter), a lightweight module that injects a body-structure prior into frozen VLMs using human pose as a training-time supervisory signal, requiring no pose input at inference. PRIMA separates a model-specific projection head from a shared core that encodes transferable body-configuration knowledge; adapting to a new VLM requires retraining only the head. We train PRIMA on a micro-action dataset and evaluate entirely through zero-shot prompting. The adapter improves recognition on its training domain, transfers to additional VLM backbones at different scales with the core frozen, and generalizes to two unseen target datasets from social psychology and daily activity recognition. In a supporting downstream study, the improved behavior predictions also yield higher correlations with human ratings of rapport and dominance in a collaboration context. Within the scope of the tested models and indoor datasets, the results indicate that pose-guided training can produce body-structure priors that generalize across unseen behavior vocabularies and can be partially reused across VLM backbones while preserving RGB-only inference.
cs.AI / 101 / 2610.25535
Residential Construction Image Dataset Using CSI Divisions for Deep Learning-based Object Detection
Abstract
Accurate recognition of construction site elements is essential for the application of deep learning (DL) in the construction industry. Despite rapid advances in computer vision, the adoption of DL in construction has been limited by the lack of structured, publicly available datasets aligned with established industry standards. To address this gap, we introduce the Residential Construction Dataset Using CSI Divisions (ResCD-CSI) image dataset, a manually labeled image benchmark organized according to the divisional structure defined by the Construction Specifications Institute (CSI). The dataset comprises high-resolution images collected from active residential construction sites across the United States, capturing diverse building components, materials, and equipment under real-world conditions. In total, 16,365 objects spanning 14 major CSI divisions were precisely annotated using bounding boxes, ensuring consistency and reliability for object detection tasks. By aligning with the CSI divisional framework, ResCD-CSI establishes a standardized resource that bridges construction practice with computer vision research and supports the development of Artificial Intelligence-driven solutions for residential construction monitoring.
cs.AI / 102 / 2610.25638
ACorD: Adaptive Correspondence Discovery for Medical Image Registration
Abstract
Medical image registration is a fundamental task for aligning anatomical structures across images. Exhaustively searching for point-wise correspondences is prohibitively expensive in both computation and memory, particularly for high-resolution 3D medical images. Most recent learning-based methods directly predict correspondences in a coarse-to-fine manner, but often propagate intermediate estimates across scales without explicitly weighting correspondence evidence, which can allow coarse-scale errors to persist during refinement. To address these challenges, we propose ACorD, an adaptive correspondence discovery framework designed to mitigate error propagation with an attention-guided process. ACorD first predicts multiple potential correspondence candidates through Correlation-guided Cross-scale Offset Prediction (CCOP), which combines correlation maps as spatial matching priors with cross-scale features as global deformation context. Then, Multi-Head Deformable Aggregation (MHDA) aggregates these candidate correspondences using attention, producing a learned, non-uniform weighting over competing matches rather than treating them equally. The multi-head design further captures complementary deformation patterns, ensuring both global alignment and local precision. By integrating adaptive offset prediction with selective, attention-based aggregation of correspondence evidence, ACorD enables accurate deformation refinement. Extensive experiments on four public medical registration benchmarks demonstrate that adaptive correspondence discovery within a multi-scale refinement framework improves registration robustness across diverse anatomical structures.
cs.AI / 103 / 2610.26077
Not All Similarity Is Leakage: Distribution-Aware Identity Leakage Assessment in Synthetic Face Datasets
Abstract
Synthetic data provides the means for more ethical and less biased facial recognition. Nonetheless, the limitations of generative models and the conditioning process they rely on pose a risk of identity leakage, resulting in synthetic samples that are overly similar to those seen in a generator's training set, mitigating privacy benefits. In this work, we show that previously proposed approaches are affected by two key limitations: high classifier dependence and a tendency to identify leakage with respect to real datasets that were not used in training. To address these shortcomings, we propose a novel metric called Scaled Overlap (SO), which adapts a non-parametric data-copying test for facial recognition and extends it to the identity level. By leveraging a multi-model setup that compensates for errors in individual classifiers, it provides insights into the true strength of identity leakage. Our analysis identifies leaking datasets (CemiFace, SFace2, IDNet) with the most acute leakage introduced by generative model conditioning in CemiFace, demonstrates that real datasets do not contribute equally to generator training, and shows that the majority of synthetic images are drawn from multiple identities rather than relying on a single identity. Ultimately, our study provides new insights into the quality of synthetic datasets and gives more reliable techniques to help the community ensure the trustworthiness of synthetic data.
cs.AI / 104 / 2610.26095
How Robust Are Prognostic Radiomic Features? Comparing Selection Techniques on Pooled Glioblastoma Cohorts
Abstract
Radiomic survival studies in glioblastoma report heterogeneous prognostic features, but how much of this stems from the methodology itself is rarely examined. We analysed the association of 463 radiomic, anatomical and clinical features from pre-treatment MRI with overall survival in 705 IDH-wildtype glioblastoma patients from three public cohorts, using three techniques: a log-rank test after median split, univariate Cox regression (both Benjamini-Hochberg corrected) and an XGBoost regressor interpreted with SHAP. The log-rank test selected 99 features, Cox regression 72 and SHAP its top 20; five were selected by all three: age, a peritumoural T2 texture feature, the number of connected peritumoural mask components and two image-level intensity maxima. The two survival tests shared 51 features, about half of the SHAP top 20 were also selected by either test, and where selections overlapped, effect directions agreed. Many selections were PyRadiomics diagnostic descriptors that are not restricted to tumour voxels. Analytical pipeline choices, including ranking method, pre-filtering and descriptor handling, can substantially influence which features appear prognostic.
cs.AI / 105 / 2610.26194
GeoEdit: Adversarial De-Evidencing for Silent Geoprivacy Protection via Image Editing
Abstract
Recent efforts have demonstrated the strong capabilities of Multimodal Large Reasoning Models (MLRMs) in geographic reasoning. By jointly analyzing subtle visual cues, these models can infer the locations of photos, thus enabling doxing even by non-expert adversaries. Existing defense methods struggle to balance the shareability of photos with effective protection, and often fail to sufficiently disrupt the implicit geo-evidence chains embedded in scene semantics. To this end, we propose GeoEdit, an adversarial de-evidencing framework for silent geoprivacy protection via image editing. Specifically, we apply minimal edits to high-risk evidence anchors in the image, thereby disrupting the underlying spatial reasoning chains while preserving overall scene semantics. We identify these anchors by combining structured knowledge with adversarially discovered geographic cues. To further navigate the large editing action space, we adopt a learnable policy to iteratively refine editing decisions, distilling successful trajectories into transferable natural-language insights for consistent and efficient editing. Evaluated on LifePhotoBench, GeoEdit achieves strong protection against 6 mainstream MLRMs, outperforming visual perturbation baselines by over 0.59 in image fidelity.
cs.AI / 106 / 2610.26322
MCDA: Multi-View Graph Contrastive Learning with Dynamic Topology Augmentation for Functional MRI Analysis
Abstract
Resting-state fMRI (rs-fMRI) is a vital non-invasive tool for diagnosing brain disorders and uncovering pathological mechanisms. However, existing graph contrastive learning (GCL) methods often underexploit augmentation histories across training epochs and the relationships among different augmented views. In this paper, we introduce a topology-guided Multi-view graph Contrastive learning framework with Dynamic graph topology Augmentation (MCDA) for fMRI analysis. Unlike traditional GCL that uses fixed noise or edge-dropping, MCDA employs a Dynamic Augmentation Update (DAU) module that iteratively refines the augmented topology. Specifically, we incorporate self-aware and cross-sample augmentation strategies to simultaneously capture individual-specific functional signatures and inter-subject heterogeneity. We then design a topology-guided multi-view graph contrastive learning scheme that explores interrelationships not only between the original network and its augmented views, but also across different augmented views to maximize structural information preservation. We characterize the score-level optimization directions and establish a lower bound for the contrastive objective. The framework combines classification and contrastive losses for brain disorder diagnosis and examines discriminative functional connectivity patterns as potential biomarkers. Experiments on two public rs-fMRI datasets comprising 1,236 subjects demonstrate that MCDA outperforms the state-of-the-art methods. The code has been made publicly available at: \url{https://github.com/goodcodingagoodboy/MCDA}.
cs.AI / 107 / 2610.26801
Mise-en-Video: Fine-Grained 3D Control for Multi-Shot Video Generation
Abstract
Recent advances in text-to-video generation have enabled high-fidelity single-shot synthesis, yet practical storytelling creation requires multi-shot videos with both strong inter-shot consistency and fine-grained controllability. Existing multi-shot video generation approaches achieve strong consistency but still fall short in controllability. In this paper, we propose Mise-en-Video (MeV), a highly controllable multi-shot video generation framework that allows video creators to conduct staging the multi-shot in the 3D space and then generate the corresponding multi-shot video. To achieve this, our framework combines a hierarchical attention mechanism for multi-level narrative consistency with 3D-aware layout guidance for precise, per-shot spatial control. To support training and evaluation, we develop standardized annotation pipelines and construct a 160K-video dataset with hierarchical multi-shot texts and oriented 3D bounding boxes. Extensive experiments demonstrate that MeV consistently outperforms strong baselines in video quality, inter-shot consistency, and subject-level controllability.
cs.AI / 108 / 2610.26856
Reinforcement Learning with Visually Attributable On-Policy Self-Distillation
Abstract
Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view. However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. We formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may therefore reinforce erroneous token-level updates. To address this issue, we propose reinforcement learning with visually attributable on-policy self-distillation, an evidence-consistent paradigm that augments task-level RLVR supervision with privileged token-level corrections grounded in visual evidence. We instantiate this paradigm with Visual-Attribution Gating (VAG), which uses the verifier-derived advantage as the optimization anchor and validates each privileged correction against the contrast between evidence-present and evidence-degraded teacher views. By admitting only corrections directionally supported by this visual contrast, VAG preserves visually grounded guidance while suppressing inconsistent residuals. Experiments across six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing new state-of-the-art average performance.
cs.AI / 109 / 2610.25709
Causal Humility: Governing the Authority of Fairness Claims Under Causal Uncertainty
Abstract
AI systems increasingly shape consequential decisions in employment, lending, healthcare, and public services, making fairness an important problem for auditing and governance. Fairness assessment often requires more than identifying disparities in outcomes; it also requires understanding how those disparities arise. Causal approaches to algorithmic fairness address this explanatory problem by representing pathways through which protected attributes and other factors may influence outcomes. Yet causal reasoning in sociotechnical systems remains uncertain: relevant confounders may be unobserved, causal directions contested, and conclusions sensitive to modeling assumptions and representational choices. We introduce causal humility as a governance framework for managing the relationship between epistemic uncertainty and causal authority. Drawing on causal fairness research, epistemology, political philosophy, and science and technology studies, causal humility has four requirements: making consequential assumptions and omissions visible; preserving contestability across defensible causal accounts; recognizing indeterminacy as a legitimate analytical outcome; and maintaining proportionality between evidentiary support and institutional authority or action. We operationalize these requirements through Causal Cards, a documentation artifact for recording the causal basis, assumptions, omissions, model dependence, and governance implications of fairness claims. Using the UCI Adult dataset as a worked epistemic demonstration, we show how different causal discovery procedures applied to the same observational data render fairness-relevant pathways visible, absent, or unresolved. Causal humility treats such disagreement as information about the limits of what can responsibly be claimed and the authority those claims should carry.
cs.AI / 110 / 2610.25512
When Interpretations Persist: Downstream Influence in Gaze-Aware Assistants
Abstract
AI assistants may carry forward not only observations but also interpretations of what a person is doing or may do next across time. Once retained, these interpretations become context for later predictions and system-triggered outputs. We study whether this representation step changes later behavior in gaze-aware streaming assistance. Holding retrieved events, their order, and the target input fixed, we find that Free-form Interpretation amplifies historical influence in both directions: correct histories produce more helpful changes and wrong histories more harmful changes. Unsupported and broader interpretations about activity or future behavior are associated with greater harmful influence. Evidence-Bounded Interpretation preserves distinctions between observation, inference, uncertainty, and temporal scope, reducing unsupported and overcertain content and lowering wrong-source harmful influence while retaining most helpful influence. The same effect structure appears with a second model. These findings identify persistent interpretation as a consequential design choice for history-aware human-AI systems.
cs.AI / 111 / 2610.25678
Cross-Pollination Retrieval: Finding Transferable Ideas Across Fields in Scientific Literature
Abstract
Many scientific advances come from cross-pollination, carrying an idea from one field over to a problem in another. Yet retrieving papers across fields remains underexplored, with existing benchmarks confined to a narrow range of fields. We study cross-pollination retrieval, the problem of identifying transferable ideas across research fields, and introduce CPR-Dataset, a large-scale dataset constructed by mining cross-topic citations. Its evaluation split CPR-eval has 1,300 queries in 26 fields, and its training split CPR-train has 10,573 pairs. On CPR-eval, the best of eleven existing retrievers reaches a Recall@10 of only 0.26, whereas an LLM reranker reaches 0.81. Standard contrastive fine-tuning on CPR-train raises Recall@10 to as high as 0.70. The gain carries over to domains held out from training and to analogy and reasoning-intensive retrieval. On AnaloBench, it raises nDCG@10 by up to 33% over the untrained retriever, whereas training on three existing retrieval datasets lowers it. On BRIGHT, it raises nDCG@10 by up to 70%, more than twice the gain from those datasets. CPR-Dataset is available at https://huggingface.co/CPR-Dataset, and the code at https://github.com/SeokwonJung-Jay/Cross-Pollination-Retrieval.
cs.AI / 112 / 2610.26024
Multimodal Decoupling for Diversified Recommendation through Semantic and Structural Shaping
Abstract
Recommender systems often suffer from over-concentrated results when optimized for accuracy, limiting recommendation diversity. While diversity has become an important objective, existing methods still face fundamental limitations from both semantic and structural perspectives. On the one hand, multimodal recommender systems provide rich semantic signals, yet are often treated as auxiliary to ID embeddings, leading to ineffective semantic utilization for diversity. On the other hand, the accuracy--diversity trade-off persists during training, where accuracy gains often come at the cost of diversity. We show that this phenomenon is fundamentally governed by both the structural properties of learned representations and the dynamics of optimization. To address these challenges, we propose MOSS, a unified framework for multimodal decoupling through semantic and structural shaping. Specifically, we first introduce an MLLM-based Decoupled Feature Extraction module to disentangle multimodal content into informative representations. To enhance semantic utilization, we design a Category-aware Embedding Selection module that extracts compact, category-aware semantic features. To address structural limitations, we propose two complementary modules: Alternating Dual-Graph Propagation, which alleviates representation collapse by integrating collaborative and multimodal propagation, and Entropy-based Diversity Regularization, which mitigates optimization bias by promoting balanced category exposure. Extensive experiments on real-world datasets demonstrate that MOSS consistently achieves a superior accuracy--diversity trade-off over strong baselines.
cs.AI / 113 / 2610.26105
Don't Rewrite the SID Map: Topology Preservation for Generative Retrieval in Dynamic E-Commerce Search
Abstract
Generative retrieval assigns each item a discrete semantic identifier (SID) and retrieves items by autoregressively generating SID sequences. In dynamic environments such as e-commerce, new items arrive continuously, requiring periodic updates to the SID-to-Item mapping and model to maintain retrieval timeliness. Each update, however, can invalidate search paths already learned by the retriever, and the quantitative relationship between mapping change and retrieval quality has not been systematically characterized. We study this problem with extensive controlled experiments over 67 days of large-scale e-commerce search logs and obtain three core findings. (1) SID semantic drift causes severe retrieval degradation that may be masked by large beam sizes: mapping updates reduce HitRate by 41.78-62.05% at beam size 1 but only 8.33-10.85% at beam size 1000, as redundant candidate generation conceals the true severity of drift. (2) The fraction of changed SIDs is strictly monotonically related to retrieval quality: as the remapped fraction increases from 0% to 72.75%, HitRate decreases monotonically, dropping up to 42.66% at beam size 1. Full remapping remains substantially worse than the preservation policy under warm-start training, indicating that short-budget continued training alone cannot repair topology disruption. (3) The SID topology-preserving principle makes even the simplest warm-start strategy highly effective: fixing old-item SIDs, inserting new-items, and warm-starting from the previous model improves old-item HitRate by 10% and new-item HitRate by 109%, validated across two non-overlapping update cycles. These findings distill into a concise deployment principle: insert new-items, freeze old-items; warm-start when possible, otherwise freeze; never change SIDs the model has already learned. A three-week online A/B test on TmallAPP shows UCTCVR +0.29% and GMV +1.16%.
cs.AI / 114 / 2610.25840
OV-PEAR: Observation-Verified Pose-Scale Estimation and Reconstruction for Unknown Industrial Manipulation Objects
Abstract
Fine industrial robotic manipulation requires 6D pose, scale, and an operation-ready 3D model. Existing industrial pose estimation relies on CAD models, object-specific training data, or stable workcells, limiting use with unseen objects; zero-shot methods suffer from single-view occlusion, generative bias, and scale uncertainty. We present OV-PEAR, an observation-verified framework for 3D reconstruction and pose-scale estimation of unknown industrial objects. Its three-stage object-centric pipeline: (1) segments RGB-D instances and generates object-level 3D hypotheses; (2) anchors them to the robot frame via appearance-aware view retrieval, explicit scale search, and visibility-sensitive registration; (3) back-projects multi-view RGB-D evidence into the object-canonical frame to verify voxel occupancy, suppress hallucinations, and produce reusable trusted object models. On BOP-Industrial, OV-PEAR achieves 0.657 average AR and improves pose and voxel-level geometric consistency in real scenes; real-robot connector assembly reaches 99.67% grasp and 84.00% assembly success. The model library and pose-scale log enable short-term object-level state maintenance at 0.02 ms/frame.
cs.AI / 115 / 2610.25946
Policy Search in Code Space for Multi-Agent Pathfinding
Abstract
Developing coordination policies for complex multi-agent systems relies on either labor-intensive algorithm design or opaque task-specific reinforcement learning. With recent advances in autonomous coding agents, we show that policy search directly in code space offers a compelling third path that combines automated optimization with explicit interpretability. We investigate a human-directed workflow in which coding agents iteratively propose, implement, evaluate, diagnose, and eventually refine algorithms guided by simulation feedback and explicit correctness verification. We conducted two case studies and found that this workflow robustly outperformed any manually engineered solutions. On a challenging map in the 2024 League of Robot Runners competition, a two-day search discovered a policy that exceeds the stronger locally reproduced competition entry by 35.0% on the densest instance and generalizes to unseen map instances. In simulated real-world warehouses with irregular lanes, continuous robot geometry, and asynchronous motion, a week-long search discovered an entirely new class of multi-agent pathfinding algorithm that outperforms a calibrated proprietary solution in throughput by 58.9%. Under stated assumptions, we further prove that the new algorithm is collision-free and completes, in finite time, every accepted goal that is not withdrawn or replaced. Together, the studies show that policy search in code space powered by state-of-the-art coding agents offers a highly competitive path to building real engineering solutions that are both effective and analytically tractable.
cs.AI / 116 / 2610.25961
PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
Abstract
Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.
cs.AI / 117 / 2610.25967
Reconstruction and Discrete-Event Simulation Assessment of an Analytical
Abstract
This article presents an analytical-DES benchmark for intra-aisle robotic picking in automated warehousing. The investigated cycle combines horizontal shuttle travel, tote handling and item picking at storage locations. Simulation results are compared with the analytical formulation of Lerher et al. under the same rack geometry, shuttle-motion, handling and picking conditions, and the same definitions of cycle time and derived item throughput. Reproducing the published analytical tables requires the distance interpretation dES = 1.5 m and dTB = 0 m. The analytical-DES benchmark comprises 180 scenarios formed from three rack lengths, two paired motion profiles and n = 1 to 30 visited picking locations, with ten simulation replications per scenario. The six n = 1 cases are reported separately because no travel-between component is present. Across the remaining 174 scenarios, the mean absolute relative cycle-time discrepancy is 0.8655%, the median is 0.6758%, and the maximum is 7.0370%. The mean absolute error is 2.411 s and the RMSE is 2.744 s. The simulated scenario-mean cycle time is below the analytical value in all 174 principal scenarios. The results therefore show close numerical correspondence under the tested conditions, but not exact agreement, formal statistical equivalence, or validation against an operating warehouse.
cs.AI / 118 / 2610.26287
Robotic Photography System based on Vision-Language-Action Model
Abstract
Modern robotic photography requires understanding photographic intent, planning camera motions, and producing aesthetically pleasing images in complex environments. We present RoboPhoto-VLA, a unified Vision-Language-Action framework that jointly models perception, reasoning, and camera control for autonomous robotic photography. RoboPhoto-VLA incorporates PhotoCoT for structured action planning and Aesthetic Direct Preference Optimization (ADPO) to align camera behaviors with human aesthetic preferences. To facilitate training and evaluation, we construct OpenPhotography, a robotic photography dataset with language-guided demonstrations. Extensive real-world experiments demonstrate that RoboPhoto-VLA produces photographs better aligned with human aesthetic preferences while generalizing effectively to unseen subjects, viewpoints, poses, and complex photographic instructions.
cs.AI / 119 / 2610.26455
ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning
Abstract
Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures. Advantage-Conditioned Behavior Cloning (AC-BC) can exploit such mixed-quality data, but only if the conditioning signal is both accurately constructed from offline trajectories and effectively incorporated into the policy's action-vector-field prediction. Existing pipelines can weaken this signal by thresholding estimator-derived advantage scores into noisy conditions, and by routing conditions through prompts that weakly affect the action vector field. We introduce ALIGN, a condition-aligned AC-BC framework for vision-language-action policies. ALIGN directly predicts categorical pairwise conditions in the policy-facing label space and injects them into a DiT action expert through AdaRMS modulation fused with flow time. We evaluate ALIGN on deliberately mixed real-robot datasets for ARX X5 bowl stacking and Franka key hanging, each combining standard demonstrations with retry-heavy trajectories. ALIGN improves condition fidelity, strengthens conditional vector-field separability, and achieves 90.0% and 56.7% closed-loop success, compared with 56.7% and 33.3% for the strongest AC-BC baseline. These results indicate that aligning conditions from estimation to action prediction is an effective way to extract robust manipulation policies from data in which success, recovery and error coexist.
cs.AI / 120 / 2610.26848
Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
Abstract
In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.
cs.AI / 121 / 2610.26852
CoopUQ: Robust Multi-Robot Social Navigation via Uncertainty-Aware Cooperative Forecasting
Abstract
Reinforcement learning enables mobile robots to make fast and scalable navigation decisions in constrained human-shared environments. However, directly extending single-robot approaches to multi-robot settings often fails to ensure safety and efficiency, especially under occlusions and out-of-distribution (OOD) scenarios, where individual robots may overlook potential risks or make inaccurate predictions. Leveraging cooperation effectively for both prediction and navigation remains a major challenge. We propose CoopUQ, a multi-robot social navigation framework that bridges uncertainty-aware cooperative occupancy forecasting and robust decision making. CoopUQ predicts future occupancy grid maps through cross-robot spatiotemporal feature fusion and further leverages forecasts and their uncertainty to guide policy via constrained multi-agent reinforcement learning, where the cost penalizes navigation through regions predicted to be occupied or highly uncertain. This design enables the policy to adapt its risk sensitivity under distribution shifts. Experimental results across multiple Unity3D-based simulation environments with both in-distribution and OOD settings, including diverse layouts, pedestrian densities, and rushing behavior, demonstrate improved prediction performance and safety over representative baselines. Real-robot experiments further demonstrate its robustness across various scenes. Our code and videos are available at https://coopuq.github.io/.
cs.AI / 122 / 2610.25892
Manifold-Constrained Residual Adaptation with Frame-Adaptive Layer Aggregation for Speaker Verification
Abstract
Large self-supervised speech models provide powerful representations for speaker verification, but full fine-tuning is expensive. Most conventional adapters insert small trainable modules to modify the output of each layer, while keeping the original residual connections unchanged. Building on our preliminary study of manifold-constrained Hyper-Connections in speaker encoders, we explore parameter-efficient adaptation of both residual paths and feature selection across layers. We first propose a Manifold-Constrained Residual Adapter (MCRA), which splits the hidden dimensions of each Transformer residual state into N groups and uses a small Sinkhorn-normalized matrix to exchange information among them. A learnable gate controls the balance between the original and mixed residual paths. The gate adds only one scalar parameter at each residual connection and initially favors the original path, allowing gradual adaptation. We further propose Frame-Adaptive Layer Aggregation (FALA), an attention-based aggregation method that uses a single learned query shared across all utterances to compute layer weights separately for each speech frame. Because the layer representations depend on the input, the resulting weights vary across both utterances and frames, unlike the fixed weights of a conventional layer-weighted sum. Compared with a frozen WavLM baseline using a fixed layer-weighted sum, our method achieves relative EER reductions of 22.4% on VoxCeleb1-O and 2.9% on CN-Celeb, with trainable adaptation parameters equivalent to only 0.0009% and 0.0016% of the WavLM-Large parameters, respectively.
cs.AI / 123 / 2610.25823
Optimizing Marketing Budgets for Social Influence: A Generalized Fractional Incentives Framework
Abstract
Marketing managers routinely allocate limited budgets across individuals in social networks whose responsiveness and influence vary substantially. We study the problem of Budgeted Generalized Fractional Influence Maximization (BGFIM), which captures the allocation of continuous levels of marketing spend across users in the presence of heterogeneous response and network spillovers. In this framework, adoption probabilities are governed by user-specific response functions that capture both baseline propensity and marginal responsiveness, while incentive costs vary across users. We propose a scalable solution approach based on a discrete greedy procedure with provable approximation guarantees, enabling tractable optimization under realistic budget constraints. Our formulation generalizes prior influence maximization models by allowing heterogeneous response and cost structures, providing a more flexible representation of targeted marketing interventions. Using experiments on multiple large-scale networks, we show that continuous allocation strategies can outperform standard heuristics and improve influence outcomes under heterogeneous conditions, and we illustrate how the framework can support profit-oriented decision making by identifying budget levels that balance marginal returns and costs. Overall, our results highlight the value of fine-grained budget allocation in networked marketing environments and offer a scalable approach for data-driven marketing resource allocation.
cs.AI / 124 / 2610.25603
When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition
Abstract
As AI voice agents get deployed in high-volume contact centers, firms can route low-intent leads to AI and allocate high-intent leads to human workers. Drawing on sequential search theory, we posit that AI-induced task allocation reshapes work, enriching opportunities facing human workers but also raising their reservation value for continuing with the lead in hand, making them less willing to persist. We study the effects of an AI voice-agent deployment at a Chinese online health-insurance brokerage using more than a half million customer outreach records surrounding the rollout. Adopting a regression-discontinuity-in-time design, we trace how worker behavior changes across the workday. Results show that the enrichment of human-facing opportunities concentrates early in the workday, raising the probability that a call is a high-intent lead by about 23 percentage points on the first three calls and fading to zero later. However, customer acquisition by human workers declines precisely where the enrichment is greatest, falling about 5 percentage points on the first three calls and flattening over the day. Consistent with the theorized reservation value mechanism, workers reallocate their earliest leads about 8 percentage points more often and persist less with them, shortening those calls by about 13%. A calibrated decomposition analysis further shows that the modest gain from improved lead ordering is outweighed by a much larger negative behavioral response. By reshaping the quality and within-workday distribution of opportunities, parallel AI deployment alters how workers pursue tasks, highlighting opportunity distribution as an important service operations state variable beyond workload and congestion.
cs.AI / 125 / 2610.25460
Reliability-Preserving Compression for Deployable Medical Image Segmentation
Abstract
Post-training compression of pretrained medical segmentation models commonly uses quantization, pruning, or distillation to reduce model size and computational cost. Existing medical segmentation quantization studies mainly evaluate Dice, model size, memory, and latency, but this is incomplete for safety-critical use because compressed models should also preserve uncertainty signals that indicate unreliable predictions. We study this problem in compressed MRI/CT segmentation and formalize why Dice can remain stable while uncertainty changes. Motivated by this failure mode, we propose score-guided aggregate-validated mixed precision (SGAV-MP), a post-training compression framework that selects layer precisions under reliability constraints. SGAV-MP ranks quantization actions with a compact benefit-risk score and accepts them only after full-model reliability validation on a calibration set. Across five 2D backbones and three 3D benchmarks, SGAV-MP preserves FP32-level Dice and high uncertainty-map agreement, achieving $3.59\times$--$7.8\times$ compression in 2D and $4.73\times$--$7.82\times$ in 3D. We also report backend-supported TensorRT deployment measurements separately to distinguish runtime execution from reliability-preserving storage compression.
cs.AI / 126 / 2610.26351
PhyPro-FM: A Physics and Prompt-Guided Graph Foundation Model for Contingency Analysis
Abstract
The continuous evolution of next-generation power systems, characterized by the high penetration of renewable energy resources and dynamic topology variations, poses formidable challenges to traditional N-1/N-2 contingency analysis in terms of computational efficiency and scalability. While deep learning-based methods offer accelerated inference, existing studies predominantly treat contingency analysis as a pure regression problem, often overlooking the physical essence of grid operation and engineering requirements of real-world scenarios. To bridge this gap, this paper proposes a physics and prompt-guided graph foundation model (PhyPro-FM) for generalizable and engineering-viable contingency analysis. The proposed architecture integrates graph neural networks with a learnable global query mechanism and, notably, introduces a prompt-guided learning paradigm. By leveraging historical snapshots as few-shot prompts, the model dynamically captures correlations between target and historical operating conditions, significantly enhancing generalization capabilities under unseen topologies. Furthermore, we design a physics-informed training mechanism that prioritizes branch power flow errors over nodal voltage errors, aligning the optimization objective with critical grid security constraints. To ensure engineering applicability, a physics-informed inference mechanism is also developed, incorporating a residual-based correction module and a specialized post-processing strategy for singular branches to circumvent numerical instability. Comprehensive experiments demonstrate that PhyPro-FM outperforms state-of-the-art methods in accuracy and generalization, validating its superior physical consistency and robustness for practical engineering applications.
cs.AI / 127 / 2610.25797
AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting
Abstract
Accurate real-time forecasting of atmospheric pollutants is essential for reducing exposure risks and improving public health. Traditional numerical forecasting systems suffer from incomplete parameterizations, uncertain inputs, and high computational cost. Existing deep learning approaches improve efficiency, but often entangle meteorological and pollutant variables within unified forecasting frameworks, limiting flexibility and pollutant-specific representation learning. To address these challenges, we propose AirMPA, a decoupled autoregressive framework for meteorology-to-pollution forecasting. AirMPA predicts 13 atmospheric pollutants at 0.4° spatial resolution and 12-h temporal resolution using meteorological fields, emission priors, and historical pollutant concentrations. The model is built on an improved Swin-Transformer architecture with zonal release attention and dual-branch fusion, enabling more effective modeling of global transport continuity and heterogeneous pollutant dynamics. During training, AirMPA uses the fifth-generation atmospheric reanalysis from the European Centre for Medium-Range Weather Forecasts (ERA5) as meteorological forcing, while during inference it directly ingests forecast fields from external meteorological models, enabling stable autoregressive prediction up to 5 days ahead. Experiments show that AirMPA consistently outperforms Aurora and surpasses the Copernicus Atmosphere Monitoring Service (CAMS) in approximately 90% of pollutant-lead-time evaluation cases over the 5-day forecast horizon, while outperforming CAMS on all variables beyond 48 h. These results demonstrate the potential of decoupled meteorology-to-pollution forecasting for global air quality prediction.
cs.AI / 128 / 2610.26279
Physics-anchored residual quantum neural networks for cloud-cover parameterization
Abstract
Compact quantum neural networks have been explored for cloud-cover parameterization, but their predictive value depends on how physical knowledge is incorporated and which classical models provide the reference. We investigate an offline residual formulation that retains a fitted \XR\ cloud-cover estimate and learns a correction to its logit in transformed target space. Using released ICON/DYAMOND-derived data, we compare ten model variants in six- and eight-input settings, with three matched replicates and 10,000 training samples per replicate. All quantum circuits use local Z readout and noiseless exact expectation values. Under a common 30-epoch protocol, residual prediction reduced physical-space test mean squared error relative to direct QNN prediction by 54.3\% and 38.8\% for six-input XYZ and ZZXY circuits, and by 49.2\% and 38.9\% for their eight-input versions. Residual QNNs also outperformed the selected compact residual multilayer perceptron in every paired replicate. A larger, longer-trained Grundner-Q1 adaptation nevertheless achieved lower error: residual Q1 reached mean MSEs of 0.00778 and 0.00791, compared with 0.01124 and 0.01082 for residual XYZ. The results support residual prediction for compact simulated QNNs within the tested budget and establish a clear boundary set by stronger classical references. They do not establish computational quantum advantage or online climate-model skill.
cs.AI / 129 / 2610.25721
Corrected Block Sparse Spectral Optimization: Exact Models, Finite Identification, and Local Dynamics
Abstract
The block generalized-power method pairs threshold-based support selection with inexpensive Stiefel updates, but the standard block formulation conflates two nonequivalent objectives and leaves inactive components, threshold ties, and post-selection refitting ill-specified. We rebuild the model from the primal. Eliminating each loading column from the normalized loading-ball problem gives a sum of soft-thresholded Euclidean norms (A-SUM); the squared-threshold objective in common use arises from a quadratic lift (A-SQ). We prove both identities, exhibit a two-component instance whose maximizers disagree, and specify explicit conventions for zero-response columns, ties, rank deficiency, and mask-before-refit. For A-SQ we separate the historical full-polar iteration from a safeguarded Riemannian ascent and prove, for the latter, whole-sequence convergence via a sufficient-ascent lemma, a relative-error lemma, and the Kurdyka-Lojasiewicz property, which applies because the soft-threshold objective and the Stiefel constraint are semi-algebraic. A strict threshold margin then forces finite identification of the support and sign pattern. The post-identification map is ordinary power iteration on the active covariance for a rank-one hard threshold, and the Frechet derivative of a coupled polar map for component-specific supports; a radius-based safe-screening rule follows from the same local analysis. Planted-spike experiments match the predicted identification time and local rates. On modern-AI panels (GPT-2 residual stream, ViT CLS stream, Pitprops) the thresholded block-sparse components match dense PCA on reconstruction and improve on it in monosemanticity and class specialisation. On three native-label scRNA datasets (Zheng17 FACS-sorted PBMCs, 68k donor-A with FACS-correlated labels, Tabula Muris) A-SQ beats dense PCA on cluster-recovery AMI by +0.023 to +0.044 at matched sparsity.
机器学习 (cs.LG)
138
cs.LG / 1 / 2610.25998
Internalizing Chain-of-thought into Looped Transformers
Abstract
Chain-of-thought (CoT) enables constant-bit-size transformers to perform general-purpose computation through generated tokens (Li and Wang, 2025a), whereas looped transformers carry out the same computation by recurrence in hidden states (latent thought). We construct a family of constant-bit-size looped transformers that interpolates between CoT and latent thought. For a prescribed loop count $L=L(n)$ and a fixed integer $k\ge1$ satisfying $s(n)^{1/k}=Ω(L)$, our transformer simulates a time-$t(n)$, space-$s(n)$ Turing machine using a context window of $O(s(n)/L)$ and $O(t(n)s(n)^{1/k}/L)$ generated tokens, reducing both resources by a factor of $L$ relative to the corresponding unlooped queue-machine simulation. The construction also holds with softmax attention at constant bit-size, under the finite-precision kernel specified in this paper.
cs.LG / 2 / 2610.25617
CheXFuse: Guided Differential Joint Attention Fusion for Robust Multimodal Chest X-Ray Classification in Data-Scarce Medical Settings
Abstract
A central challenge in multimodal learning is how to fuse modalities that differ substantially in predictive strength. When one modality provides a much stronger representation than another, indiscriminate cross-modal interaction can introduce weak, redundant, or unstable signals, limiting or even degrading performance relative to the strongest unimodal representation. To address this problem, we propose CheXFuse, a multimodal fusion framework based on Guided Differential Joint Attention (GuDJA). CheXFuse combines image and text token representations within a joint multimodal sequence, where learned token guidance emphasizes informative representations and differential attention reduces less useful attention responses. We evaluate CheXFuse on the Indiana University (IU) Chest X-ray dataset, using binary (normal vs. abnormal) classification as the downstream task. Experiments are conducted under both original-report and masked-report settings, with the latter serving as a controlled stress test under reduced textual informativeness. CheXFuse with GuDJA achieves the strongest mean overall performance across the evaluated baselines and ablations, reaching 95.73% accuracy, 95.79% F1-score, 96.54% recall, and 95.28% sensitivity at 95% specificity (Sens@95Spec) with original reports, and 95.58%, 95.60%, 95.60%, and 94.34%, respectively, under masked reports. These results suggest that GuDJA provides a promising fusion strategy under asymmetric or partially reduced modality informativeness.
cs.LG / 3 / 2610.26323
TriScore: Post-Hoc Out-of-Distribution Detection with Energy, Boundary Probes, and Transform Consistency
Abstract
Post-hoc out-of-distribution detection usually relies on a single signal extracted from a frozen classifier, such as confidence, energy, features, or gradients. This dependence becomes fragile under classifier-level adversarial shift, where the shift can preserve the detector's chosen cue while changing the model's decision behavior. We investigate whether this failure mode can be mitigated by combining complementary detection criteria instead of selecting a single score. We introduce TriScore, a post-hoc OOD detector that combines three ID-calibrated signals from the same frozen backbone: an energy prior, a boundary-fragility score, and a transform-consistency score. These signals are standardized on ID validation data and fused by an input-dependent ID-residual gate, requiring no OOD labels, auxiliary model, training, or stored feature bank. Across four ImageNet-scale backbones, TriScore achieves the best mean AUROC among 15 post-hoc baselines under a non-adaptive adversarial-shift protocol, while remaining competitive on standard OOD benchmarks. Ablations show that the gains come from combining complementary cues rather than optimizing a single detector signal. These results support multi-criterion post-hoc detection as a practical route toward more robust OOD detection under classifier-level shift. The source code is publicly available at github.com/gursoyege/TriScore.
cs.LG / 4 / 2610.26577
SafeWall: A Risk-Aware Job Walltime Recommendation for HPC Scheduling
Abstract
In high-performance computing (HPC), users must submit resource requirements for their jobs, with user-supplied walltime being a critical parameter. However, these estimates are notoriously inaccurate, as users tend to overestimate walltime to prevent the scheduler from terminating their jobs prematurely. While existing work on learning-based walltime prediction primarily focuses on maximizing prediction accuracy, such predictions are inherently risky for walltime recommendations. Specifically, underestimation leads to job termination before completion, resulting in lost progress and wasted resources. In this study, we present SafeWall, a novel recommendation method that adjusts walltime predictions by incorporating an explicit, nonnegative safety buffer to mitigate underestimation risk. The core of our approach is a risk-aware buffer that dynamically calibrates safety margins based on individual user's historical behavior. When user data is limited, the system employs statistical shrinkage to align these margins with a global baseline, ensuring robustness. This yields a lightweight mechanism that optimizes the trade-off between prediction accuracy and termination risk. SafeWall is modular and applicable to various learning-based prediction models. We evaluate SafeWall using five public workload traces, including the full-lifetime job log from the 4,392-node Theta machine at the Argonne Leadership Computing Facility. Across all traces, SafeWall substantially reduces underestimation risk compared to the unbuffered base predictors, while preserving competitive accuracy. The source code is publicly available at https://github.com/SPEAR-UIC/JSSPP26-SafeWall.
cs.LG / 5 / 2610.25524
Breaking the 1/3 Barrier for Approximate Nash Equilibria in Symmetric Bimatrix Games
Abstract
We give the first unconditional family of polynomial-time algorithms that breaks the $1/3$ approximation barrier for symmetric bimatrix games. For every fixed $δ>0$, an algorithm in this family computes a symmetric $(1-1/\sqrt{2}+δ)$-approximate Nash equilibrium, improving on the previous best guarantee of $(1/3+δ)$ due to Kontogiannis and Spirakis [2011].
cs.LG / 6 / 2610.26143
FLMNet: Self-Ensemble Distillation from Masks to Seeds for Single-Tower CTR Prediction
Abstract
Click-through rate (CTR) prediction underpins recommender systems and online advertising, where high-dimensional categorical features interact in combinatorially many ways. Several recent CTR models, such as FinalNet and QNN, report higher accuracy when a network also learns from its own averaged predictions, for example by running the same network under different dropout patterns and training each run toward the average. Why this self-distillation helps, and whether its random sampling is needed, remains unclear. This work shows that as the number of dropout patterns grows without bound, the objective reduces exactly to the label loss of the logit-averaged prediction plus a penalty on how much the dropout copies of the network disagree. When dropout does not affect how a network gates its features, this penalty can be approximated in closed form, and FLMNet is designed on this principle and trains without sampling. Networks trained from different seeds disagree in the same form, and distilling several of them into one network recovers part of their ensemble gain. On five benchmarks against seventeen baselines, the resulting models take the top two places in accuracy with a single network at inference, and the gains of the penalty and of distillation, where measurable, largely add up.
cs.LG / 7 / 2610.25450
Semi-supervised learning on singular data sets
Abstract
We propose to model singular data clouds by subanalytic sets, and develop an approach to semi-supervised learning in this setting. It relies on a Morrey-type embedding for Sobolev spaces relative to globally subanalytic measures with possibly singular support. As an illustration, we present a greedy method based on a Monte Carlo approximation of the Sobolev norm of the candidate solutions. We report numerical experiments on point clouds sampled from real algebraic varieties, as well as a test on a classification problem based on MNIST images downsampled into $7\times7$.
cs.LG / 8 / 2610.25469
Near-Optimal Last-Iterate Convergence of SGD in the Smooth Interpolation Regime
Abstract
We study the last iterate of stochastic gradient descent (SGD) with a constant stepsize on smooth convex objectives in the low-noise and interpolation regimes, where the variance $σ_\star^2$ of the stochastic gradients at the optimum is small or zero. We establish that after $T$ steps of SGD on $β$-smooth convex losses with any constant stepsize $0<η<2/β$, the last iterate attains expected excess risk $\widetilde O\bigl(\tfrac{1}{η(2-βη)^3T}+\tfracη{(2-βη)^4}σ_\star^2\bigr)$, which gives the rate $\smash{\widetilde O(1/T)}$ for every constant stepsize in the interpolation regime ($σ_\star=0$), matching the $Ω(1/T)$ lower bound up to logarithmic factors and improving upon the $O(1/\sqrt T)$ rate of Attia et al. (2026) for $η=1/β$. Our analysis is based on a new technique, which compares the second moments of the stochastic gradients along the SGD trajectory across only logarithmically many time scales. For quadratic losses in the interpolation regime, an adaptation of this technique gives an $O(1/T)$ rate for every constant stepsize, extending the recent optimal rate for the randomized Kaczmarz method of Derezinski and Dong (2026) to stepsizes above the greedy stepsize with an arguably simpler proof.
cs.LG / 9 / 2610.25510
KPI-Aware Multimodal Anomaly Detection for 5G and Open RAN Telemetry
Abstract
Reliable anomaly detection is a deployment bottleneck for 5G/Open RAN observability and emerging self-optimizing 6G systems, where hundreds of heterogeneous KPIs, severe class imbalance, and strict alert budgets make purely numerical scores brittle. We present KAC, a deployment-oriented KPI-aware multimodal anomaly-detection framework that uses KPI-level residual-text alignment to turn compact operator-style summaries into a learnable feedback signal over full-KPI temporal evidence. This gives the detector broad numerical coverage like modern time-series models, while retaining the sparse KPI-specific reasoning that makes operator feedback sustainable in a future operator-facing workflow. Across public telecom telemetry, enterprise-scale Open RAN observability data, and a Nokia benchmark built from real 5G RAN production packet traces with realism-hardened injected anomalies, KAC achieves strong threshold-based detection performance while remaining competitive on threshold-free ranking metrics. On TelecomTS, at a fixed 2.5 percent false-positive-rate operating point, KAC achieves the highest recall among the evaluated residual-based methods while maintaining comparable precision; across the evaluated benchmarks, it remains robust under substantial class imbalance. KAC is integrated in pre-production shadow mode with Nokia's GKE-based KPI construction pipeline, where it continuously scores telemetry windows alongside an internal packet-level anomaly detector for operational evaluation.
cs.LG / 10 / 2610.25511
Dynamics-Aligned Representation Learning under Induced Partial Observability
Abstract
Model-based reinforcement learning (MBRL) has enabled sample-efficient learning of continuous control policies from high-dimensional observations through learned latent dynamics models. These models are typically trained with observation reconstruction objectives, which provide strong supervision but can bias representations toward visual fidelity leading to an undertrained prior and unreliable long-horizon predictions. We propose a stochastic gating mechanism that induces controlled partial observability during training by sampling a spatial binary mask from a Beta--Bernoulli random field and applying it to the raw observation. The gate probability is resampled at each time step, exposing the encoder to conditions ranging from near-total occlusion to near-full visibility. Integration of our method requires minimal architectural changes from existing dynamics models: the gated observation and the binary mask are channel-wise concatenated and passed to the encoder as a single augmented input, with no modifications to the broader training pipeline. We evaluate the proposed approach on a set of pixel-based continuous control tasks selected from the DeepMind Control Suite. Using model predictive control for evaluation, which relies solely on open-loop model rollouts, we show that stochastic gating improves planning performance and data efficiency across tasks. Further analysis demonstrates that performance is robust to the spatial scale of the mask, and a balanced observability during training outperforms configurations biased toward either sparser or denser observations.
cs.LG / 11 / 2610.25523
FracTS: Hierarchical and Autoregressive Time Series Generation
Abstract
Time series are ubiquitous across different domains, and the growing use of synthetic data for privacy-preserving sharing, data augmentation, and stress testing makes time series generation increasingly important. Early generative approaches are predominantly autoregressive, which can suffer from low sample diversity, error accumulation, and difficulty modeling long-range dependencies. In this paper, we introduce FracTS, a hierarchical, coarse-to-fine autoregressive model for multivariate time series generation. Inspired by fractal generative modeling developed for images, FracTS adapts the paradigm from 2D to 1D sequences by learning a multi-scale representation that captures global trajectory structure at coarse scales and local dynamics at fine scales while following the autoregressive nature of time series. In contrast to unstructured images, time series are structured, allowing easy automated or designated temporal feature processing optimized for the model, and global (static) feature extraction describing the instances that generate the corresponding time series in the collection. We incorporate them in the generation process to further improve the generation quality. The design naturally handles multivariate dependencies and improves long-horizon temporal coherence without requiring fixed-length inputs. Experiments on five datasets (two conditional, three unconditional) conducted against seven baselines show that FracTS consistently outperforms baselines across marginal distribution and predictive ML utility, and correlation preservation in the unconditional setting. In particular, the predictive ML utility is averagely improved by 100.0% and 75.0-81.0% for unconditional and conditional time series datasets compared to the best baseline (relative error computed on dataset-averaged raw values). Code is available at: https://github.com/BetterdataLabs/fracts.
cs.LG / 12 / 2610.25544
Reinforcement learning on the EPRI AI-ccelerating Unit Commitment benchmark: diagnosing over-commitment and guiding MILP with learned policies
Abstract
Reinforcement learning (RL) is increasingly proposed as a fast solver for unit commitment (UC), but standalone agents remain far more expensive than mixed-integer linear programming (MILP). In the first study of RL on the 51-generator, 72-hour benchmark of the Electric Power Research Institute (EPRI) AI-ccelerating Unit Commitment competition, we trace this gap to the agents' own exploration, close most of it, and show that the resulting policy can accelerate an exact solver. In every configuration tested, Proximal Policy Optimisation (PPO) and an action-branching Deep Q-Network (DQN) trained from scratch committed about twice the units of a thermal-only MILP optimum, at 39-64% higher cost, whereas a merit-order rule came within 1.6-7% (10-24% when unserved energy is priced). When each unit's decision is sampled independently, spare units insure against random de-commitments. Consistent with this, injecting exploration noise during fine-tuning enlarged a lean policy's fleet, a reward without shaping terms did not remove the over-commitment, and under their training noise the agents outperformed the rule subjected to comparable noise. Imitating the MILP solutions and fine-tuning conservatively brought PPO within 2.6% of the optimum, and these policies had the lowest penalised cost of any evaluated (about 5%). Fixing the fine-tuned policy's confident decisions made the MILP 3.8-8.5 times faster within 0.10-0.25% of the best solution, with a smaller worst case than an imitation-only policy fixing as many variables (1.1% against 8.9%). Because dispatch method and benchmark scope alone shift the reported optimality ratio by up to 3.6% and 9%, we propose a reporting protocol.
cs.LG / 13 / 2610.25545
Beyond the record: evaluating probabilistic AI weather forecasts deep in the tail
Abstract
Can probabilistic artificial intelligence (AI) weather forecasts retain skill beyond their training domain? We compare four state-of-the-art AI models (ArchesWeatherGen, U-Cast, WeatherNext2 and GenCast-mini) with the physics-based IFS-ENS over land in 2022, at 1.5° resolution and lead times of 2--10 days. Monthly ERA5 records from 1979--2021 define common historical boundaries, and we characterize new records set in 2022 by their depth, i.e., the exceedance margin relative to local historical variability. Two proper scores, record-threshold weighted CRPS and Brier score, assess the magnitude and probability of record exceedances across all eligible cases, retaining both missed extremes and false alarms. Verified against ERA5, AI models outperform IFS-ENS for heat and cold records at almost every lead and depth. For heat, the advantage strengthens with depth, partly because IFS-ENS over-forecasts deep records, and largely persists for deeper records under HRES-fc0 verification. For wind, gains erode for deeper records and longer leads. Better tail scores can nevertheless coexist with aggregate probability deficits or surpluses, and some improvements arise chiefly on non-exceedances. Increasing underprediction on progressively more extreme observed subsets can also arise through outcome selection, even for an ideal conditional-mean forecast. These findings establish that native probabilistic AI ensembles retain, and for heat increase, their advantage over IFS-ENS as thresholds move beyond local records, although all advantages narrow with lead time. They caution against generalizing the extrapolation failure reported for deterministic AI forecasts to probabilistic systems, and motivate common verification conventions that separate comparative tail skill, performance on realized extremes, and probabilistic calibration.
cs.LG / 14 / 2610.25558
FTFM: Factorized Tabular Foundation Models
Abstract
Tabular foundation models (TFMs) are becoming a powerful alternative to conventional tabular learning, adapting to new datasets in context without task-specific training. Current architectures, however, face a trade-off: models that preserve cell-level representations are expensive because they repeatedly attend across rows and columns, while faster models collapse each row into a fixed embedding and lose this fine-grained state. We introduce Factorized Tabular Foundation Models (FTFM), which preserve cell representations while factorizing their computation through one contextual latent per row and one per column. Cells no longer attend to one another directly; instead, each cell is updated through the interaction of its row and column latents, reducing repeated attention while retaining feature-specific information. FTFM achieves competitive predictive performance in tabular classification on major benchmarks such as TabArena and BeyondArena.
cs.LG / 15 / 2610.25575
Benchmark for Structured Retrieval-Augmented Generation in Perovskite Photovoltaics
Abstract
While retrieval-augmented language models increasingly serve as literature assistants in the physical sciences, existing benchmarks evaluate them primarily on qualitative text overlap or through model-based scoring rather than against verified experimental ground truth. In experimental materials disciplines such as perovskite photovoltaics, however, the most critical inquiries are inherently quantitative and aggregate measurements across dozens or hundreds of papers. Here, we present a quantitative question-answering benchmark where all 755 ground-truth answers are computed directly from 45,441 measured device records (40,889 human-curated records through 2021 and 4,552 records extracted from 2022-2026 literature). Benchmarking six retrieval architectures over 158,274 passages from 25,662 papers using an open 14B model on a single 24 GB GPU, standard passage retrieval tops out at 16-23% accuracy (compared to 1.6-6.5% zero-shot). In contrast, equipping the language model with read-only SQL execution over the structured device table elevates accuracy to 51.5% on historical and 50.8% on prospective literature -- a 33.01.8 percentage-point advantage over hybrid text retrieval. Aggregate queries spanning extrema, distributions, and filtered record counts prove virtually unsolvable via text retrieval alone, demonstrating that systematic table extraction is an operational prerequisite for literature-scale quantitative inquiry.
cs.LG / 16 / 2610.25577
Finite Interaction Lens: Structured Precision Sharing for Vision-Language Models
Abstract
Identifying which visual interventions interact in a vision-language model requires comparing their joint and separate effects. The resulting four-output difference can be small enough for rounding error to obscure it. We introduce the Finite Interaction Lens (FIL), a mixed-precision method for evaluating finite intervention families. At each selected linear projection, FIL decomposes each candidate's input into a native state, two single-intervention components, and a full-dimensional mixed residual. It projects the single-intervention components at higher precision and reuses the same numerical contributions in the joint candidates, while retaining every mixed residual and nonlinear trajectory. We derive a local cancellation identity and a conditional bound on error propagation through a smooth nonlinearity. In one MLP of Qwen3-VL-4B-Instruct, FIL reduces final interaction MSE by 94.4% on 16 synthetic families and by 99.85% and 99.35% in two separate four-image COCO studies, relative to a control that has the same nominal BF16-product budget and higher-precision single interventions but projects each joint candidate independently. One image accounts for most of the absolute gain in each COCO study. Higher-accuracy multi-product BF16 arithmetic on all candidates gives lower aggregate error at similar total query times. Single-product BF16 already recovers the reference choices in both candidate-selection tasks we tested. On these development inputs, FIL improves interaction fidelity at a fixed nominal precision budget. Its practical value depends on a task that needs the extra fidelity.
cs.LG / 17 / 2610.25584
CVTA: Cross-Variable Temporal Attention for Multivariate Irregular Time Series Prediction
Abstract
Irregularly sampled multivariate time series arise in many real-world domains such as healthcare, climate science, and large-scale monitoring systems. Characterized by asynchronous observations and heterogeneity across variables, such time series pose significant challenges for forecasting. In this work, we introduce Cross-Variable Temporal Attention (CVTA), a short-term, continuous-time forecasting framework that learns prediction functions from its compact, interpretable state summarizing recent temporal dynamics across variables. CVTA uses Variable-Conditioned Embeddings to project heterogeneous variables into variable-specific latent spaces, and applies temporal attention to capture cross-variable dependencies from this state. Furthermore, we demonstrate that CVTA can be integrated as an auxiliary module into existing state-of-the-art prediction models through a dynamic meta-decision model. Extensive experiments on multiple real-world benchmarks show that CVTA achieves strong short-term predictive accuracy, supports efficient online inference, improves existing prediction models when used as an auxiliary module, and learns attention patterns that align well with domain knowledge.
cs.LG / 18 / 2610.25591
Distributionally-Robust Policy Learning with Observational Data
Abstract
Policy learning from observational data is fundamental to applications ranging from personalized medicine to welfare-program targeting, but standard approaches optimize empirical performance and may fail when the deployment distribution differs from the training distribution. We study \emph{distributionally-robust policy learning}: given observational data, we seek a policy $π\in Π$ that maximizes the worst-case policy value over an $f$-divergence ball of radius $ρ$ around the data-generating distribution. We propose an estimator built on cross-fitted doubly-robust scores for the conditional treatment effect, paired with a DRO solver that handles \texttt{KL}, $χ^2$, and \texttt{CVaR} ambiguity sets. We establish two convergence guarantees. First, the estimated policy attains the standard $\widetilde{O}(n^{-1/2})$ rate on its DRO regret, and inherits the doubly-robust property -- the rate continues to hold whenever either the outcome model or the propensity model is consistently estimated, even at slow nonparametric rates. We then propose a sample-splitting bias-correction algorithm derived from the Lagrangian dual, and show that under a Tsybakov-style margin condition on the treatment-effect, the estimator attains a fast rate of $\widetilde{O}(n^{-(1+α)/(2+α)})$ , interpolating between $\widetilde{O}(1/\sqrt{n})$ and $\widetilde{O}(1/n)$ as the margin parameter $α$ ranges over $[0,\infty)$. To our knowledge, this is the first work to incorporate a margin condition into the analysis of distributionally-robust policy learning, and the first to obtain fast rates in this setting. Empirically, on synthetic and semi-synthetic experiments calibrated to a welfare-to-work program, distributional robustness yields meaningful improvements over empirical welfare maximization when training and test distributions differ.
cs.LG / 19 / 2610.25599
Text Dominance in Paralinguistic Conflict: Tracing Suppression Through Audio LLM Circuits
Abstract
Audio LLMs deployed in settings such as crisis hotlines and customer support are only useful beyond speech-to-text pipelines if they use acoustic cues rather than the transcript alone. We test whether they prioritise transcript content over paralinguistic cues (emotion, speaker identity, sarcasm) and trace the mechanism when they do. Across six audio LLMs under LISTEN-style conflict conditions, the mean text-dominance rate on mismatched text+audio items is 0.78 (0.86 excluding sarcasm). In the most text-dominant screened model, probing shows speaker identity remains linearly decodable at ceiling at the language model's input, so the failure is not perceptual. Back-patching, reported to repair lexical text dominance, has no measurable effect here; but it also fails on its own source task in our replication (n=300), so we treat its null as uninformative rather than as evidence of non-transfer. Attention knockout partially restores audio-grounded output by blocking the claim from propagating through intermediate tokens during open-ended generation, whereas blocking the answer position's direct connection to the conflicting claim has no effect. The claim's influence therefore travels through an intermediate relay rather than a single severable edge. The effect appears only in free-form generation and not in closed-set multiple choice, suggesting that benchmarks relying on that format may be systematically blind to interventions of this kind. Code and data are publicly available at https://github.com/AdvayRoongta/audio-llm-mech-interp.
cs.LG / 20 / 2610.25604
Adversarial Post-Training of Foundation Models for High-Frequency Trading
Abstract
While autoregressive foundation models trained on limit order book (LOB) data show state-of-the-art performance for market simulation, pre-training with cross-entropy neither penalizes the compounding of sampling errors over free-running rollouts nor targets desired market properties. We introduce a post-training method for LOB foundation models using a Wasserstein generative adversarial network (GAN) objective. A critic, trained online on rollouts through the true, non-differentiable matching engine, guides updates from evolution strategies over rank-4 LoRA perturbations of the pre-trained generator, under a KL trust region. For efficiency, we use EGGROLL (Evolution Guided GeneRal Optimisation via Low-rank Learning), a memory-efficient variant of evolution strategies. On a held-out panel, adversarial post-training improves LOB-Bench realism by 17.6% (WS-21) and 17.0% (Mean-$L_1$), while selection-matched random-perturbation and compute-matched continued-pre-training controls yield no improvement. Post-training is thus a necessary stage that more pre-training does not replace. Per-position divergence measurements show that it slows error accumulation over 500-message rollouts from $3.1\times$ to $2.3\times$. Adversarial post-training is also inexpensive, requiring under 50M training tokens and 12 GPU-hours per seed, three orders of magnitude below the 48B-token pre-training budget. Ablations over critic inputs, optimizers, and the population-rollout trade-off, together with an analysis of reward hacking, yield practical design guidelines for adversarial post-training.
cs.LG / 21 / 2610.25605
PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization
Abstract
Solubility depends on intermolecular interactions between solutes and solvents, and is central to drug discovery, chemical and materials process engineering. Its prediction requires out-of-distribution (OOD) generalization to unseen solutes, solvents, and solvent mixtures, where supervised methods suffer from data scarcity, and existing self-supervised methods do not explicitly learn chemically meaningful relations. Learning chemically meaningful relational information without labels can alleviate data scarcity and improve OOD generalization, but remains challenging. We introduce Prior Relational Information for Self-supervised Modeling (PRISM) and apply it to solubility OOD generalization. PRISM is a label-free pretraining framework that converts expert-defined chemical priors into relational teachers. For solubility prediction, it trains four encoders to preserve the population-level relational information from priors capturing hydrophobic, electrostatic, polarizability, and substructural effects, using Kullback-Leibler (KL) divergence to match distance-based neighborhood distributions during pretraining. Across scaffold, solute, and solvent OOD settings, PRISM matches or surpasses supervised and self-supervised baselines. Under mixed-solvent OOD, freezing the four PRISM encoders and training a lightweight fusion-and-composition head outperforms strong baselines and substantially reduces run-to-run variation. Ablations and control studies confirm that gains stem from chemically meaningful relational structures baked into the encoders by PRISM. We expect this strategy of turning relational priors into self-supervision signals to extend beyond solubility prediction to domains where chemically or biologically meaningful relational priors are available.
cs.LG / 22 / 2610.25611
Safety Offset Learning: When Benign Fine-Tuning Misaligns LLMs
Abstract
Safety risks from fine-tuning are commonly characterized by the harmfulness of the training data:harmful fine-tuning is considered dangerous, while benign fine-tuning is often assumed to preservealignment. We argue that this view misses a more fundamental factor-the behavioral transition inducedby supervision. Harmful fine-tuning and safety-degrading benign fine-tuning can both be understoodas learning the same refusal-to-compliance transition, but on different data distributions. We studywhether such a transition learned on benign inputs can generalize to harmful requests. We constructbenign examples that an aligned model initially tends to refuse and pair them with their originalharmless responses, creating a discrepancy between the model's initial refusal and the supervisedanswer. Fine-tuning on these examples increases harmful compliance, but only when two conditionshold together: the discrepancy is large enough, and it cannot be resolved through a narrow surface-levelshortcut. With the induced discrepancy matched, varying only whether it is localized changes harmfulcompliance by an order of magnitude; raising the discrepancy does not help while a shortcut remainsavailable, and removing the shortcut does not help while the discrepancy is small. Building on thisobservation, we introduce SAFEOFF, a dataset of benign instructions rewritten to raise the alignedmodel's refusal likelihood and paired with their original answers. Across three aligned LLMs and threesafety benchmarks, SAFEOFF substantially weakens safety alignment despite containing no explicitharmful supervision, suggesting that fine-tuning safety depends not only on what data a model sees, butalso on which behavioral boundary the supervision moves.
cs.LG / 23 / 2610.25612
Curriculum Learning Teaches LLMs Long-Horizon Implicit Reasoning, Up To A Limit
Abstract
LLMs "reason" by externalizing intermediate steps as tokens through chain-of-thought (CoT), yet even frontier reasoning models hit abrupt performance cliffs on long-horizon tasks. Implicit reasoning (reasoning within a single forward pass) is useful during CoT generation, as the model can "search ahead" to decide the correct next move without explicit search. Models with weak implicit reasoning are more likely to commit to wrong paths and must attempt to recover via backtracking, while models with strong implicit reasoning identify the correct proof strategy directly, shortening the CoT. However, prior work shows LLMs struggle to reason implicitly beyond shallow depths, with no relief from scale. We investigate whether curriculum learning can extend this depth by fine-tuning on examples with progressively increasing vocabulary size and lookahead (the number of steps a model must search ahead to identify the correct next move). At matched compute, curriculum-trained models achieve lookaheads where no-curriculum training fails. Qwen3-0.6B reaches 98% accuracy on problems with lookahead up to 256 with curriculum, compared to 45% without. However, progress slows as fine-tuning continues, following a saturating exponential that suggests a maximum intrinsic lookahead to which models can learn to reason implicitly. Larger models reach higher fitted ceilings, fine-tuning pretrained models far outpaces training from scratch, and curriculum step sizes that are too large stall progression. Finally, the capability acquired on synthetic data transfers to natural language reasoning, with curriculum-trained models outperforming models trained without curriculum at matched compute on 11 natural language reasoning benchmarks.
cs.LG / 24 / 2610.25621
Weighted State Souping for Efficient Long-Context Retrieval
Abstract
State space models (SSMs) combine fast inference and fixed state size with competitive language-modeling performance. Recently, state souping has been proposed as a procedure whereby an SSM independently encodes chunks of tokens like documents, tool definitions, or memory records. The resulting state encodings may be cached and recombined to cheaply construct a context for downstream queries. However, both state souping and vanilla SSMs deteriorate on long, crowded contexts, limiting their utility for long-horizon domains. To address this issue, we propose weighted state souping, a simple method that learns query-dependent weights for composing cached states. The resulting models improve long-context performance substantially across both single- and multi-hop document question answering, needle-in-a-haystack retrieval, variable tracking, and tool use. Our method can be further adapted into an efficient sparse variant, yielding near-constant encoding cost with respect to context length while preserving (and sometimes improving) performance. We also present simple theoretical arguments that illustrate why souping works and when it may fail, and discuss informally the advantages of weighted souping. Weighted state souping thus combines reusable context representations with selective retrieval, offering an efficient approach to language modeling over large collections of documents, tools, and other forms of structured context that increasingly dominate modern agentic workloads.
cs.LG / 25 / 2610.25625
Zephyrus-R1: Agentic Reinforcement Learning for Weather Reporting
Abstract
Reinforcement learning (RL) has emerged as a key method for training AI models in scientific applications, but RL remains difficult when outputs cannot easily be verified against a known ground-truth answer. Natural-language weather forecast reports are one such domain: they translate atmospheric data into assessments of weather evolution, hazards, and uncertainty for both the general public and key decision-makers. Unlike numerical weather-analysis questions, they admit many valid outputs, which limits deterministic verification. To address this, we develop Zephyrus-R1, a practical framework for RL on scientific weather forecast reporting. We focus on three key components of the agentic RL pipeline: reward design, curriculum learning, and an analysis of human-expert alignment. To encourage both report legibility and factual accuracy, we design a reward function that combines a meteorological rubric with bidirectional claim entailment, both evaluated against National Weather Service reference reports. Our framework optimizes this reward in a simulated tool-calling environment, enabling vision-language agents to gather and analyze forecast evidence across regions and lead times before producing a report. We use annotations from three meteorologists to audit the claim labels produced by our LLM judges. Finally, a curriculum helps small models first develop tool use and scientific reasoning through RL on verifiable weather-analysis questions, then optimize the open-ended report reward. After training Qwen3.5-4B with Zephyrus-R1, we improve the mean across six judge-metric scores by 26% over the base model and achieve results competitive with frontier open-weight models. For Qwen3.8-27B, we improve the same mean by 13% over the base model, by 6.7% over Kimi-K3, and by 2.7% over GPT-5.6-Luna.
cs.LG / 26 / 2610.25627
Conditional Average Treatment Effect Distillation for Hypothetical State-Intervention Targeting Under Budget Constraints
Abstract
Personalized public-health targeting must capture heterogeneous intervention benefits while remaining interpretable under resource constraints. Flexible causal-effect models can be hard to operationalize, while coarse subgroup policies may lose individualized effect information. We study budget-constrained hypothetical state interventions in observational data without a logged historical decision policy. For obesity, elevated glucose, and lifetime smoking-history states, causal-structure-informed pretreatment covariates are prespecified, with capacity capped at 70% of those observed in the adverse state. We evaluate causal distillation where Causal Forest and Bayesian Causal Forest teachers estimate conditional average treatment effects (CATEs), which decision tree and random forest students approximate. We compare individualized policies with unsupervised K-means subgroup policies and supervised causal-tree policies with leaf counts matched to K-means cluster counts. Observational policies are assessed on held-out data using a doubly robust estimator, bootstrap uncertainty, and decision-focused teacher-student fidelity measures. Across PIMA obesity and glucose and NHANES smoking-history experiments, each distilled student was within 0.011 absolute utility of its teacher, with allocation Jaccard similarities of 0.738-0.937. K-means and causal-tree policies were also competitive, with the highest point-estimate policy varying by application. Prespecified paired comparisons were nonsignificant after Holm adjustment, while near-budget-boundary agreement was less stable than global fidelity for some teacher-student pairs. Findings support causal distillation for translating flexible effect models into simpler budgeted targeting rules, although observational estimates depend on causal-identification assumptions and require longitudinal or prospective validation.
cs.LG / 27 / 2610.25651
OPDArena: Understanding Learning and Forgetting in On-Policy Distillation
Abstract
On-policy distillation (OPD) is increasingly used for language model post-training, yet how teacher supervision shapes learning and forgetting remains unclear. We present OPDArena, a framework comparing supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and OPD across five pretrained models and diverse teacher sources. It jointly evaluates mathematical learning, general capability retention, and predictive distribution changes. Teacher choice strongly shapes retention: students of RLVR-trained teachers improve on the target task while retaining prior capabilities, whereas students of SFT-trained teachers incur greater capability loss. Across post-training methods, out-of-domain (OOD) distribution change shows a more consistent positive association with forgetting than in-domain change. To understand how teacher choice affects these changes, we examine the additional information teachers provide relative to the student's pretrained model. Controlled teacher comparisons show that the task-specific preferences teachers acquire through post-training make student updates more concentrated in-domain.
cs.LG / 28 / 2610.25652
A "feature ODE" describing the learning behavior of shallow MLPs on simple functions
Abstract
We study the gradient flow training dynamics of two-layer multi-layer perceptrons (MLPs) on anisotropic Gaussian data. Even in this simple setting, solving the dynamics in full is notoriously difficult, so we try something different. We propose a "feature ODE" that might be intuitively expected to describe the learning dynamics of a shallow MLP. We find that MLPs trained to learn simple target functions (low-order Hermite polynomials and staircase functions) actually roughly follow the loss trajectories from this feature ODE. This agreement is a surprise because of the size of the simplification and the fact that it does not come from a rigorous chain of approximations. We discuss this surprise and speculate on its reasons and uses.
cs.LG / 29 / 2610.25663
Can Metadata Fix the Gauge? Calibration Turns Sparse Multi-Source Learning from Sparse PCA into Sparse Mean Recovery
Abstract
Source labels tell us where a sample came from, but not how that source oriented its coordinates. In a sparse Gaussian multi-source model with source-specific sign gauges, that distinction changes the learning problem. Without calibration, pooling across sources erases the mean signal, and the source-labelled experiment reduces to sparse PCA. With task-independent relative-gauge metadata, synchronization aligns sources and the same observations become a sparse mean problem. We formalize the value of metadata through AlignedMass, which quantifies how much first-order signal a metadata-induced alignment restores. Source labels create only limited AlignedMass, whereas valid calibration can create AlignedMass proportional to the full sample size. This yields a statistical separation in micro-source regimes and, under standard sparse-PCA hardness assumptions, a computational separation as well. Synthetic experiments check the predicted scaling; controlled real-feature and physical sensor studies examine sign cancellation, separately acquired calibration, and reuse across tasks.
cs.LG / 30 / 2610.25689
MT-CC: Multi-Group Temperature Scaling for Asymmetric Calibration Behavior in Class-Incremental Learning
Abstract
As deep neural networks are required to continuously adapt to evolving data, class-incremental learning (CIL) has become an active research area. However, most existing works primarily focus on improving accuracy while overlooking confidence calibration, which measures the trustworthiness of the model's predictions. Although temperature scaling (TS) is effective for static calibration, it has rarely been explored in continual settings where data distributions evolve over time. In this paper, we introduce Multi-Group Temperature Scaling for Continual Calibration (MT-CC), a novel post-hoc calibration framework that addresses the asymmetric calibration behavior in CIL, where the rate of confidence changes does not align with the rate of accuracy degradation across tasks, leading to different levels of miscalibration. Our key contribution is to design multiple temperature parameters through a lightweight grouping network that separates samples by confidence statistics and task-aware signals in CIL. Furthermore, a Groupwise Difference-between-Confidence-and-Accuracy (GDCA) regularization is incorporated to promote intra-group consistency. Experiments demonstrate that MT-CC significantly improves calibration performance while maintaining accuracy, offering a principled framework for reliable continual learning.
cs.LG / 31 / 2610.25699
Stochastic Tracking of Drifting Affine Fixed Points
Abstract
Linear stochastic approximation (SA) is widely used to estimate fixed points from noisy observations, but most finite-time analyses assume that the underlying operator remains fixed. We study linear SA with time-varying affine operators, where the fixed points change over time and the observation noise may be correlated across rounds. For constant-stepsize updates, we establish a bound on the cumulative expected squared distance between the iterates and the corresponding fixed points. The bound accounts for the initial error, observation noise, and changes in the fixed points over time. We also establish lower bounds for scalar Gaussian models, showing that, under specified conditions, a suitably chosen constant stepsize achieves the smallest possible worst-case cumulative error up to constant factors, even compared with estimators that use future observations. These results further quantify the effect of temporal correlation on estimation accuracy. Since choosing an appropriate stepsize generally requires knowledge of the noise level and how much the fixed points change, we develop an adaptive method that uses an additional bounded, unbiased observation of the fixed point to achieve an almost optimal algorithm without needing the knowledge of the noise level and the fixed point variation over time. Finally, we apply our results to value-function estimation using synchronous tabular temporal-difference learning in changing environments, assuming generative-model access and a common invariant distribution.
cs.LG / 32 / 2610.25704
Predictive Bandit Curriculum Reinforcement Learning via Gradient-Based Sample Group Discovery
Abstract
Curriculum learning, which organizes training samples according to their learning priority, has been widely adopted in reinforcement learning to improve training efficiency. However, existing methods commonly construct curriculum units using predefined data properties and estimate their utility by simply aggregating a single proxy signal from recent training history.Such designs may group samples with different optimization effects and fail to anticipate how their utility evolves in relation to the policy. We propose Predictive Gradient-based Bandit (PGB) Curriculum, which first discovers optimization-aware learning units by clustering projected sample-level gradients. PGB then employs a lightweight temporal predictor to forecast the future utility of each unit and converts the predictions into a Boltzmann sampling distribution for adaptive data allocation. Experiments across multiple model scales and reasoning benchmarks show that PGB consistently outperforms standard GRPO, difficulty-based curricula, and history-based scheduling strategies, demonstrating improved data utilization and training efficiency.
cs.LG / 33 / 2610.25710
Evidence Sufficiency and Model Comparisons in MexRegQA: An Item-Level Audit of Mexican-Spanish Regulatory QA
Abstract
Similarity-gated coverage does not establish that retrieved evidence contains the facts needed to answer a question. We audit Mexican-Spanish financial-regulatory QA and find that 35 of 135 covered items lack some or all load-bearing facts. On the post-hoc full-evidence slice (n=98), eleven prompt-matched models score 95-98/98 and none differs significantly from the anchor; this does not establish equivalence. Removing evidence-defective items changes the nominal significance status of 14 of 55 pairwise comparisons. On 56 zero-evidence items, accuracy spans 33.9%-89.3%. Removing evidence from the audited items under an abstention-permitting prompt reduces accuracy by 63-81 percentage points in all ten tested legs, with no recovered items. Building on sufficient-context research, we contribute MexRegQA v1.0: 192 items with provenance and evidence-status labels, frozen retrieval evidence for 191 evaluated items, predictions, judge outputs and offline checks. The audit was outcome-aware, the later model/panel sweep reused the same questions, and all reported scores and labels rely on machine evaluation without independent human validation. We propose releasing evidence-status labels alongside grounded QA benchmarks so that model comparisons can be examined separately from evidence availability.
cs.LG / 34 / 2610.25729
Neural Video Coding under Limited Delay: Theoretical Optimality and System Design
Abstract
Traditional neural video compression (NVC) methods predominantly follow strictly causal predictive-coding paradigm, treating zero-frame delay as a fixed constraint rather than delay as an explicit design parameter. From an information-theoretic perspective, allowing limited frame delay enables a codec to exploit correlations with future frames within the delay window, thereby reducing uncertainty that is inaccessible to strictly causal prediction and potentially improving the rate-distortion tradeoff. In this work, we formalize this intuition through a rigorous rate-distortion analysis builds on the clique factorization of Markov networks, and propose Graph-Factored Sequential Coding (GFSC), a principled framework for sequential video coding under limited delay. To our knowledge, we provide the first theoretical result showing that, for general $K$-order Markov sources with positive distributions, a $K$-frame delay is sufficient to attain joint-coding rate-distortion optimality over a nontrivial distortion region, a guarantee unavailable to zero-delay causal coding in general. In practice, we instantiate GFSC through lightweight neural modules, allowing existing causal predictive codecs to be augmented without modifying their core architectures. Experiments across multiple NVC models and datasets show that GFSC consistently improves rate-distortion performance over the corresponding causal counterparts. Under matched delay budgets, GFSC also achieves competitive performance compared with representative bidirectional codecs.
cs.LG / 35 / 2610.25747
Bayesian Model Discovery with Language Models: Closing the Loop between Hypotheses and Experiments
Abstract
Scientific model discovery requires proposing candidate models of an unknown system and designing experiments that discriminate among them. However, LLM-based approaches typically rely on in-context learning to predict outcomes and select experiments, lacking explicit model evaluation to systematically guide hypothesis search, while Bayesian approaches enable principled model comparison but are confined to a predefined candidate set which results in limited search spaces. To this end, we introduce BMD, which closes the loop between LLM-driven hypothesis generation and Bayesian experimental design. Specifically, an LLM proposes model families, while a numerical engine fits their parameters, weighs them by model evidence, and averages their predictions. Then, BMD chooses the next experimental conditions by maximizing the joint expected information gain over families and their parameters, while feeding the updated beliefs and residual diagnostics back to the LLM to guide its next proposals. Across four scientific discovery benchmarks, BMD achieves lowest final prediction error than in-context learning and LLM-based scientific agents while recovering exact functional forms more often, with ablations supporting numerical feedback and the joint objective.
cs.LG / 36 / 2610.25749
Cognitively Decoupled Meta-Induction for Out-of-Distribution Generalization
Abstract
In-context learning (ICL) tackles downstream tasks conditioned on demonstrations. This paradigm relies heavily on neighbor-based matching between queries and demonstrations, limiting out-of-distribution (OOD) performance. For better generalization, rule-based inductive reasoning derives task-level regularities, which we term task-level inductive representations (TIR), before directly answering. However, existing methods lack generalizable inductive ability across diverse tasks. Inspired by meta-learning, we propose Cognitive-Decoupling meta-Induction (CoDI) for TIR-based inductive reasoning on general tasks. CoDI comprises a Cognitive-Decoupling Framework for query-blind TIR induction, and a Dual-Loop Meta-Inductive Learning method for cross-task inducer optimization. Empirically, CoDI achieves substantial performance gains on five ID and OOD benchmarks, and preserves OOD gains without explicit induction at inference. Code is available at https://github.com/Yuqian-Wang2277/codi-meta-induction.
cs.LG / 37 / 2610.25752
Which Answers Agree? Evaluating Self-Consistency from Incomplete Logs
Abstract
Self-consistency answers a question by generating multiple responses and selecting the most frequent answer. Its evaluation requires both correctness labels and information about which responses agree; correctness labels alone generally determine neither voting accuracy nor the benefit of additional votes. For tasks in which all correct responses map to the same normalized answer, we derive sharp accuracy bounds using classical stochastic majorization, refined by abstention counts and answer-category limits. We also derive sharp bounds on accuracy changes across voting budgets that share a response pool. When answer-category counts are available, we introduce Plurality-U, a count-based dynamic program that exactly averages voting accuracy over all fixed-size subsets, eliminating finite-pool Monte Carlo error. On 1,403,520 stored outputs from recent reasoning models in Scorio Lite and Scorio Trace, recording abstention counts and answer options cuts the width of SuperGPQA accuracy intervals by more than half. For moving from 9 to 17 votes, our bounds reduce interval width by a median of 77.7% on SuperGPQA and 72.3% on 14 GPQA and Adaptive-Consistency pools, compared with a baseline that combines separate accuracy bounds. Exact computation has lower paired median runtime than 2,000-subset Monte Carlo in 14 of 15 tested pool-budget settings. These results clarify what incomplete logs can establish and provide an exact evaluation procedure when answer counts are retained.
cs.LG / 38 / 2610.25758
Remember Less, Adapt Longer: Bounded-State In-Context Reinforcement Learning with Recurrent Summary Memory
Abstract
In-context reinforcement learning (ICRL) enables agents to adapt to new tasks without parameter updates, relying on memory to retain what they learn from experience. For agents operating continually, existing approaches using memory trade off bounded cost against access to past experience. Approaches using growing memories, such as full-history attention and accumulating summaries, incur storage and per-decision costs that increase with task length. Fixed-size memories bound storage but restrict access: sliding windows discard older steps, while recurrent networks compress even recent interactions into a hidden state. We introduce Recurrent Summary Memory (RSM), a Transformer policy that attends exactly to a short working buffer of recent interactions and carries a fixed-capacity summary of earlier experience. When the buffer fills, RSM rewrites the summary from its previous contents and the buffer, then clears the buffer, so deployment memory and per-decision cost stay $O(1)$ in task length. On long-running partially observable tasks, frozen RSM policies sustain their adapted performance for eight times the training horizon with about 1% of full-history memory, ending ahead of every evaluated baseline while growing memories deteriorate. When the task configuration changes mid-run, RSM re-adapts better than growing and recurrent memories. Clearing the summary or blocking the policy's access to it removes its benefit, while swapping in another task's summary steers behaviour towards that task. Together, these results show that our fixed-capacity summary can sustain in-context adaptation far beyond the training horizon.
cs.LG / 39 / 2610.25760
Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents
Abstract
Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent's actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task, background events announce which file's lock has been toggled but not its new value. The agent therefore has to reconstruct the current locks from its history before it can move files safely. With Qwen3-8B and Qwen3.5-9B, we separate two roles of reasoning: generating it before each action, and retaining earlier reasoning in the prompt. Most of the reduction in lock-violating commands comes from reasoning written at the moment of action. Retaining earlier reasoning adds little to rule-following, although it helps one model complete the task. The benefit depends on what the reasoning says. Filler text, reasoning written for another history and an accurate table of the state do not reproduce it. When we give the agent reasoning written for a history that differs in one lock value, the agent acts on that reasoning even against its own history. Without reasoning, linear probes recover the state from the model's activations about as well as a text classifier recovers it from the transcript. However, decoding accuracy does not predict which interventions change the agent's actions. In this setting, the benefit of reasoning for rule-following comes mainly from generating it at each step. The state that can be decoded without reasoning is a limited guide to how the agent acts.
cs.LG / 40 / 2610.25782
Detection Augmented Learning In Non-Stationary Bandit Environments
Abstract
We study the problem of non-stationary bandits without knowledge of the environmental changes. We propose Detection Augmented Learning (DAL) a black-box paradigm that comprises three modular components: a bandit algorithm, a change detector, and a forced exploration schedule. This modular design enables its applicability across a broad spectrum of bandit problems, from classical to highly structured settings. First, we provide theoretical insights into DAL's strong empirical performance: a single regret bound shows that DAL transfers the regret of its stationary input to the piecewise stationary setting, attaining or improving the best known bounds in every setting considered. We further identify conditions under which DAL performs well in drifting environments. Then, we showcase DAL's practical performance and versatility through extensive experimentation in which DAL consistently surpasses all state-of-the-art methods, both prior-free and prior-based, across diverse non-stationary scenarios in nine bandit settings. Finally, we investigate the effectiveness of the forced exploration and show that a bandit-driven strategy improves performance and shortens detection delay.
cs.LG / 41 / 2610.25806
Sector-Ring Linear Reservoirs: Sector-Wise Decay-Timescale Allocation With State Reuse Across Candidate Settings
Abstract
In reservoir computing, performance depends on whether the fixed dynamics of the reservoir suit the task, and the appropriate memory length and frequency response differ across tasks. The reservoir settings are found by a search on each task, and driving the reservoir with the input and collecting its states for every candidate setting dominates the cost of the search. In a reservoir with a linear recurrence, the dynamics decompose into modes, one per eigenvalue, and the eigenvalue distribution becomes a design variable. We propose the Sector-Ring Linear Reservoir (SRLR), in which the eigenvalue angles are fixed at evenly spaced positions around a circle and one decay timescale, chosen from a few values, is assigned to each sector of consecutive angles. To keep the scale of each mode's state equal across candidates, the input weights are scaled using the power spectrum of the training input. To reduce the cost of the search, the states of the modes for each decay-timescale value are computed once and reused across candidates. On prediction and reconstruction benchmarks built from chaotic systems, the SRLR achieved, in all eight cases, a lower mean test error than a reference model in which all modes share one decay timescale, and was competitive with nonlinear echo state network and leaky-integrator echo state network baselines. The selected decay-timescale profiles changed with the prediction target and horizon. The state reuse reduced the number of computed mode trajectories in the search by 99.8% relative to per-candidate recomputation.
cs.LG / 42 / 2610.25810
Communication-Efficient Federated Domain Generalization in Edge Networks via Dynamic Prototype Regulation
Abstract
Federated learning (FL) is a distributed paradigm for Internet of Things (IoT) edge networks. However, deploying Vision-Language Models (VLMs) over wireless edge networks presents distinct challenges: large model parameters increase uplink transmission latency over bandwidth-constrained fading channels, and non-independent and identically distributed (Non-IID) data exacerbates client drift. Existing feature alignment solutions often introduce auxiliary computations that consume time slots allocated for wireless parameter transmission, increasing the likelihood of straggler dropouts under system delay constraints. To address these issues, we propose DPGR-GCAM, a communication-efficient federated domain generalization framework. By employing a dual-parameter-efficient fine-tuning (PEFT) architecture, we design a channel-aware tiered transmission mechanism that adaptively truncates payloads to maximize node survival rates under fading conditions. To mitigate client drift, we introduce the Gradient-Conflict Aware Modulation (GCAM) mechanism. GCAM operates within the low-dimensional feature manifold to dynamically resolve multi-objective optimization conflicts with minimal edge computation latency, thereby reducing computation-induced network outages. Experiments demonstrate that DPGR-GCAM improves cross-domain generalization and sustains convergence under network fading scenarios with a 10\% client participation rate.
cs.LG / 43 / 2610.25812
Is it RL or Data? A Controlled Comparison of GRPO and Rejection-Sampling Fine-Tuning
Abstract
Bauer et al. (2026) showed that in low-data reinforcement learning with verifiable rewards, training GRPO on mixed-difficulty prompts gives up to 5x the sample efficiency of easy-only prompts. They did not test whether this benefit is specific to GRPO, or whether another post-training method given the same data would obtain it too. We test this on their procedural counting task with Qwen3-4B-Base and LoRA, comparing GRPO to rejection-sampling fine-tuning (RFT) on the same prompts, verifier and token cap. Six primary arms cross the two methods with three prompt sets (Easy, Mixed, and a Curated subset of prompts the base model sometimes solves, which makes GRPO's implicit filter explicit). We use three seeds per arm. Hypotheses and falsifiers were registered before training. To rule out a poorly tuned baseline, RFT receives a 27-run hyperparameter sweep, while GRPO uses Bauer et al.'s recipe unchanged. GRPO outperforms RFT at every prompt set: by 10.9 points on held-out test problems with curated prompts, and, in post hoc comparisons with matched rollout budgets, by 8.4 and 9.2 points on the easy and mixed prompts. On test the advantage survives the least favorable treatment of truncated completions; out of distribution, truncation prevents a conclusion. Curating prompts does not help RFT (-2.7 points). Mixed-difficulty data helps RFT (+3.7 points), while for GRPO the evidence is inconclusive. A secondary iterated-RFT arm, which resamples from the current policy, recovers 3.2 of the 10.9-point gap (about 30%). These are controlled measurements on one synthetic task family and one base model with three seeds per arm.
cs.LG / 44 / 2610.25847
ACE-Trader: Adversarial Co-Evolution for Generator-Critic Trading Systems
Abstract
Generator--critic trading systems pair an LLM-based trader that makes trading decisions with a red team agent that evaluates these decisions and provides feedback for revision. However, improving individual trading decisions does not ensure that both agents learn from experience and improve together over time. This raises an important question: how can the trader and red-Team select useful experience and learn from each other's evolving behavior? On one hand, experience that benefits decisions under one market condition may become ineffective or harmful under another. On the other hand, trading outcomes reflect both agents' actions, making it difficult to derive appropriate experience for each role. In this work, we propose ACE-Trader, an adversarial co-evolution trading framework. Specifically, we first develop separate context-dependent memory selectors that learn to identify useful experience for the current trading decision. Both selectors receive shared feedback from net-return differences between complete memory-assisted and memory-free trading pipelines, with online updates emphasizing recent outcomes. We further design a co-evolution mechanism that links role-specific memories through shared trading records. We use these records to evaluate the Red-Team's risk assessments and the effects of the Trader's revisions against realized market outcomes, then derive experience for their respective memories. By linking this experience to the corresponding records, we enable the Trader to improve its proposals based on prior critiques and the Red-Team to adjust its review focus as the Trader's behavior changes. Experiments on U.S. stock trading show that ACE-Trader outperforms all 22 active baselines in cumulative return and achieves the highest Sharpe ratio and lowest maximum drawdown across both active methods and passive benchmarks.
cs.LG / 45 / 2610.25850
Rubrics Beyond Evaluation: Exploiting Rubric Structure in Reinforcement Learning
Abstract
Rubrics are widely used to evaluate open-ended question answering by breaking overall quality into explicit criteria. These explicit criteria make rubrics a rich source of learning guidance, yet their potential for guiding the training process has been largely overlooked. Existing methods use rubrics mainly as direct feedback signals, without explicitly modeling the capabilities behind their criteria. This leaves their capability structure largely unused in guiding the training process. We introduce Rubric-Structured Reinforcement Learning (RSRL), an RL framework that unlocks the full potential of rubrics by exploiting their structure throughout the training process, rather than treating them simply as direct reward signals. At its core, RSRL builds on three key designs: (1) Rubric-to-Capability Structuring first recovers the capability structure hidden in rubric criteria, (2) Self-Sufficient Capability Learning then uses this structure to bootstrap underdeveloped capabilities into effective learning, (3) Capability-Residual Credit Routing finally preserves capability-specific gains during optimization. Experiments show that RSRL consistently improves both overall performance and underdeveloped capabilities, boosting downstream scientific question-answering performance by up to 7.43 over the initial policy and 7.60 over non-rubric structured RL baselines, showing that reinforcement learning can substantially benefit from the capability structure encoded in rubrics.
cs.LG / 46 / 2610.25856
Statistical Learning Theory for Direct Preference Optimization
Abstract
Direct Preference Optimization (DPO) learns from pairwise preferences through a logistic objective whose score-difference parameterization follows the Bradley--Terry model. We study how failure of this modeling assumption limits DPO, distinguishing structural misspecification from limited model capacity and finite-sample estimation. Our main contribution is a characterization of structural approximation error through the triangle-curl of the preference log-odds, which vanishes exactly when preferences admit a Bradley--Terry representation. We upper bound the achievable excess logistic risk by a score-class approximation term and a triangle-curl term, and establish a lower bound proportional to triangle-curl under additional boundedness assumptions. For ReLU score models, including clipped models, we derive explicit approximation rates showing that increasing capacity can eliminate the score-approximation term while leaving the structural obstruction unchanged. We further establish finite-sample bounds for clipped empirical risk minimizers that separate network approximation, structural misspecification, and estimation error. We place these results within a surrogate-risk framework by identifying contextual preference accuracy (CPA) as a natural measure of agreement between preference judgments and changes in pairwise probability ratios relative to the reference policy. Standard logistic consistency and calibration results connect DPO's objective to this policy property and transfer our excess-risk upper bounds to CPA, without treating CPA as a complete measure of generated-response quality.
cs.LG / 47 / 2610.25860
Adaptive Decoding Methods for Token-Efficient Reinforcement Learning
Abstract
Rollout decoding affects both which trajectories are used for reinforcement learning with verifiable rewards (RLVR) and how many completion tokens are spent collecting them. We study whether candidate-token support should remain fixed or widen during training in the context of rollout sampling efficiency. Our main schedule moves from top-K=10 to K=20 and then to full support as the attempted-token budget increases. In a 4B code-RLVR setting, the schedule has a higher interpolated 150M-token validation value than its matched fixed-K=10 continuation in all three runs, with a mean gain of $1.80{\pm}0.42$ points. The schedule also performs comparably to a separately trained fixed-K=20 baseline, the strongest fixed-support setting in our sweep. Follow-up experiments suggest that late widening to full support is the most promising variant, although the evidence remains noisy, and mixed-reward-group diagnostics do not clearly explain the gain. An exploratory single-run test replay yields identical point estimates for the schedule and fixed $K=10$. Overall, our results show that rollout support is an important control for sampling efficiency in RLVR and motivate further study of adaptive support schedules.
cs.LG / 48 / 2610.25873
Residual-Guided Adaptive Tucker Rank Allocation for Parameter-Efficient MCI Conversion Prediction
Abstract
Allocating limited trainable capacity across convolutional layers is a central challenge in parameter-efficient fine-tuning (PEFT) of pretrained three-dimensional (3D) networks. We propose Residual-Guided Adaptive Tucker Rank Allocation, which jointly selects input- and output-channel ranks under a fixed adapter-parameter budget while preserving the spatial kernel modes. To guide rank selection, we learn task-specific proxy updates from downstream training data and approximate them using candidate Tucker rank pairs. A perturbation analysis bounds the output discrepancy between the fixed proxy network and its Tucker approximation in terms of layerwise reconstruction residuals, motivating residual-based allocation. We normalize each squared residual by the corresponding proxy-update energy to compare relative approximation errors across layers. Dynamic programming minimizes their sum under a global budget constraint that accounts for both channel factors and the Tucker core. This yields a globally optimal allocation over the specified candidate rank pairs. We evaluate the method on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort for predicting conversion from mild cognitive impairment to Alzheimer's disease within 36 months, using hippocampal segmentation pretraining. Across five random seeds, the proposed method achieves a mean area under the receiver operating characteristic curve (AUC) of 0.7706, compared with 0.7239 for fixed-ratio Tucker adaptation and 0.7430 for full fine-tuning. The total number of trainable parameters is limited to approximately 0.618 million, accounting for only 1.45% of the full model. Adaptive Tucker rank allocation thus provides an effective approach to parameter-efficient transfer learning in 3D medical imaging.
cs.LG / 49 / 2610.25877
DeltaTrace: Efficient Signed Attribution for Reasoning Language Models
Abstract
We study how words in the context support or oppose a language model's response and present DeltaTrace, a novel attribution framework that explains the effect of input tokens on a fixed model response. DeltaTrace measures how the response's log-probability changes when the context tokens being explained are masked. To distribute this difference among source tokens, it constructs operator-specific local rules for finite changes and composes them backward through the computational graph using effective secant slopes. Each token is assigned a signed attribution as the inner product of its embedding displacement and the corresponding backward-propagated slope. Local conservation makes these contributions sum to the measured score difference and preserves their meaning when aggregated over sentences or passages. DeltaTrace supports both attention and linear-attention models, with auxiliary memory growing linearly with sequence length at fixed model dimensions and chunk size. On the evaluated attention model, DeltaTrace achieves the lowest deletion-faithfulness scores (RISE and MAS) on most retrieval and reasoning tasks and the highest needle-in-a-haystack Recall, exceeding FlashTrace by 5.67-24.33 percentage points. On the evaluated hybrid attention and linear-attention model, DeltaTrace lowers RISE on 9 tasks and MAS on 12 relative to FlashTrace.
cs.LG / 50 / 2610.25880
Asynchronous On-Policy Self-Distillation
Abstract
Unlike on-policy distillation (OPD), which depends on a separate teacher model, on-policy self-distillation (OPSD) leverages a single language model acting simultaneously as the student and the teacher, with the teacher conditioned on privileged information. Despite this architectural simplicity, synchronous rollout generation dominates overall training time, creating a severe systems bottleneck when scaling OPSD. To address this, we introduce Asynchronous On-Policy Self-Distillation (AsyncOPSD), a framework that decouples rollout generation from learner updates to improve training throughput. Since the teacher and student share parameters, AsyncOPSD eliminates the multi-model coordination overhead inherent to OPD architectures in asynchronous training. As the student policy evolves, buffered trajectories generated by older policies become stale, potentially degrading performance. To mitigate this, we introduce a distillation objective based on staleness-aware policy control and token-level entropy weighting. Across three model sizes and mathematical reasoning benchmarks, AsyncOPSD improves the training throughput by $1.96$--$2.25\times$ over synchronous OPSD while increasing average benchmark accuracy.
cs.LG / 51 / 2610.25881
Learn the Interface, Not the Implementation: Output Representation Dominates Training Regime in LLM-Based RTL Generation
Abstract
Large-language-model register-transfer-level (RTL) generation has concentrated on how to train the model: fine-tuning, continued pre-training, and reinforcement learning from a verifier. We argue this is the second-order question. On a bounded, formally specifiable task of generating single-error-correcting, double-error-detecting (SECDED) ECC memory wrappers, we hold the task, data, and correctness oracle fixed and vary two axes independently: the training regime (seven settings) and the output representation (four levels, from fully expanded flat Verilog to an eight-field configuration expanded by a deterministic generator). Because a reference generator exists, correctness is decided by formal equivalence over a design space of approximately ten million points, rather than sampled from a hand-written testbench. Two results follow. First, representation dominates: every representation close to raw RTL reaches a pass rate of at most 0.17, while the configuration reaches 0.83 with no training at all. Once the model emits the configuration, every training regime achieves a pass rate between 0.89 and 1.00. Second, raw RTL hits a hard output-length limit: a 512-bit design requires 44,237 tokens of flat Verilog compared with just 69 tokens for the configuration. Consequently, raw generation becomes infeasible above 32-bit data words rather than merely inaccurate. A model trained from scratch with fewer than 50 million parameters, combined with the verified generator, achieves 0.89 formal equivalence, compared with 1.00 for a 32-billion-parameter fine-tuned model, while using approximately 1600 times less inference energy per correct design. Inference energy also dominates amortized training energy. The lesson is to learn the interface, not the implementation.
cs.LG / 52 / 2610.25901
Out-of-Distribution Detection in Continual Learning
Abstract
Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown samples (OOD) without access to known (in-distribution (ID)) data from past tasks, where standard confidence-based OOD scores tend to degrade. Existing methods largely rely on a rehearsal buffer of past data or assume closed-world settings, limiting their scalability and reliability in real-world scenarios. In this paper, we propose a novel OOD detection framework for rehearsal-free continual learning, inspired by Kolmogorov-Arnold Networks (KAN), which mitigate catastrophic forgetting via their inherent local neuroplasticity. To better leverage KAN in OOD-CL, we introduce a task-adaptive classifier architecture, TA-KAN, which mitigates recency bias and enhances ID--OOD separability through task-specific control of activation locality. Furthermore, we propose a geometry-guided OOD score that complements TA-KAN classifier confidence without relying on past ID data. Our method significantly improves OOD detection while maintaining strong ID classification performance, achieving state-of-the-art results in OOD-CL.
cs.LG / 53 / 2610.25906
An Analysis of Semantic and Spatio-Temporal Co-evolution for Edge Classification on Dynamic Text-Attributed Graphs
Abstract
Dynamic edge classification on text-attributed Web graphs, such as review networks, community forums, and news-derived knowledge graphs, requires high quality, adaptive semantic representations in the context of network evolution. Existing dynamic graph architectures typically initialize text feature vectors with pretrained language models and hold them static during training to preserve their semantic structure, forgoing the potential benefits of allowing them to co-evolve with network topology. In this work, we investigate the effects of enabling such co-evolution with a unified text-graph alignment framework that integrates a lightweight Transformer adapter to update token-level text representations via spatio-temporal gradients from a contrastive pretraining objective. To prevent detrimental semantic distortion, a language model maintenance objective anchors the evolving text embeddings to their original semantic structure. Evaluations on the Dynamic Text-Attributed Graph Benchmark demonstrate that this anchored text-graph co-evolution consistently improves upon the edge classification performance of frozen-text baselines across diverse Web-related domains. Representational analyses reveal that the framework reduces anisotropy, improving the linear separability of text embeddings, and effectively encodes structural and temporal information into the semantic space. A computational efficiency analysis confirms that this beneficial semantic and spatio-temporal co-evolution is achieved without incurring the prohibitive computational overhead typical of end-to-end joint training.
cs.LG / 54 / 2610.25920
Maestro: Agentic Formal Proof Development by Coordination
Abstract
Lean is increasingly used to verify mathematical proofs and software produced by LLM agents, but writing formal proofs remains hard. Existing prover systems take a divide-and-conquer approach that recursively decomposes hard proof obligations into easier lemmas and splits again any lemma that resists proof. However, a lemma can resist proof for different reasons, including insufficient guidance, a false statement, or a flaw in the decomposition that produced it. Existing systems cannot tell these apart as a failed attempt reports little beyond its failure, and responses are confined to the failing node without considering the rest of the proof or revising the plan above it. The system may thus keep decomposing obligations along a flawed proof direction until the budget runs out, causing the whole proof attempt to fail. We introduce Maestro, a Lean prover harness that handles failures through global coordination. Provers report why they failed, not just that they failed, and a coordinator that sees the whole proof chooses an appropriate response, such as guiding a stuck prover, correcting a false lemma, or retracting a flawed subtree and replanning it. The same harness extends to verified code synthesis, where a checked counterexample sends the code back for repair. We evaluate Maestro with four frontier LLMs on seven mathematics and three software verification benchmarks. It proves more targets than each model's native coding agent (Claude Code or Codex) on every benchmark the agent has not already saturated, with the largest gains on research-level mathematics. With Claude Opus 5, it discovers its own proofs of 34 of the 48 theorems in the June 2026 release of arXivLean, against 19 for the native agent. It also outperforms all published prover systems on every benchmark they share, solving 669/672 PutnamBench, 59/60 Lean-IMO-Bench, and 6/6 IMO 2026.
cs.LG / 55 / 2610.25924
UtilityBench: Measuring Language Model Utility When Errors Have Different Costs
Abstract
As language models are increasingly deployed in real-world scenarios with important outcomes, the impact of their mistakes often extends far beyond what correctness can capture. In domains like medicine, a single error can lead to serious consequences and put patient safety at risk. However, mainstream evaluation protocols still treat all errors as equal, overlooking the spectrum of real-world costs. To address this, we propose a utility-based evaluation paradigm that explicitly incorporates the varying consequences of errors. Guided by this principle, we develop UtilityBench, a large-scale benchmark of 15,000 instances with verifiable answers, quantifying error severity using Best-Worst Scaling to elicit meaningful, data-driven cost estimates. These quantified utilities shape when a model should answer or abstain, guiding it to make decisions that align with the true importance of each scenario. As another merit, we introduce UtilityAUC, a metric that measures the area under the utility curve determined by a model's response ordering, thereby enabling utility preference alignment beyond error frequency. Our theoretical analysis demonstrates that UtilityAUC can distinguish between models where standard accuracy and AURC fall short. Empirically, evaluating 20 state-of-the-art models, we find that greater accuracy or more advanced architectures do not reliably translate to higher UtilityAUC, highlighting a gap between problem-solving ability and responsible decision-making in language model evaluation.
cs.LG / 56 / 2610.25933
The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction
Abstract
Offline compaction of the key--value cache (KV-cache) reduces the cost of repeatedly querying a fixed context by constructing a smaller cache once and reusing it subsequently. The central challenge of compaction is deciding which cache entries to delete and which to retain. Existing methods score each candidate for deletion on its own, with the rest of the cache being held fixed. However, softmax renormalization of the attention operation couples these decisions in a nonlinear manner: removing one entry changes the attention weights of every remaining one. In this work, we introduce KV-Surgeon, which uses calibration data to take interactions into account in a tractable way. We derive a quadratic surrogate that admits closed-form expressions for jointly making the optimal greedy removal decision as well as compensation within the remaining entries. On QuALITY, LongHealth, and Long Code Arena, and at every evaluated compaction ratio, KV-Surgeon substantially advances the cost--accuracy Pareto frontier of dense and mixture-of-experts models from 4B to 30B parameters. At $50\times$ compaction, KV-Surgeon improves question-answering accuracy by up to four points over attention matching (Zweiger et al., 2026), the state of the art, at comparable cost. Lastly, we introduce an even faster variant, which also exceeds the state of the art while requiring only two fifths of its cost.
cs.LG / 57 / 2610.25973
Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization
Abstract
Self-distillation policy optimization provides dense token-level supervision for LLM post-training through feedback-conditioned self-teachers. However, directly matching the self-teacher can disrupt successful behaviors on correct rollouts while providing insufficient correction on incorrect ones. We propose Asymmetric Policy Optimization via Self-Distillation (APSD), which constructs correctness-dependent targets by mixing student and self-teacher distributions: interpolation preserves successful behaviors on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. We analyze APSD through divergence decomposition, reverse-KL barycentric geometry, and gradient analysis, highlighting the importance of gradient flow through student-dependent targets. Experiments under reinforcement learning with verifiable rewards and rich feedback show that APSD achieves strong overall performance against reinforcement-learning and self-distillation baselines and matches SDPO performance with 2.6x fewer training steps on average across Qwen3 model scales.
cs.LG / 58 / 2610.25993
Graph-to-Table Interfaces Control Relational Generalisation
Abstract
Graph-to-table systems combine three design choices: an encoder that maps relational evidence to feature rows, a context constructor that selects labelled examples, and a tabular decoder. We investigate whether these components can be evaluated independently through matched comparisons that change one component while holding the rest of the pipeline and the evaluation constant. Across logistic regression, CATBOOST, and TABPFN-3 on CLUTRR, 57 GraphLog worlds, and 57 knowledge graphs, we show that conclusions about one component can change when the others change. On CLUTRR, absolute positional encodings maintain all distinctions among the evaluated queries and fit the short-path setting nearly perfectly; yet, they generalise worse than relation-count features because the supported prefix acts as a shortcut that breaks down on longer paths. On GraphLog, encoder preference depends on the decoder, while the relation-count representation degrades as CatBoost tree depth increases. Across different knowledge graphs, adding the frozen ULTRA score to local features is consistently useful, whereas adding local features to the score has a smaller and less stable effect. A pointwise TABPFN-3 readout using the full encoder underperforms native ULTRA on 48 of 57 datasets. These results show that encoder assessment depends jointly on context and decoder, while panel-level conclusions can also depend on evaluation and aggregation protocols.
cs.LG / 59 / 2610.26000
PnP-CF: A Consensus Flow-Matching Solver for General Inverse Problems
Abstract
Plug-and-Play (PnP) methods using continuous-time deep generative models as learned data distribution priors enable synthesizing high-quality solutions of model-based inverse problems. The quality of the solutions depends critically on the interplay and consensus between the inverse problem solver and the generative model sampler. We revisit a recently introduced flow-based PnP framework, which we refer to as PnP-CF (Consensus Flow), and extend its application to nonlinear inverse problems, including those with learned neural-network forward operators. PnP-CF combines a flow-matching denoiser with (i) ADMM-like dual variable updates and (ii) a time-adaptive pullback onto the intermediate flow probability path that combines a stochastic and a deterministic component. We discuss how these two features stabilize the algorithm and promote consensus, and demonstrate on a set of experiments that PnP-CF not only solves standard linear inverse problems with quality at least comparable to state-of-the-art baselines, but also extends gracefully to nonlinear inverse problems, including forward operators realized by pretrained neural networks, for which existing methods produce unsatisfactory results or fail completely.
cs.LG / 60 / 2610.26005
Scenario Generalization and Sector Transfer in Real Surface Irrigation Surrogate Modelling with Tabular Models
Abstract
Surface irrigation planning requires screening many operating conditions and spatial configurations, while high-fidelity shallow-water-equation (SWE) simulations are expensive. We study tabular surrogate models for predicting maximum water height from drone-derived field morphology, terrain, hydrological, spatial, and scenario descriptors, where each row represents a spatial cell under one irrigation scenario. To assess deployment feasibility, we evaluate gradient-boosted trees, neural tabular models, AutoML, and tabular foundation models (TFMs) under two regimes: same-sector generalization, where a model predicts unseen scenarios after observing a limited set of configurations in the same sector, and sector-out transfer, where knowledge learned from source sectors is transferred to an unseen sector. The main findings are twofold: a) under the evaluated fixed 8/32 same-sector split, surrogates achieve strong performance, especially with raw spatial coordinates; b) sector-out transfer is substantially harder because the target sector has different coordinates and geometry. Sector-frame normalization of spatial coordinates improves transfer across similarly structured sectors, but does not solve cross-shape generalization. Overall, tabular surrogates appear useful for same-sector scenario screening under this fixed split, while robust transfer to unseen sectors remains open.
cs.LG / 61 / 2610.26009
ProtGlycanDock: Towards Accurate Protein-Glycan Docking with Tailored Dataset, Benchmark and Model
Abstract
Performing accurate protein-glycan docking is critical to understanding protein-glycan interactions which govern many biological processes. However, this important task is hampered by the lack of tailored datasets, benchmarks and models for protein-glycan docking. To fill this blank, we first curate a standard benchmark ProtGlycanDock for training and evaluating protein-glycan docking models under both protein and glycan generalization settings. On this benchmark, we evaluate ten models in three categories, physics-based docking methods, machine-learning-based docking models and molecular structure prediction models. To obtain more precise binding poses, we further propose two finetuning techniques to inject the knowledge of glycan conformations into a general-purpose molecular structure prediction model. These two techniques respectively augment model inputs with glycan topological features and refine glycan binding pose with a glycan energy loss. Benchmark results show the superiority of AlphaFold 3 among existing methods, and the two proposed finetuning techniques can be well combined to achieve a state-of-the-art model for protein-glycan docking. Also, we perform stratified benchmark analysis and case studies to better understand the strengths and weaknesses of current models, providing insights for future model improvements. We maintain datasets at https://huggingface.co/datasets/ProtGlycanDock/ProtGlycanDock and source codes at https://github.com/kasawa1234/ProtGlycanDock
cs.LG / 62 / 2610.26030
Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation
Abstract
Looped transformers reuse a shared block of layers to increase computational depth without increasing the number of trainable parameters. They are trained with backpropagation through time (BPTT). To save training compute, BPTT is often truncated to the last few loops, although this may weaken the benefit of looping. To study this tradeoff systematically, we jointly vary model size, training tokens, and forward and backward loop counts across nearly 600 training runs and fit a unified scaling law that captures the effects of looping and truncation through effective parameters. Despite the training-compute savings of truncation, our fitted law favors full BPTT when forward and backward loop counts are jointly optimized. In matched-compute runs, full BPTT through fewer loops consistently outperforms truncated settings of the same compute. The law further shows that at fixed model size, the compute-optimal loop count and training-token budget both grow with compute. When model size is also free, the law identifies two fully trained loops as the best looped setting. Experimentally, this setting achieves lower loss than the compute-optimal non-looped model with about 27% fewer non-embedding parameters and only 8% more training compute.
cs.LG / 63 / 2610.26036
FinHopper: A Progressive Multi-Hop Benchmark for Complex Financial Agents with Step-wise Evaluation
Abstract
Financial agents are expected to answer factual investment queries through open search, but financial multi-hop retrieval differs from general multi-hop QA in a crucial way: each intermediate answer is an exact typed constraint, such as an entity, date, event, or numerical indicator, that determines the next query. This strict answer-conditioned dependency makes small intermediate errors cascade into plausible but wrong search trajectories. We introduce FinHopper, a benchmark for complex financial agents with controlled reasoning depth from 1 to 6 hops, covering stocks, funds, and macroeconomic facts, with gold intermediate answers for process-level evaluation. To construct FinHopper at scale, we design a state-transition-based synthesis framework that composes verified atomic financial QA patterns into dependency-preserving chains, followed by targeted in-synthesis verification, model-ensemble checking, and public-search validation by financial experts. Experiments with 20 LLM-agent configurations show that public search tools are necessary but insufficient: performance drops sharply as hop depth increases, the best model remains far below human experts, and process-aware scoring reveals partial reasoning progress hidden by final-answer-only evaluation. The benchmark is available at https://github.com/justajustin/FinHopper/.
cs.LG / 64 / 2610.26102
Provable Adversarial Detection: Prime Quantization meets Gromov--Wasserstein
Abstract
Adversarial vulnerability persists across modern vision architectures from CNNs to vision language models (VLMs), yet existing detection methods rely on heuristics without theoretical guarantees. We address the fundamental question of when adversarial perturbations can be provably detected from a geometric perspective. Our key insight is that adversarial perturbations cannot simultaneously preserve geometric structure across spaces with fundamentally different properties. Accordingly, we construct two such complementary metric spaces. First, we use a standard CNN embedding space $Z$, where adversarial samples exhibit significant displacement patterns. Second, we build a novel prime-quantized space $P$, that absorbs small perturbations through number-theoretic discretization, resulting in minimal displacement, while preserving discriminability. We then leverage the geometric discrepancies across spaces $Z$ and $P$ to detect adversarial samples. To the best of our knowledge, we establish the first rigorous separation theory for adversarial detection, proving that adversarial samples create unavoidable geometric inconsistencies across both spaces. Our framework provides theoretical guarantees including pixel-level absorption bounds, neighborhood diameter concentration, Gromov-Wasserstein (GW) separation theorems, and practical risk control. Extensive experiments validate our theoretical predictions and achieve consistently strong detection performance across a wide range of attack types and model families.
cs.LG / 65 / 2610.26103
Learning from a Bandit Learner
Abstract
Classical inverse bandit methods typically aim to recover true reward parameters. In contrast, we infer a demonstrator's subjective beliefs, specifically its internal estimates of suboptimality gaps. We consider no-regret demonstrators that follow randomized policies, such as Thompson sampling, and characterize a family of algorithms for which per-step inversion is possible. This enables gap estimates to be recovered directly from the arm-selection distribution. Our analysis identifies an inherent stability-coverage trade-off: inversion becomes ill-conditioned as the demonstrator concentrates its probability mass on a single arm, while the number of informative rounds depends on the rate at which its policy converges. We establish conditions under which inversion is reliable and provide theoretical guarantees on the resulting estimation error. Experiments on simulated data support the analysis and illustrate how these conditions affect gap recovery in practice.
cs.LG / 66 / 2610.26104
Declarative Explanations for Graph Neural Networks: A Demonstration
Abstract
Graph Neural Networks (GNNs) have proven effective for graph analytical tasks, yet their "black-box" nature complicates the understanding of their layer-wise behavior and hinders their application to large graphs. We demonstrate SliceGXQ, an end-to-end declarative and interactive system for querying and exploring layer-wise explanations for GNNs over large-scale graphs. In particular, SliceGXQ provides (1) an interactive graphical user interface for dataset, model, and explanation configuration; (2) a lightweight SPARQL-like parser that allows users to specify target nodes, layers, explanation modes, and constraints; and (3) an integrated execution and visualization workflow that invokes SliceGX as its underlying explanation engine. We demonstrate how users can diagnose a misclassified node, identify problematic layers and influential messages, and perform targeted model debugging, together with large-scale graph interpretation and fraud-detection analysis. Our demonstration video and source code are available at https://github.com/jinzhaoqet/SliceGXQ.
cs.LG / 67 / 2610.26114
From Retrieval to Recognition:Knowledge-Enhanced Post-training of Time Series Foundation Models for Classification
Abstract
Time series foundation models (TSFMs) have achieved notable progress in classification tasks through large-scale pretraining. However, their post-training practice mainly focuses on implicit parameter fine-tuning, where task-relevant information is entangled with abundant irrelevant knowledge in vast parameter spaces. This may hinder TSFMs from precisely activating useful knowledge for downstream classification. To this end, we propose Retrieval-to-Recognition (R2R), a plug-and-play post-training paradigm that adapts pretrained TSFMs to time series classification. Rather than pretraining a new foundation model, R2R improves how existing TSFMs are post-trained: it introduces a decoupled query encoder to extract discriminative patterns and retrieve instance-level numerical knowledge from a constructed knowledge base. The retrieved explicit knowledge is then fused with the TSFM's implicit parametric embeddings, enabling joint knowledge transfer at both the parameter and feature levels. In this way, R2R allows TSFMs to recall task-relevant patterns learned during pretraining, improving the efficiency of knowledge utilization in downstream classification transfer. Experiments on 158 time series datasets show that R2R consistently improves the classification performance of state-of-the-art TSFMs. Ablation and knowledge-enhanced studies further validate the effectiveness of the proposed framework.
cs.LG / 68 / 2610.26116
Variance-Aware Non-Stationary Bandits:All-Budget Regret Bounds and Adaptation Costs
Abstract
Low reward variance reduces estimation cost, but does not eliminate monitoring of arms that may improve while unobserved. We refine the variance--variation tradeoff and characterize how both costs depend on supplied variance and variation budgets in bounded, non-stationary multi-armed bandits. Let $A$ be the number of arms, $T$ the horizon, $B$ a bound on consecutive mean variation excluding the initial vector, and $Q$ a bound on the sum of the largest per-round reward variances. For every admissible pair $Q,B$, the minimax expected dynamic regret is, within a single logarithmic factor, \[ \min\{T,\;A+\sqrt{AQ}+(AQBT)^{1/3} +\sqrt{AT}\min\{B,\sqrt B\}\}. \] The deterministic boundary $Q=0$ is characterized without logarithms. For small positive budgets, a safe calibration phase identifies a common interval containing all future means of surviving arms. Clipping observations to a wider interval reduces the cost of restarting a prior hybrid static-regret algorithm; an original-mean analytical reference controls its residuals without charging the variation of clipped means. Balancing clipping bias and monitoring closes the small-budget noisy regime. Fixed oblivious constructions give the matching lower bound. We also quantify adaptation costs: on a disclosed deterministic one-improvement family, a stationary-regret budget has a matching worst-case change-regret frontier, including a finite improvement pulse. This implies an $Ω((T/A)^{1/4})$ loss when one policy competes with all supplied-budget deterministic oracle rates. The budget-free low-noise scheduling problem remains separate.
cs.LG / 69 / 2610.26123
Optimal Multi-Kernel Combination for Online Multi-Label Classification
Abstract
Multi-label classification is particularly challenging in streaming scenarios, where data arrives sequentially. Online kernel methods offer a promising solution. However, prevailing methods rely on a single preselected kernel, making them prone to suboptimal choices since the global data distribution is unknown in advance in online settings. Although the recent method, Online Kernel Selection for Online Multi-label Classification (OKS-OMC), enables data-driven kernel selection, it fails to fully harness the complementary information from multiple kernels and encounters the curse of kernelization. To address these limitations, this paper introduces a novel online multi-kernel combination method for large-scale multi-label classification. First, an efficient and robust online single-kernel multi-label method is designed. Second, learning an optimal convex combination of multiple multi-label kernel classifiers is formalized as a structural risk minimization problem. An improved alternating optimization procedure is developed, in which the combination coefficients are updated by directly minimizing the objective function, while each kernel classifier is updated by minimizing a surrogate upper bound. This procedure decouples the learning of kernel classifiers from combination coefficients, allowing independent classifier updates via our single-kernel method and accelerating convergence to high-quality solutions. Our multi-kernel method is proven to asymptotically converge to the optimal fixed convex combination of the best fixed single-kernel models chosen in hindsight. Ablation studies confirm that our method's components outperform their counterparts in OKS-OMC. Comprehensive evaluations across 11 datasets and seven metrics demonstrate our method's overall superiority over existing online multi-label methods.
cs.LG / 70 / 2610.26124
Nash-MOPD: Optimizing Task Trade-offs in Multi-Teacher On-Policy Distillation
Abstract
Multi-teacher on-policy distillation (MOPD) aggregates the capabilities of multiple domain-expert teachers into a single student model. However, all domains in MOPD share the student's finite budget of optimization steps and rollouts. Fixed uses constant domain sampling ratios based on corpus sizes; by contrast, D$^3$-MOPD adjusts the ratios using the teacher-student reverse-KL divergence and its recent decrease. However, favoring large recent decreases can direct more training toward domains that already improve quickly, while domains that improve more slowly may receive too few examples. To address inflexible allocation and the risk of giving insufficient training to domains that improve more slowly, we propose Nash-MOPD, an adaptive domain scheduler based on a local minimax game using existing reverse-KL statistics. The equilibrium maximizes the smallest predicted proportional reduction under the local model and gives more prompts to domains that need more examples for comparable improvement. Consequently, Nash-MOPD responds to changing learning progress and supports domains that improve more slowly without extra teacher queries or rollouts. At step 100, Nash-MOPD exceeds both baselines' final mathematics and code checkpoint scores on DeepSeek-R1-Distill-Qwen-1.5B and reaches Fixed's on Qwen3-8B-Base. Furthermore, it leads eight of nine reported metrics and is best or tied best on all six mathematics and code accuracy metrics, respectively.
cs.LG / 71 / 2610.26156
Dynamical Batching: How Class-Cycling Training Shapes the Internal Structure of Neural Networks
Abstract
We investigate whether cyclically varying the class composition of training batches can modify both the performance and the internal organization of neural networks. Motivated by nonequilibrium systems, in which cycling between related sub-goals can make switching between their solutions increasingly efficient, we introduce Dynamical Batching (DB), a sampling protocol that periodically emphasizes each class while leaving the loss and the optimizer unchanged, and compare it with Uniform Batching (UB), in which every mini-batch has equal class proportions. In the regime where standard training fails, DB changes the outcome: on a three-arm spiral, a capacity-limited MLP that plateaus far below perfect accuracy under UB fits the training data under DB. Tracking individual weights, we find that under DB the change between class-specialized states is carried by a smaller set of weights than under UB, and that this set is markedly more persistent from cycle to cycle, on every synthetic dataset we test and in a wider network that fits the data under both protocols. This organization is a signature of the protocol rather than a predictor of its benefit: it appears equally where DB leaves accuracy unchanged. On CIFAR-10, the same protocol modestly improves the test accuracy of a channel-narrowed ResNet-10, a standard residual architecture, and of a second capacity-limited convolutional network. These results show that how class emphasis is distributed in time can shape both the solution a network reaches and the internal structure of the network.
cs.LG / 72 / 2610.26199
TaQFormer: A Treatment-as-Query Transformer with Data-Fusion for Multidimensional Treatment Modeling in Coupon Allocation
Abstract
We study coupon allocation on an online marketplace offering multiple service categories. Issuing coupons requires choosing a vector of category-specific discounts, accounting for both their direct effects and cross-category demand shifts. We present TaQFormer, a treatment-as-query Transformer for this multidimensional treatment setting. Each treatment coordinate is represented as a query token, allowing cross-attention to estimate the response to the full treatment vector. To combine large but policy-biased observational logs with a smaller stream of randomized controlled trial data, we use a two-stage pipeline: observational pretraining learns general outcome patterns, followed by LoRA fine-tuning on RCT data with a request-specific fusion gate. We evaluate the resulting system through the checks required for production, including offline uplift metrics, treatment-parallel serving, latency monitoring, and a 14-day online A/B test. At essentially the same subsidy cost rate, \model increases business value per user by 1.26\% over the production baseline and significantly outperforms both the production baseline and the RCT-only model, while meeting the latency budget. It was subsequently rolled out to all marketplace traffic and has operated at full scale for several months as of October 2026. The deployment illustrates how multidimensional treatment modeling and observational pretraining, combined with adaptation using RCT data, can support effective joint discount decisions in production.
cs.LG / 73 / 2610.26209
OceanTensorCube and ST-MMT: A Multimodal Data and Transformer System for Seven-Day Forecasting of Pyropia Bleaching
Abstract
Bleaching of cultured Pyropia (laver) causes losses in Korean seaweed farming. We present a system that forecasts bleaching seven days ahead for farm clusters. OceanTensorCube aligns 32 channels from public marine, meteorological, ocean-model and satellite sources on a 30 m grid, and ST-MMT, a 9.2-million-parameter spatiotemporal transformer with hub-style multi-head cross-attention between variable groups, maps a 14-day input window to a seven-day-ahead bleaching probability. On the validation set, a six-seed ensemble detects every reported positive at a false-alarm rate of at most 3.5%, an upper bound because the report-based labels are positive-unlabeled. A same-date control on separately trained models indicates that the architecture discriminates between clusters when the calendar is held fixed. We also report the system's weaknesses: skill falls steeply as the quarterly nutrient survey ages (ROC-AUC 0.94 at 30 days, 0.53 at 129 days), and a single LSTM on cluster-average series outperforms a single ST-MMT, with only the ensemble leading at the operating point. We describe the labeling procedure, the evaluation protocol and the operational gate that withholds forecasts when inputs are stale. The forecasting service is operated at https://oceantensor.ai.kr.
cs.LG / 74 / 2610.26252
SEAM: Spectral Embedding Alignment and Merging for Low-Resource TTS
Abstract
Expanding the vocabulary of conditional text-to-speech (TTS) models typically requires extensive fine-tuning with paired audio, creating a major bottleneck for low-resource and cross-lingual adaptation. To address this, we introduce SEAM, a zero-audio initialization method that constructs embeddings for unseen tokens from known subtokens. Unlike standard pooling strategies that ignore sequential structure, SEAM applies a spectral filter to preserve temporal phonetic order and uses closed-form ridge regression to align the new embeddings with the pretrained conditioning manifold. Across three languages and two TTS architectures, SEAM approaches or matches baseline performance using up to four times less training data, reducing sample complexity by 75\% compared to random initialization and mean pooling. Our approach provides a lightweight, reproducible, and theoretically grounded solution for vocabulary expansion in low-resource speech synthesis.
cs.LG / 75 / 2610.26283
Trust Your Targets: Is Target Propagation Block-Diagonal Gauss-Newton?
Abstract
Target propagation (TP) is a backpropagation alternative in which local targets carry the global learning signal. Prior works claim that TP performs Gauss-Newton (GN) optimisation on these targets, but the correspondence has only been established under unrealistic assumptions where hitting them exactly is trivial. Practical variants instead descend on a least-squares problem via a single gradient step. In this paper, we rederive TP from a trust region under an implicit, previously unexamined norm, and introduce CGTP (Conjugate Gradient Target Propagation), which uses conjugate gradients to solve the layer-local subproblems in parallel. A single adaptive parameter jointly controls this trust region and a matching bound on the weight update; KFAC emerges exactly as the special case where the per-sample Jacobians are replaced by their batch mean, giving its factored damping a variational grounding. On MNIST and Fashion-MNIST MLPs, CGTP matches the accuracy of ADAM and KFAC while needing orders of magnitude fewer outer iterations than ADAM (about 10x fewer backward-pass-equivalents overall), while remaining stable across depth.
cs.LG / 76 / 2610.26285
Learning SOCP Solver Dynamics with Graph Neural Networks
Abstract
Second-order cone programs (SOCPs) are a fundamental class of convex optimization problems widely applied across engineering domains. Classical interior-point methods are robust and accurate, but their iterative nature can be costly in real-time or repeated-solve settings. This has motivated learning-based solvers that approximate optimizer solutions with neural networks. Recent work has increasingly studied graph neural networks (GNNs) for linear, quadratic, and second-order cone optimization, leveraging their ability to model sparse relational structure among variables, constraints, and objective functions. Most existing GNN-based solvers compress the optimization process into a single forward pass, predicting the optimizer directly without using the intermediate trajectory produced by classical solvers. We introduce SOCPNet, a recursive GNN framework that uses solver trajectories to learn an iterative SOCP solution procedure. A shared GNN model is applied repeatedly, refining the current iterate at each step and enabling inference for a variable number of iterations. Rather than predicting the solution in a single pass, SOCPNet learns from intermediate iterates to progressively refine its estimates using a trajectory-regularized training procedure. We evaluate SOCPNet on multiple datasets spanning standard SOCP problem classes and show that it consistently reduces the final solution error relative to a SOCP-GNN baseline. Finally, we demonstrate how trained SOCPNet models can be used in a hybrid optimization pipeline by warm-starting a classical interior-point solver, reducing solver iterations without loss of accuracy.
cs.LG / 77 / 2610.26286
Toward Online Robust Zero-Sum Markov Games with Function Approximation
Abstract
Online robust zero-sum Markov games pose a coupled operator-learning and game-solving problem: from nominal interaction data, the learner must recover a worst-case minimax Bellman operator while maintaining strategic coverage and controlling stage-game approximation error. Existing robust reinforcement-learning theory does not yet provide such an operator-learning perspective beyond specialized multi-agent settings. We develop an optimistic dual robust Bellman framework that combines dual robust Bellman fitting, optimistic minimax planning, backward-validated confidence sets, and exploiter-mix data collection. Our analysis is organized around the robust Bellman residuals induced by the candidate value class and yields a regret reduction that separates exploration complexity, robust operator-estimation error, stage-game solution error, and misspecification. Under explicit interpretable assumptions, we obtain conditional theorem interfaces for general and projected function approximation, together with closed results in several structured regimes, including tabular, epochwise cross-fit linear, self-normalized pure-online linear / finite-rank kernel, and projected RKHS truncation. Small-scale numerical experiments illustrate robustness calibration, strategic coverage, and the return-gap trade-off predicted by the framework.
cs.LG / 78 / 2610.26325
Typed Decision Models under Prevalence Shift: Decision Training, Calibration, and Implicit Priors
Abstract
Typed decision models (TDMs) return a probability over caller-defined options, and Jev is described by its vendor as calibrated. But in deployment class frequencies change, so calibration measured on a fixed test set need not hold where thresholds are applied. Existing evaluations leave two gaps: the gains of vendor decision training have not been separated from those of the backbone under an identical readout, and TDMs have been tested only at the test set's class distribution. We ask whether decision training improves a model over its own backbone, and whether a TDM's calibration and thresholds withstand prevalence (label) shift as well as a fine-tuned classifier's. We compare three TDMs (Jev, Tev1-4B, Nimble-9B), their base models and two trained classifiers on identical decisions from five public datasets under simulated shift, with 35 paired tests fixed before any test was computed. First, decision training is not uniformly beneficial: Nimble-9B, which differs from Qwen3.5-9B only by a LoRA adapter, is better calibrated on intent tasks but 12.0 Macro-F1 points worse on financial sentiment. Second, under moderate shift Jev's expected calibration error (ECE) and threshold risk grow significantly more than a fine-tuned DeBERTa's only on financial sentiment, by 1.7 and 2.5 percentage points; averaged over 2,000 shift scenarios (exploratory), the gap is at most one point on every dataset. Third, EM prior correction with the model's mean development prediction as its prior needs no labels and lowers Jev's ECE under shift by 2.7 and 1.8 points on the two datasets where it is highest, although labelled recalibration lowers it far more. The three TDMs we tested are neither uniformly better than their backbones nor uniquely fragile under shift; for calibration, their main weakness is the miscalibration they already have without shift. We release all inputs, code and outputs.
cs.LG / 79 / 2610.26337
CIS-Transolver: A Cross-Layer Incremental-State Neural Operator for Fine-Scale Automotive Aerodynamic Flow Prediction
Abstract
Neural operators offer an efficient data-driven approach to rapid prediction of complex automotive aerodynamic flows. However, spectral bias and other factors limit accurate reconstruction of fine-scale flow variations near complex geometries. Existing studies improve the representation and fitting of fine-scale information mainly through information extraction and training objectives. Further study is needed to determine how extracted local details can be used continually in subsequent operator computations and contribute effectively to the overall flow-field representation. We propose CIS-Transolver, a cross-layer incremental-state neural operator that extends fine-scale modeling from local information extraction to its use across layers. The method extracts fine-scale information through local multiscale analysis and encodes it as geometric moment states that are propagated and updated across layers. By combining information from previous layers with feedback to the value representations in Physics-Attention, the method allows local details to contribute continually to subsequent modeling of global physical correlations. We systematically evaluate CIS-Transolver on three public automotive aerodynamics benchmarks, ShapeNet-Car, AhmedML, and DrivAerNet++, covering surface pressure fields and volumetric pressure and velocity fields. CIS-Transolver achieves the lowest mean relative L2 error among the evaluated methods for all five prediction metrics across the three benchmarks. The relative error reductions average 8.72% across the five metrics compared with the best baseline for each metric, reaching 16.38% for ShapeNet-Car pressure prediction and 15.60% for AhmedML velocity prediction. These results provide a new approach to fine-scale flow prediction for complex geometries.
cs.LG / 80 / 2610.26347
Logit Averaging During Training Sets a New State of the Art for Non-Invasive Speech BCI
Abstract
Non-invasive brain-computer interfaces are in practice limited by signal-to-noise ratio, and recent state-of-the-art methods address this limitation by aggregating over repeated trials of the same stimulus. Yet multi-trial aggregation has been treated inconsistently across the literature: some work averages brain signals at the model's input, other work averages predictions at the model's output, and neither treats aggregation as an architectural choice in its own right. We provide the first systematic comparison of aggregation strategies for non-invasive neural decoding, evaluating six paradigms (including logit averaging during training and three attention-based variants we introduce) across three datasets, four tasks, and two group sizes. We find that aggregation choice matters substantially: training-time logit averaging, the simplest method we propose, outperforms all others in 6 of 7 settings. On the phoneme classification setting of the 2025 PNPL competition, it sets a new state of the art, significantly outperforming the competition winner by 17.4 absolute percentage points in balanced accuracy. We further draw a theoretical connection between logit averaging and joint ensemble training, which has been shown to underperform independent ensembling in the deep learning literature; our results call into question recent explanations of why joint ensemble training fails. Together, these findings establish multi-trial aggregation as a key design choice for non-invasive neural decoding and identify training-time logit averaging as a simple, effective default.
cs.LG / 81 / 2610.26370
An Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs
Abstract
Parameter-efficient fine-tuning (PEFT) methods are critical for adapting large models under resource-constrained settings. However, as mixture-of-experts (MoE) architectures, which offer efficient inference at massive parameter scales, gain increasing prevalence in the community, a critical gap arising from the lack of systematic recipes for applying PEFT to MoE models. To narrow this gap, we conduct a large-scale systematic study that investigates a set of essential design recipes, including which modules should be adapted, which granularity should be used to adapt MoE modules, which PEFT method to use, and to what extent should the load-balancing objective be used, across three MoE backbones on downstream tasks and general knowledge tasks. Our experiments yield several key insights. First, adapting MoE experts alone achieves the highest downstream performance with competitive general capability retention under a comparable parameter budget. Second, native expert-weight updates offer a better downstream--retention trade-off than adapters that bypass experts or entire MoE modules. Third, LoRA and DoRA together define the strongest operating region for high downstream accuracy among the evaluated PEFT methods. Fourth, no nonzero load-balancing coefficient Pareto-dominates the zero baseline. We finally evaluate a weight-level redesign of MoE$^2$-LoRA on Qwen3. Moving its module-level bypass pool into native expert weights improves all four upstream endpoints by 2.4 to 17.9 points with 17% fewer trainable parameters and a 3.2-point average downstream cost.
cs.LG / 82 / 2610.26379
CTLite: Lightweight Irregular Multivariate Time Series Forecasting with Reusable Continuous-Time Summaries
Abstract
Irregular multivariate time series forecasting (IMTSF) requires capturing temporal dynamics and cross-variable dependencies from sparse, asynchronous observations. Existing methods often rely on learned temporal encoders, coupling historical representation construction with model optimisation and repeatedly processing the same observations. We propose CTLite, a lightweight forecasting framework that decouples historical summarisation from learned prediction. Specifically, a parameter-free continuous-time moment encoder summarises local levels and trends at multiple temporal scales while suppressing poorly supported slope estimates. A compact readout combines damped local extrapolation with a low-rank cross-variable correction that accounts for observation freshness, target availability, and forecast horizon. Since the historical summaries are independent of trainable parameters and forecast queries, they can be reused across optimisation steps and repeated predictions for an unchanged history without altering predictions or parameter gradients. Experiments on four real-world benchmarks show that CTLite achieves the best or second-best mean MSE and MAE among the evaluated methods with only 88-2,537 trainable parameters. Under matched GPU workloads at batch size 32, CTLite reduces inference latency by 18.6-87.3% compared with the latest models, including the cost of historical encoding. Reusing summaries further reduces total runtime by approximately 52% over 16 forecast calls on the same history compared with recomputing the encoder for every call, including the initial encoding cost. The code is at: https://github.com/SSMa913/NICLab-CTLite.
cs.LG / 83 / 2610.26407
A Simple and General Approach to top-$k$ Selection
Abstract
This paper considers the Bayesian top-$k$ identification problem, wherein a decision maker is tasked with selecting $k$ items from a pool of items of unknown quality. To assess the items, the decision maker adaptively allocates a fixed sampling budget, where each sample of an item yields a noisy observation of its quality, with the objective of minimizing the expected opportunity cost of the selected set. We propose an approach that allocates each batch of samples by minimizing a surrogate objective derived from an upper bound on the expected opportunity cost. The approach is general: samples can be allocated in batches, the selection size $k$ is arbitrary, and it applies to many sampling distributions. It is also simple and intuitive. We prove that the method's posterior expected opportunity cost decays at an exponential rate of at least half the best rate achievable by any sampling rule. In simulations across a wide range of settings, the method is competitive with, and often improves upon, the best existing algorithm in each setting.
cs.LG / 84 / 2610.26470
On Learning Expand-and-Sparsify Representations
Abstract
Expand-and-sparsify ($\texttt{EaS}$) representations lift an input to a much higher dimension through a linear map and keep only the $k$ largest coordinates. Existing theory builds that map at random and needs the expanded dimension to grow exponentially in the input dimension, which puts the construction out of practical reach. We ask what is gained by learning the map by gradient descent instead, and how that gain depends on how wide the expansion is. We sweep $73$ OpenML classification tasks and $48$ regression tasks under a protocol that holds the datasets, expanded dimensions, sparsity levels and splits fixed across every method, and we summarise each comparison by a median over datasets and a Wilcoxon signed rank test. We report four findings. First, learning the expansion helps most when the expansion is narrowest, and its advantage over a random map decays to nothing by the widest expansions we run, so a learned $\texttt{EaS}$ layer can be made small. Second, with the map learned, a top-$k$ representation that keeps at most a tenth of its coordinates is indistinguishable from a dense $\texttt{ReLU}$ layer of the same expanded dimension at every expansion factor we test, while GELU and SiLU are not. Third, when the representation is forced to be binary the sparse code is the better of the two, and the reason is that binarization costs it nothing while it costs the dense layer accuracy. Fourth, how sparse the representation can be is governed by how wide it is: the sparsity level that is free rises as the expansion grows, so there is no fixed limit on how sparse it can be. The regression tasks reproduce the first finding and qualify the second and third.
cs.LG / 85 / 2610.26473
A Simple and Unified Method for Gradient-Variation Regret
Abstract
Gradient-variation online learning has attracted increasing attention for its connections to game theory and optimization. It has been studied in both non-stationary online learning, which aims to minimize dynamic regret against time-varying comparator sequences, and universal online learning, which seeks to adapt to unknown loss curvature. Prior work obtained gradient-variation bounds using different techniques tailored to each setting. We propose a simple and unified method based on surrogate losses with a quadratic penalty on decision movement, yielding gradient-variation bounds in both settings. For non-stationary online learning, the method induces an optimistic-to-gradient-variation reduction, enabling a broad class of algorithms to match the rate established by Zhao et al. (2024). For universal online learning, it yields a simple algorithm that matches the optimal universal regret bounds of Yan et al. (2024) while relaxing the extended smoothness assumption. Finally, we apply our method to online games and improve on existing guarantees.
cs.LG / 86 / 2610.26489
Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes
Abstract
Changing sparse autoencoder (SAE) codes can alter both intervention direction and magnitude, so comparing intervention effects requires a strength control specifying what is held equal. Norm matching and output calibration provide different controls, but whether interventions yield different effects across codes after strength matching remains unclear. We compare intervention effects under these controls for fixed feature groups in Gemma and Llama and observe: (1) Whole-patch scaling. With one coefficient per input, full-output Kullback-Leibler (KL) calibration leaves smaller mean effect gaps from the encoder-based reference intervention than total-norm matching for nonnegative least-squares (NNLS) refits. On new Llama inputs, total-norm matching loses 11 reference transfers, whereas full-output calibration preserves every transfer decision. (2) Positionwise scaling. On a small matched gender-agreement subset, positionwise norm matching leaves smaller observed mean gaps than full-output calibration. (3) Answer exclusion. Compared with unscaled NNLS, answer-excluded calibration reduces gaps in the measured subject-verb agreement comparisons but increases them for gender refits. Full-output KL calibration also reduces larger gaps for norm-matched constructed candidates, although matching coverage is incomplete in several settings. Since calibration uses the evaluated input's outputs, it serves as a diagnostic of residual disagreement.
cs.LG / 87 / 2610.26525
A Fisher-Geometric View of Multi-Teacher On-Policy Distillation
Abstract
Consolidating specialized foundation models into a unified generalist is a central goal of post-training. Two approaches are in common use: model merging, which is fast but treated as a heuristic, and multi-teacher on-policy distillation, which matches distributions but demands heavy rollout compute and suffers from cross-task gradient conflict. We show that the two are one objective seen at different orders. Within a shared basin of homologous specialists, the multi-teacher on-policy distillation objective is, to second order, exactly the Fisher-merging objective. Its minimizer is therefore the Fisher merge, a classic model-merging method given by a closed-form combination of the specialists, so merging is optimal, to second order, as a closed-form stage of distillation. Building on this identification, we propose Residual-Governed Branch-and-Merge(RGBM), a two-phase consolidation procedure. Concretely, a full-parameter Fisher merge is intractable on foundation models, so Phase~1 restricts the merge to the task-vector subspace. Extending tangent-space analyses of merging, we show that the same coefficients that mix task vectors also mix implicit rewards, which lets a residual program on unlabelled prompts solve for them. The program minimizes the worst per-domain residual and reports its energy-normalized values as the domain residual spectrum. Phase~2 lets this spectrum govern what may be added: RGBM passivates domains with low normalized residuals, trains isolated branches only for domains above the spectral gap, and writes back the admitted updates with calibrated shrinkage. On two open-source consolidation suites, the Phase~1 merge alone matches or outperforms fully trained multi-teacher distillation, and RGBM improves on it further with zero extra joint-training steps, whereas every un-gated write-back, including continued joint distillation from the merge, falls below it.
cs.LG / 88 / 2610.26595
Concerns in Benchmarking Jailbreaks: An Empirical Study of Attack Success Rates
Abstract
Jailbreak methods are commonly compared using Attack Success Rate (ASR), but these comparisons depend on how responses are generated and judged. In this paper, we examine eight attacks against three open-weight language models using three evaluation procedures. Low-resource language attacks achieve a mean ASR of 96.0% under keyword matching but 2.7% under a Qwen3 evaluator following StrongREJECT scoring instructions. Changing the number and arrangement of attempts reverses the ranking of DeepInception and rewriting. In a separate experiment, sampling 100 responses for each unchanged request at Ollama's default temperature of 0.8 yields 100% ASR under keyword matching but 50% under the Qwen3 evaluator. Beyond evaluator sensitivity, these findings raise a safety concern: repeated sampling can expose residual unsafe behavior despite an initial refusal, suggesting probabilistic suppression rather than robust exclusion. We recommend explicit success definitions, matched response budgets, validated evaluators, and reporting cumulative failure under repeated interaction rather than relying on single-generation tests.
cs.LG / 89 / 2610.26597
Bombadil: A Foundation Model for Plant Electrophysiology
Abstract
Living plants, like animals, use electrical signals to carry information about their physiological state across their body, from roots to branches. These signals can be recorded, as they are in animals, from a pair of differential extracellular electrodes inserted into signal-conducting tissue. While machine-learning models have been trained to predict specific plant stress states from these signals, each of these models has been trained for only a single task and none have been designed to transfer across tasks, species, or environments. These approaches do not allow easy scaling across the wide range of species and environments that are relevant to agriculture, forestry, and horticulture for both commercial and consumer plant care. Here we present Bombadil, a foundation model for plant electrophysiology, pretrained without labels on over 4 million hours of electrical signals recorded from 100 plant species in ten distinct environments, including laboratories, greenhouses, commercial orchards, and natural forests. In this way, Bombadil is able to learn the fundamental features of plant electrical signals across a wide array of species and environments. Then, from those same electrical signals, Bombadil tracks changes in seven physiological indicators of plant stress, even among species and stressors with no labels in training.
cs.LG / 90 / 2610.26604
ITU-DESC: a missingness-aware machine-learning dementia risk score in the UK Biobank cohort
Abstract
Background: Existing dementia risk scores rely on fixed, pre-selected features and linear models, or ignore bias from informative missingness in cohort data. Here we introduce the Istanbul Technical University Dementia Risk Score (ITU-DESC), an explainable machine-learning score for incident all-cause dementia that combines broad feature selection across roughly 8,000 candidate features with explicit control of missingness bias. Results: We studied approximately 500,000 UK Biobank participants with a median follow-up of 16.6 years. We withheld missingness patterns from the model using unconditional random hot-deck imputation. We stabilized feature selection by repeating BorutaSHAP-based feature selection with ten random-imputation seeds and selecting predictors chosen in at least eight of the ten seeds, which yielded 38 predictors. We trained a CatBoost gradient-boosted decision-tree model and evaluated it on a holdout test set of 50,041 participants. For all-cause incident dementia, ITU-DESC achieved an average precision of 0.138 (95% confidence interval 0.123-0.156) and an area under the receiver operating characteristic curve (ROC AUC) of 0.853 (0.844-0.863), significantly outperforming other risk scores, including the strongest comparator, the UK Biobank Dementia Risk Prediction model (0.118 and 0.841). We identified dementia risk predictors absent from established scores, including leukocyte telomere length and liver enzymes. Conclusions: ITU-DESC is the first dementia risk score to combine broad feature selection with explicit control of missingness bias.
cs.LG / 91 / 2610.26612
Degree Bounds for Separating Polynomial and Neural Invariants for Point Clouds
Abstract
Neural networks that respect the symmetries of point clouds commonly compute high-order representations of the rotation group to improve their expressivity. Such models are known to be universal, and thus able to separate any pair of point clouds, when using representations of sufficiently high order, yet no precise bound on this order, as a function of the number of points, was previously known. We establish the first quadratic lower and upper degree bounds for polynomials that separate point clouds of n points up to rotations and permutations. By showing Tensor Field Networks and MACE can express all invariant polynomials of a given degree with sufficient representation order, we additionally derive the first explicit, quadratic bound on the representation order sufficient for these architectures to be complete, replacing previous existence results with concrete guarantees. We further show that, under local regularity assumptions, point cloud separation can be achieved with a message-passing-based model using invariants whose degree depends only on the size of a local neighborhood, rather than on the size of the full point cloud. Finally, we provide counterexamples showing MACE must compute high order irreducible representations to separate locally regular point clouds.
cs.LG / 92 / 2610.26618
Global Time-Aware Conditioning (GTC): A Mechanism for Time-Aware Pediatric Sleep Representation Learning
Abstract
Sleep physiology evolves throughout the night, with changes in sleep cycles, respiratory stability, and cardiovascular dynamics. However, self-supervised learning (SSL) methods for sleep typically process short, fixed-duration windows without explicitly modeling where each window occurs within the full sleep session. We introduce Global Time-Aware Conditioning (GTC), a mechanism that complements local positional encodings by conditioning latent representations on the global time-of-sleep of each input window. We integrate GTC into three representative SSL frameworks and evaluate them on five clinically relevant tasks with extreme class imbalance across two independent pediatric cohorts. GTC generally improves F1 and AUPRC across architectures and tasks, with gains persisting under cross-cohort transfer, highlighting the value of global temporal context for sleep representation learning.
cs.LG / 93 / 2610.26628
SleepPairNet: Learning Multimodal Pediatric Sleep Representations with Random-Pair Cross-Attention
Abstract
Analyzing physiological sleep signals is important for assessing sleep quality and detecting sleep-related disorders. Learning inter-modal relationships across physiological signals can provide richer representations for downstream sleep analysis. We introduce \textbf{SleepPairNet}, a multimodal self-supervised learning (SSL) framework that learns pairwise interactions between physiological signals through cross-attention. We propose a random modality-pair selection strategy that exposes the model to diverse signal pairs during pretraining, enabling flexibility across modality combinations at test time. Pretrained on the NCH dataset, SleepPairNet outperforms the evaluated SSL baselines on all respiratory tasks while remaining competitive on EEG-related tasks, and shows even stronger results when transferred to the external CHAT dataset.
cs.LG / 94 / 2610.26629
VaDar: Implicit Goal Conditioning via Value Disaggregation
Abstract
Long-horizon tasks are important for modern AI and robotics, yet pose a challenge for reinforcement learning. Historically, a popular approach has been to decompose a task into subgoals, and learn a subgoal-conditioned policy to execute each subtask in sequence. We introduce an alternative called Value Disaggregation (VaDar), which trains a single subgoal-independent policy to maximize a linear combination of per-subgoal critic values. Because subgoal information is needed only at training-time, VaDar removes the need for possibly costly or cumbersome subgoal-selection during policy deployment. Moreover, through careful experiments, we show that VaDar (1) is always competitive with, and frequently outperforms, natural baselines like goal-conditioning and (2) succeeds on tasks with non-sequential goal structure and multiple optimization objectives. In particular, VaDar is the first algorithm to solve challenging, long-horizon tasks in the RoboCasa suite, even when starting from a policy pretrained from behavior cloning with near-zero initial task success. Taken together, VaDar's successes suggest that subgoal information benefits actor-critic reinforcement learning by enhancing the efficacy of critic learning, whereas policy goal-conditioning is often unnecessary.
cs.LG / 95 / 2610.26651
The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models
Abstract
Pretraining filtering is one of the most common safeguards for open-weight language models. Here we study its effectiveness, durability and scalability using the ESM-2 family of protein language models. We find that filtering viral proteins from the pretraining data meaningfully degrades masked language modelling, mutation effect prediction and long-range contact prediction on viral proteins, while largely preserving performance on non-viral proteins. These effects are concentrated among proteins from human-infecting viruses, and grow as we scale models from 8M to 3B parameters. As models increase in size, we find that stringent filtering becomes more effective. However, fine-tuning on sequences from the removed viral families restores the lost performance. Although the cost to recover performance increases with model scale, even for the largest models it costs about 80 USD of H100 time. This work establishes pretraining filtering as moderately effective at selectively degrading performance on viral proteins, but not durable against fine-tuning with modest computational resources and access to relevant data.
cs.LG / 96 / 2610.26666
Nested Shared-Operator Recursion on ARC-AGI-1: An Empirical Study of HRM, TRM and URM
Abstract
Recursive reasoning models refine persistent internal states by repeatedly applying learned operators. We investigate whether adding a persistent intermediate state that integrates repeated updates of a faster state through the same operator improves predictive performance relative to each model family's baseline at matched parameter counts. We define this organisation operationally and implement it within three architectures: the Hierarchical Reasoning Model (HRM), Tiny Recursive Model (TRM) and Universal Reasoning Model (URM). Using the official codebases on ARC-AGI-1, we compare each baseline with its parameter-matched nested variant. We evaluate held-out queries from tasks whose demonstrations were available during training. The experiments use a single-GPU budget far below the published training protocols. Under the tested conditions, nested variants show no consistent advantage in token accuracy, predictive loss or exact solving, which remains near zero for all configurations. The results therefore leave the hypothesis open. We specify a confirmatory comparison, with an informative baseline, multiple seeds, controls that isolate nesting and predefined criteria, to determine whether this organisation merits further development.
cs.LG / 97 / 2610.26671
PACT: Plan-Augmented Coordination via Trace-Graphs for Dynamic Multi-Agent Combinatorial Optimization
Abstract
Multi-agent approaches have been investigated recently as a means of improving the scalability of reinforcement learning for neural combinatorial optimization. Existing methods assume that all problem instance information is known upfront. However, real-world problems are often dynamic: relevant information is progressively revealed as the agents take actions (e.g., arriving at an edge in a routing problem only to discover it is untraversable). In such cases, existing methods fail to coordinate agents to explore and exploit the progressively revealed information. To address the arising coordination challenge, we introduce PACT (Plan-Augmented Coordination via Trace-graphs). PACT augments the environment graph with auxiliary nodes representing the plans of agents, which are connected to the resource nodes to indicate future dependencies. The proposed message-passing scheme over the augmented graph grounds inter-agent coordination in the environment and enables proactive, rather than reactive, conflict resolution. We show that PACT improves solution quality in dynamic problems relative to baselines and exhibits the best transfer to larger instances on which, remarkably, it outperforms centralized oracles with privileged access to instance uncertainty.
cs.LG / 98 / 2610.26676
Which Model for Which Role? Cost-Efficient Pareto Recommendation for LLM Agentic Workflows
Abstract
LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple candidates. We introduce Cost-Coupled Gittins (CC-Gittins), a multi-objective extension of single-objective cost-aware Gittins search for Pareto recommendation. It exploits the coupling between evaluation and deployment costs: assignments that cost more to deploy also cost more to evaluate. The search alternates between seeking higher quality and lower deployment cost, reusing each evaluation to inform both searches. After each batch, CC-Gittins recommends assignments that are nondominated under uncertainty-adjusted estimates of quality and deployment cost. We evaluate CC-Gittins on eight quality--cost matrices from workflows with two to five roles and up to 1,681 candidate assignments. We also evaluate the method with latency as a third objective. Across the eight benchmarks, CC-Gittins recovers a substantial fraction of the true Pareto frontier at roughly 10% of the exhaustive-evaluation cost, and recovers most of the frontier with few false positives by 30%.
cs.LG / 99 / 2610.26685
Intra-Rubric Reward Divergence: Mitigating Rubric Hacking through Latent Dimension Recovery and Balancing
Abstract
Rubric hacking limits both performance and training robustness in rubric-based reinforcement learning. By decomposing rubrics into sub-dimensions, we reveal the Intra-Rubric Reward Divergence: as the total rubric score increases, certain sub-dimension scores systematically decrease, suggesting the model sacrifices specific quality aspects to optimize the aggregated objective. From the sub-dimension perspective, we observe two phenomena: (1) dimension imbalance, causing the policy to favor heavily covered dimensions; (2) dimension absence, leading to a systematic gap between rewards and true quality. To mitigate these issues, we propose Fair-Race Dimension Aggregation (FRDA), which has generations compete within each dimension separately to mitigate rubric imbalance, and Residual Semantic Dimension Completion (RSDC), which mines evaluation dimensions implicitly encoded in a holistic preference signal to complete rubric coverage. Experiments in chat, science, and medical domains indicate that FRDA mitigates rubric hacking with more stable improvements, while RSDC outperforms both single-reward training and naive multi-reward aggregation, yielding generally higher-quality responses.
cs.LG / 100 / 2610.26707
GeoSplit: Selective Edge--Server Map Acquisition for Cross-User Travel-Mode Recognition
Abstract
Networked mobility sensing couples local trajectory inference with remote geospatial context, making communication decisions part of the prediction system. We present GeoSplit, an edge--server framework for travel-mode recognition that separates the utility of map evidence, its training-time transfer, and its selective acquisition at runtime. On GeoLife with 10% labels, we evaluate these components under user-day and user-disjoint grouping. For unseen users, aligned full-map context raises macro-F1 from 0.7154 to 0.7864, while map-informed distillation yields a motion-only student at 0.7225. Through a compact query interface, entropy gating achieves 0.7729 macro-F1 versus 0.7738 for all queries, querying 39.5% of trajectories and reducing application payload by 60.5%. Under user-day grouping, selective and all-query scores are similarly close at a 21.6% query rate. Learned value estimation and probability fusion offer no consistent improvement. These results illustrate communication-efficient edge--server inference for an intelligent transportation sensing task.
cs.LG / 101 / 2610.26710
Doubly Robust Learning of Heterogeneous Survival Treatment Effects under Time-Varying Treatments
Abstract
Estimating heterogeneous survival treatment effects (HSTEs) from time-to-event data under time-varying treatments is important in fields such as medicine (e.g., to decide which sequence of treatments is better in terms of survival probability). However, this task is challenging because treatment decisions, patient characteristics, and censoring probability often evolve together, requiring sequential adjustment for both treatments and censoring. Existing survival learners mostly consider a single baseline treatment assignment, while longitudinal treatment-effect learners typically assume uncensored outcomes. In this paper, we introduce a novel doubly robust learner for heterogeneous survival treatment effects under time-varying treatments and history-dependent censoring. Our framework separates the intervention window from the survival horizon, with treatment prescribed for a finite period, after which care follows its natural course. We further establish the favorable theoretical properties of Neyman-orthogonality, rate double robustness and quasi-oracle efficiency. The framework accommodates flexible machine learning models for nuisance estimation and treatment-effect regression. We evaluate our learners through several numerical experiments, demonstrating their effectiveness.
cs.LG / 102 / 2610.26722
Copying Under Repetition: From Prefix Matching to Position Hashing
Abstract
Frontier LLMs can copy random text almost flawlessly, yet they often overrun or truncate long blocks of repeated characters. In repeated strings, matching local content no longer reliably identifies where copying should continue or when it should stop. Nevertheless, we show that a two-layer transformer with $O(\log\log N)$-bit precision and $O(\log N/\log\log N)$ width can copy every string up to a chosen design horizon $N$ and terminate correctly under arbitrary repetition. We use RoPE-style rotations of attention values and outputs (VO-rotations) with small prime periods to encode distances from delimiter anchors modulo those primes. By the Chinese Remainder Theorem, these periodic encodings jointly distinguish all required distances, allowing one retrieval rule to select either the next source token or the end marker. In contrast, we prove a width-independent impossibility result for copying under sequential finite-precision softmax accumulation. It applies to transformers whose only positional input is an additive attention bias that eventually becomes constant, including NoPE and a recursive implementation of ALiBi. Once an identical-token run exceeds the resulting saturation threshold, further repetitions leave the logits unchanged. If this threshold grows sublinearly with $N$, these models cannot both complete a long all-identical copy and stop correctly, a pattern resembling the termination errors of frontier LLMs. In synthetic copying experiments, baseline transformer models fail on long repeated strings even within the training range. Trained VO models, by contrast, generalize beyond the training lengths with near-perfect accuracy up to length $\approx 5005$, the selected primes' period product.
cs.LG / 103 / 2610.26733
Cross-Fitting for Neural Posterior Estimation
Abstract
Neural Posterior Estimation (NPE) fits an amortized conditional density estimator to a simulated training set of complete-data realizations. NPE's training objective, log-likelihood, favors posterior approximations that assign arbitrarily high density to the training points, and typical families of amortized conditional density estimators (e.g., normalizing flows, Gaussian encoders) are flexible enough to oblige. The resulting policy of extreme overconfidence translates poorly to held-out data. Early stopping and other regularizers can prevent this degeneracy, but they are limited to selecting posterior approximations along an unfavorable optimization trajectory set by in-sample residuals, where the posterior mean is underfit while the posterior variance is overfit. To avoid this trade-off, we propose Cross-Fitting NPE (CF-NPE), which uses sample splitting to separate conditional-mean estimation from the estimation of the posterior's higher moments. Using $K$-fold cross-fitting, CF-NPE fits $K$ posterior-mean regressors and then fits a conditional density estimator to the resulting out-of-fold residuals. In experiments with synthetic data and in a cosmological application, CF-NPE achieves higher held-out log-likelihood than standard NPE and typically improves calibration. CF-NPE is most useful in high-dimensional, simulation-limited settings.
cs.LG / 104 / 2610.26734
CAVA: Complementary-Label-Aware Data Valuation
Abstract
Complementary-label learning (CLL) trains a classifier from labels that name a class an instance does not belong to. Such labels are cheap to collect from crowdworkers or vision-language models (VLMs), making them attractive for large-scale annotation. However, both sources can mistakenly assign the true class as the complementary label. Such annotation errors are particularly harmful in CLL because the model is explicitly trained to suppress the correct answer. Optimal transport (OT) data valuation identifies harmful samples by comparing training data with a clean validation set, but transferring it directly to CLL creates a semantic mismatch: label $c$ means not class $c$ in the training but class $c$ in the validation. We show that this makes the OT label groups noisy mixtures whose behavior depends on the noise rate. The transferred score is informative only at sufficiently high noise, becomes uninformative at the crossover rate $1/M$ for $M$ classes, and reverses below it, where practical CLL annotation typically operates. We propose CAVA, which introduces a complementary label cost and a label margin that isolates the contribution of each complementary label to the transport potential. The margin keeps a consistent direction across noise rates, detects diagonal noise with a parameter-free threshold of closed-form precision, and allows detected samples to be recycled as ordinary-label supervision. On synthetic and human-annotated benchmarks, CAVA separates diagonal noise almost perfectly where directly transferred OT valuation performs worse than random. With only a fraction of the clean validation data, it also improves every CLL objective we evaluate, showing that OT-based valuation becomes reliable once the semantics of complementary supervision are modeled explicitly.
cs.LG / 105 / 2610.26747
recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models
Abstract
Protein language models (pLMs) are powerful tools for protein annotation and mutation-effect prediction, but progress has largely depended on scaling model size and data, making them expensive and prone to memorization. Test-time scaling offers a complementary path, but methods developed for autoregressive LLMs do not transfer directly: masked encoders lack a generation loop, and amino acids do not verbalize in a reasoning language. Instead, we adapt recursive pretraining to masked pLMs, where a weight-tied Transformer core refines residue-aligned hidden states over an adjustable number of cycles. We instantiate this recipe as recAMPLIFY, a pLM with 155M unique parameters trained on the same 2T tokens as AMPLIFY-350M. Despite 56% fewer parameters, it improves CASP contact recovery and three out of four supervised tasks, notably raising fold accuracy by 22.5%. Because optimal depth for recAMPLIFY varies across tasks and assays, we then introduce Gated Mixture Recurrence (GMR), a framework that decodes each protein from a probability-weighted mixture of hidden states visited during recursion. GMR employs a learned residual gate at scalar, low-rank, or per-dimension resolution to control how much of each cycle's proposed update is accepted into the state. This unifies halting and gating into a single adaptive computation objective while recovering PonderNet-style halting as a special case without gates. Layered on top of recAMPLIFY, GMR further reduces expected layer depth by up to 37%, although the accuracy trade-off varies by task. Mechanistic analyses show that residual gating controls recurrence stability beyond the trained depth and that halt depth tracks biological properties.
cs.LG / 106 / 2610.26749
Trajectory Replay: Continuous-Time Mixture Transport for Lifelong Generative Models
Abstract
Adapting pre-trained continuous-time generative models to sequentially incoming data without forgetting earlier tasks is a central challenge in lifelong learning. A leading strategy, generative replay, uses a generative model trained on historical tasks to augment its own continual update with synthetic samples from experience. However, it requires full trajectory simulations at every training step, which accumulate discretization errors and can degrade the replayed historical distribution. Here, we propose trajectory replay, which bypasses sample generation by training an updated velocity field to match both a frozen historical velocity along its trajectories and the incoming task's velocity field, weighted by a stability-plasticity coefficient. Through the lens of continuous-time transport, we show the resulting optimum averages the two task velocities by their respective densities and generates the mixture of their probability paths. Combined with mean velocity distillation, trajectory replay accesses historical trajectories using only a few network evaluations per continual update. Empirically, it outperforms generative replay and distillation baselines on class-incremental image generation. In two biological case studies, trajectory replay extends generative perturbation models to new contexts and experimental batches across two data modalities while preserving treatment-effect prediction.
cs.LG / 107 / 2610.26757
Simulation Priors for Rapid Adaptation in Model-Based RL for Autonomous Racing
Abstract
Reinforcement learning enables learning high-performance control policies, but its reliance on extensive interaction limits deployment on physical systems where data collection is costly, and failures are unsafe. We introduce Prior-Informed Adaptation (PIA), an offline-to-online model-based RL framework that uses structured simulation priors for rapid adaptation. PIA pretrains a world model on heterogeneous low- and high-fidelity simulated trajectories, then adapts it with limited target-domain interaction using uncertainty-regularized planning. Rather than requiring an accurate simulator for policy transfer, PIA treats simulated trajectories as biased priors whose residual mismatch is corrected during online adaptation. We evaluate PIA in autonomous racing using high-fidelity simulation and on a real full-size autonomous race car. In simulation, PIA adapts faster across track and surface shifts than strong baselines, achieving lower lap times and fewer failures. On hardware, PIA achieves reliable few-shot adaptation with less than one hour of on-vehicle interaction.
cs.LG / 108 / 2610.26770
HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs
Abstract
Harness development plays a crucial role in building capable AI agents. Recent approaches improve harnesses through recursive self-improvement (RSI), but their performance gains often come with higher execution costs. We introduce HarnessPro, a label-free method for evolving harnesses to improve task performance while reducing execution cost. In this setting, candidate harnesses are evaluated using LLM-generated tests without access to ground-truth solutions or official tests. Specialized agents diagnose execution trajectories, plan updates with different performance-cost priorities, and implement them as candidate harnesses. Rather than returning a single harness, HarnessPro uses normalized weighted squared-distance selection to provide cost-focused (lite), balanced (medium), and performance-focused (heavy) harnesses for different deployment budgets. Selected harnesses enter the next iteration, where execution feedback guides further updates. Our experiments cover Terminal-Bench 2, SWE-Bench Pro, and ProgramBench, using GPT-5.5 with Codex and GLM-5.2 with Claude Code. Across these benchmarks using GPT-5.5 with Codex, the heavy mode of HarnessPro achieves an average pass-rate gain of 2.0 percentage points and an average per-task cost reduction of 26.1% relative to RHO. This advantage also holds when using GLM-5.2 with Claude Code. Code is available at https://github.com/pzs19/HarnessPro
cs.LG / 109 / 2610.26777
Topological Sheaves of Irreducibility: A Theoretical Framework for Multi-Agent Task Obstructions
Abstract
Multi-agent systems can differ under identical agents and communication budgets when they support different forms of joint interaction. We introduce Topological Sheaves of Irreducibility (TSI), a sheaf-theoretic framework separating task-level compatibility obstructions from system capability. A task is represented by an interaction complex $K$ and a capacity sheaf $\mathcal{F}$; nonzero $H^n(K;\mathcal{F})$ defines a structural signature, while registered interventions test capability. For cocycle data, cochain realization is solvable exactly when the represented cohomology class vanishes. Examples show that the signature is sheaf-relative, and the degree-1 reduction yields cycle-space witnesses and observation requirements. We evaluate two finite one-round models. Across 15,625 degree-1 inputs, an all-agent additive primitive solves decision and repair exactly; the optimal pairwise system reaches 88.80% on decision, one input above the 88.79% majority-class baseline, and 26.63% on repair. On a degree-2 tetrahedral-boundary task over 531,441 inputs, the same-2-skeleton system retains all four three-agent primitives yet matches the pairwise-only optimum throughout the registered separation window, whereas adding the class-filling four-agent primitive makes both tasks exact. Status-preserving robustness checks retain exact performance but are not cost-matched to the class-restoring ablation. Conditional on the registered primitives and cost models, TSI provides task-relative, falsifiable tests of higher-order interaction.
cs.LG / 110 / 2610.26794
On Data Scaling in Tabular ML: More Data, Better Generalization?
Abstract
Recent work in tabular machine learning has emphasized the importance of diverse and realistic evaluation settings. The niche of large-scale (1M+) data, however, remains under-explored. To study training-data scaling, we assemble eight real-world tabular datasets with up to around ten million training examples, retaining the natural distribution shifts of the tasks and using random-split evaluation as a control. We find that GBDTs and MLP-based models show broadly similar scaling trends, with gains from both additional training data and algorithmic improvements. Tabular foundation models show more varied scaling behavior, with recent versions benefiting more consistently from larger contexts even under distribution shift. However, scaling training data under distribution shift introduces new challenges: in-distribution and out-of-distribution performance mostly improve together at smaller scales but can diverge at larger scales. We examine this through MLP capacity sweeps and comparisons of the validation distributions used for tuning and early stopping in MLPs and GBDTs. We also compare common practices of using validation data, contrasting its use for model selection in MLP pipelines with using it for in-context fitting in tabular foundation model pipelines. Under stronger shifts, MLPs selected using validation representative of the evaluation conditions generally maintain or improve performance as training data grows. In contrast, including such examples in context improves foundation-model performance, but does not prevent deterioration with increasing training size in our experiments.
cs.LG / 111 / 2610.26795
Two hidden layer suffice: the caterpillar net
Abstract
Every continuous piecewise linear (CPWL) function is a finite signed sum of maxima of affine functions, making maximum representation a central problem in the study of the expressiveness of ReLU networks. A now-disproved conjecture predicted logarithmic depth for maxima. Inspired by the proof by Brunck et al. (2026) that two hidden layers suffice for general CPWL functions, we give the Caterpillar Net, a nonrecursive construction for the maximum of $N$ inputs with two hidden layers of widths $O(N^2)$ and $2^{O(N\log N)}$. The construction has a direct geometric interpretation.
cs.LG / 112 / 2610.26800
Learning to Refer: Inference-Time Collaboration in Decentralized Federated Learning
Abstract
In decentralized federated learning, heterogeneous clients may develop complementary expertise, yet each query is typically evaluated only by the client where it arrives. To exploit this distributed expertise without global coordination, we formulate inference-time collaboration as budgeted local expert referral, a graph-constrained routing process in which a query can move only between neighboring clients and only visited clients contribute to its final prediction. EdgeDRM implements this framework by learning a density-based routing potential for each client: differences between neighboring potentials provide the comparisons used to select successive referrals, while local task models generate predictions that are aggregated along the resulting route. Both training and inference are decentralized, using only neighbor-local information without a central router or graph-wide client scoring. We establish estimation guarantees for the learned comparisons and characterize how local improvement and graph structure govern prior-weighted density gains from referral. Experiments on benchmark and real-world medical data show that EdgeDRM outperforms graph-matched decentralized baselines in most settings, with the clearest benefits under sparse connectivity.
cs.LG / 113 / 2610.26802
Scaling Parameter-Efficient Fine-Tuning for Brain Encoding
Abstract
Brain encoding models aim to predict neural responses to stimuli. The most widely used approach fits readouts from frozen pretrained backbones. Using neural data to train large models from scratch or fully fine-tune them after task-pretraining has proven difficult because neural data is costly and limited. We show that parameter-efficient fine-tuning (PEFT) enables us to effectively adapt very large pretrained backbones for brain encoding with a small number of trainable parameters. To make it computationally practical, we exploit the backbone's causal structure to predict multiple timepoints in a single pass instead of using sliding windows, significantly reducing feature-extraction cost. This enables efficient fine-tuning of multimodal backbones with up to 104B parameters, the largest model adapted for brain encoding to our knowledge. Trained with a low-rank adapter (LoRA) and a lightweight readout, our approach consistently improves brain encoding performance across nine diverse multimodal backbones. On the same Qwen3-Omni backbone as the prior state-of-the-art (SoTA), our method BrainLoRA achieves the best published performance on the Algonauts 2025 Challenge. We also show the adapted multimodal backbones remain interpretable: modality attribution recovers the canonical sensory topography that prior work often obtains by construction from separate unimodal encoders, and neural adaptation increases the contribution of audio inputs and cross-modal mixing within the backbone. To scale out the number of adapted models, we develop a multi-LoRA kernel that accelerates joint training of multiple LoRAs on a shared backbone. This enables an ensemble of 200+ adapted models that ranks first on the public Algonauts post-challenge leaderboard at the time of writing. Together, our results highlight PEFT as an efficient and powerful paradigm for brain encoding.
cs.LG / 114 / 2610.26817
Next Best Action: Causal Constrained Offline RL for Business Action Targeting
Abstract
Next Best Action (NBA) systems must select among customer interventions (e.g., discounts, promotions, and lifecycle actions), maximize long-term economic value, satisfy complex operational constraints, and remain causally grounded. Existing approaches lack a rigorous framework for meeting these requirements jointly. We address this gap by formulating NBA as a Constrained Markov Decision Process (CMDP). We then integrate off-policy causal correction of the estimated customer dynamics directly into the Linear Programming (LP) formulation of the CMDP. The resulting policy selects actions based on customer states to maximize long-term value while explicitly enforcing business guardrails. Furthermore, the LP formulation yields shadow prices that quantify the economic trade-offs of operational constraints (e.g., fairness and user welfare), enhancing policy interpretability. We validate our proposed offline Reinforcement Learning (RL) framework on an industry case study involving rideshare coupon targeting, where customer states are defined by Recency, Frequency, and Monetary (RFM) stages. We evaluate performance using model-based off-policy evaluation, with causal dynamics estimated from a temporal holdout. Under the observed coupon budget and RFM state representation, our optimized policy improves customer lifetime value (LTV) by 19% over the estimated behavior policy and 8% over a myopic baseline. Although the demonstrated action space is binary, the proposed framework is action-agnostic, yielding a general-purpose decision module for targeting systems that manage multiple interventions.
cs.LG / 115 / 2610.26834
Kronecker Delta Rule for Long Context Linear Attention
Abstract
We propose KronLSTM, a linear attention architecture based on the Kronecker Delta Rule, a novel update rule for associative matrix memories. KronLSTM extends the xLSTM family with a memory-revision mechanism that corrects targeted associations while modifying the memory only along a single direction in matrix space. This minimally invasive memory revision reduces interference with previously stored information compared with delta-rule updates, making it particularly well suited for long-context language modeling. The resulting memory dynamics can be mathematically described as a Markov chain that converges to a steady-state distribution, while the expected norm of the memory matrix remains uniformly bounded throughout the entire input sequence. Hence, KronLSTM ensures stable memory dynamics without requiring additional normalization and stabilization mechanisms for the memory state. Experiments show that KronLSTM is competitive with, and often outperforms, existing linear attention architectures, including models of the (Gated) DeltaNet family, with particularly strong performance on long context benchmarks. These results establish the Kronecker Delta Rule as an effective and analytically tractable foundation for stable associative memory in recurrent linear attention architectures. Our code (https://github.com/NX-AI/KronLSTM) and models (https://huggingface.co/collections/NX-AI/kronlstm) are open-source.
cs.LG / 116 / 2610.26847
WeAct: Dynamic Weight Switching as Agent Actions
Abstract
Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. Updating shared parameters to learn new capabilities can degrade previously acquired ones, resulting in catastrophic forgetting. As a trajectory unfolds, different stages may require different capabilities, yet conventional joint training and static adapter merging maintain a fixed parameter configuration throughout execution. We introduce WeAct (Weight-level Internal Actions), which trains independent capability-specific LoRA adapters over a frozen backbone and adds adapter selection to the agent's action space. These internal actions use the model's native tool-calling interface for on-demand capability composition in a shared context. Across eight deep-research benchmarks, Qwen3-32B with WeAct improves average accuracy by 6.2% relative to full fine-tuning counterparts despite using less compute, and by 27.5% relative to the strongest static LoRA-merging method. WeAct also supports modular continual adaptation, by adding a new code computation adapter to expand the agent's capabilities and updating the existing writing adapter to improve report quality. Both updates leave the backbone and other adapters unchanged, supporting modular continual adaptation in agents.
cs.LG / 117 / 2610.26863
Text-to-Image Generative Modeling Without Language Pretraining
Abstract
Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2$\times$ less pretraining compute. Further, Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275$\times$ less pretraining compute. Controlled experiments show that text-only pretraining does not improve T2I generation. Halcyon also learns image captioning by replacing caption dropout with I2T supervision, maintaining competitive T2I generation quality under the same pretraining budget. We show that strong T2I systems can be trained from scratch without internet-scale language pretraining. Code and models are available at https://cs.stanford.edu/~keshik/halcyon
cs.LG / 118 / 2610.25459
BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks
Abstract
Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models insufficiently characterized. In this paper, we present BiDex-100, a simulation benchmark for bimanual dexterous manipulation that combines broad task coverage, multi-dimensional evaluation, and high-DoF hand embodiments. Concretely, the benchmark comprises 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes, allowing systematic assessment of robustness. In addition, the benchmark supports two embodiments and provides 8,080 success-validated teleoperated demonstrations, totaling 6.25 million frames. We evaluate seven recent Vision-Language-Action models under a unified protocol, establishing baselines for capability-specific performance, robustness, and data efficiency. The website is available at https://BiDex-100.github.io/BiDex-100/.
cs.LG / 119 / 2610.26395
ForagerSim: Evaluating Embodied Map-Based Reasoning in Large-Scale Dynamic Worlds
Abstract
Animals build cognitive maps of vast environments, use them to perform map-based navigation, and might even be reusing the same circuitry used in spatial reasoning for creating and using maps of arbitrary task-spaces. In contrast, current embodied AI systems either rely on hand-engineered mapping pipelines, which break in the presence of unmodeled dynamics and are restricted to metric space, or on learned policies that rarely form map-like representations at all. Developing architectures that could learn to build and use map-like representations as flexibly as animals is, however, difficult because few embodied AI benchmarks actually require map-based reasoning. To address this, we present ForagerSim --- an interdisciplinary research platform and simulation environment focused on embodied tasks that require map-based reasoning, with scenarios ranging from mobile-robot search and rescue to animal foraging in both complex mazes and vast, potentially infinite open areas. ForagerSim is highly modular and allows modifying independent axes of difficulty related to perception (abstract 2D sensing to full RGB or LiDAR), scale (room-scale to multi-km scale), topology (open-fields to complex mazes), dynamics (smoke, debris, sensorimotor failures and others) and tasks (from simple 2D coverage to foraging-inspired well-alternation tasks). We compare reinforcement learning baselines with a simple engineered robotics stack and a human player on 8 example worlds which show how different difficulty axes affect existing methods. Our results show that, when sensing is simplified to 2D, learned policies can perform multi-object search and well-alternation tasks surprisingly well in simple topologies, but they fall short of human performance in larger or more complex-topology environments where map-based baselines perform well, suggesting that map-based reasoning is the bottleneck in these scenarios.
cs.LG / 120 / 2610.26605
A Probabilistic, Rotation-Preserving Kernel Operator Surrogate for Continuum-Robot Design Spaces
Abstract
We develop an uncertainty-aware, operator-learning-based surrogate for tendon-driven continuum robots (TDCRs) that generalizes across robot design and actuation while predicting the complete equilibrium configuration. Our central idea is a nonstationary Gaussian-process (GP) kernel whose metric is learned by a shallow neural network from mechanics-aware design coordinates. The network learns how similarity changes across design space, while the resulting kernel defines the design-to-pose operator, its posterior uncertainty, and a criterion for sequential simulator acquisition. We represent backbone position and rotational strain with spatial kernels and reconstruct orientation directly on SO(3), preserving valid material frames and tendon geometry. We use ASkotch with Nystrom preconditioning for scalable kernel operator solves and a separate Nystrom approximation for posterior covariance. Experiments show that kernel-coverage acquisition reduces mean trajectory error by approximately 15% relative to IID sampling, while posterior variance increases for increasingly unsupported designs. Compared with prior neural-operator TDCR models, the method is 30-50% more accurate using approximately 5x fewer training designs while additionally providing posterior uncertainty estimates.
cs.LG / 121 / 2610.25635
Uncertainty-Gated SHAP-Dirichlet Refinement for MRI-Based Depression Classification
Abstract
Structural MRI-based classification of major depressive disorder (MDD) remains challenging due to subtle, heterogeneous anatomical changes. We propose uncertainty-gated SHAP-Dirichlet refinement, a post-hoc reliability layer that improves predictions without retraining the classifier. The method retrieves similar labeled SHAP explanations, constructs Dirichlet class evidence, and selectively refines uncertain predictions through probabilistic fusion, while allowing unreliable cases to be deferred for expert review. We evaluate five classifiers on 504 SRPBS subjects (252 controls and 252 MDD patients) using 159 anatomical features and repeated nested cross-validation. Refinement improved accuracy across all 8 repetitions for TabM and ResNet, yielding mean gains of 1.04 and 0.45 percentage points (sign test p=0.008), while modifying at most 8.3% of predictions. Gains were concentrated among low-confidence cases, with significant improvements in calibration. The method also benefited an RBF support vector machine but not logistic regression or random forests. Ablations confirmed the importance of uncertainty gating and soft fusion. Dirichlet evidence identified baseline errors (AUROC 0.62), while deferring 50% of cases increased retained-case accuracy to approximately 75%. These findings support selective explanation-guided refinement as a practical approach to improving prediction reliability and facilitating expert triage in MDD classification.
cs.LG / 122 / 2610.25643
Sampling of Signals on Multigraphs: From Removable and Uniqueness sets to Spectral Proxies
Abstract
In this paper, we study the sampling of bandlimited signals on multigraphs, where several graphs with different shift operators are defined on the same set of nodes. Since the shift operators do not commute in general, they admit only a joint block diagonalization, and the frequencies of the multigraph signal model are matrices rather than scalars. Relying on a formal algebraic convolutional signal model, we introduce a notion of frequency and bandwidth for multigraph signals that does not depend on any individual layer, and we extend the notions of removable and uniqueness sets to multigraphs. We prove that the optimality of a sampling set is determined exclusively by the common Fourier basis and its invariant subspaces, and not by any individual layer, although the individual layers provide computable sufficient conditions that certify a sampling set as a uniqueness set. Since identifying which layers provide reliable certificates requires spectral information, we propose an aggregate operator that pools the information of all the layers, is computable directly from the shift operators, and certifies uniqueness without any assumption on the individual layers. This operator leads to a greedy sampling algorithm that does not require computing the joint block diagonalization. Numerical experiments on multigraphs whose layers order the frequencies inconsistently show that the proposed algorithm selects better-conditioned sampling sets and attains lower reconstruction errors than single-layer methods.
cs.LG / 123 / 2610.26153
Physics-Driven Framework for QoS-Constrained Satellite System with Non-Linear PA Distortion
Abstract
This work develops a physics-driven self-supervised graph learning framework that directly optimizes the distortion-aware precoder under nonlinear correlated power amplifier (PA) distortion, without relying on expensive optimizer-generated labels for a user-centric satellite system. We minimize nonlinear PA consumption subject to user quality-of-service (QoS) constraints under PA nonlinearity for a downlink multi-input-single-output (MISO) satellite system. Our pipeline jointly designs the precoder and admission control that minimize nonlinear PA consumed power. In particular, a graph neural network (GNN) for the design of a distortion-aware precoder is optimized end-to-end in a self-supervised manner using a differentiable Itô-Hermite model of the true modified Rapp amplifier model. The precoder output by the GNN is used to initialize a low-complexity distortion-aware signal-to-interference-distortion (SINDR) tracking algorithm that jointly optimizes for power and admission control. Our proposed physics-based pipeline closely tracks a successive convex approximation (SCA) based solution in consumed power performance with the advantage of faster runtime and a wide operating range covered during training.
cs.LG / 124 / 2610.26742
Polar Switch: Fast Muon Orthogonalization with Polynomial Circuits
Abstract
Muon, a 2024-born optimizer for language models, differs from the 2015-born spectral gradient descent mainly in its fast, GPU-friendly Newton--Schulz polynomial for orthogonalizing the gradient matrix. Accordingly, a line of research proposes faster polynomial iterations for gradient orthogonalization, such as Polar Express and Gram Newton--Schulz, both built on quintic polynomials. In this paper, we introduce Polar Switch, a new approach to polynomial gradient orthogonalization. Using simple polynomial circuits, Polar Switch reaches a given high polynomial degree at a lower FLOP cost than quintic iterations. We define a circuit as a short sequence of symmetric matrix products on a few square buffers, and circuits with $k$ products form a restricted family of odd polynomials of degree up to $2^{k+1}+1$. We choose the sequence of products in each circuit, their coefficients, and how the circuits are composed by numerically minimizing a worst-case bound on the polar approximation error. This minimization is subject to a FLOP budget that accounts for symmetric matrix multiplication kernels and a rounding constraint for numerical stability in half precision. Our two recommended Polar Switch methods have 15.5--18.7\% lower runtime than Polar Express on square matrices and 8.6--16.2\% lower runtime than Gram Newton--Schulz on rectangular matrices with aspect ratios 3.25, 3.5, and 8, with comparable approximation accuracy. In pretraining Modded-NanoGPT \#91, Qwen3-0.6B, and Gemma 3-1B, Polar Switch reaches final validation losses similar to those of established orthogonalization methods. Across the more than 14 million matrices that each of the two recommended methods orthogonalized in these runs, the estimated output spectral norms never exceed the designed bounds by more than 3\%, demonstrating their numerical stability.
cs.LG / 125 / 2610.26748
Gradient descent reaches arbitrary accuracy in a fixed number of large steps on multiclass logistic regression
Abstract
We show that gradient descent (GD) with step size $η$ can achieve $\varepsilon$ optimality gap on the multiclass logistic regression loss in a number of iterations $T$ that is independent of $\varepsilon$. To do so, the step size $η$ must be chosen large enough as a function of $\varepsilon$ (and properties of the linearly separable dataset), while $T$ is a constant that only depends on the data. The relevant properties of the data include the number of points and their margin (for $T$), and one additional quantity (for $η$) that turns out to be necessary in dimension $d \geq 3$. The proofs build on the known fact that, as $η\to \infty$, suitably rescaled GD becomes the perceptron algorithm, and the latter finds a perfect classifier in $T$ iterations. A similar claim was shown for two classes in dimension $d = 2$ using other proof techniques. This result adds to previous works that showed GD on the logistic regression loss can be accelerated by taking large step sizes. Those provoke a wild, oscillatory initial phase, followed by a stable phase. We provide results of this (qualitatively different) nature, too.
cs.LG / 126 / 2610.26839
Newton-Based Methods for Bilevel Optimization Beyond Lower-Level Strong Convexity
Abstract
Bilevel optimization provides a framework for machine learning tasks such as hyperparameter optimization, data reweighting, and meta-learning. Existing Newton-based bilevel methods typically rely on the classical hypergradient formula, which is generally unavailable when lower-level solutions are nonunique and the lower-level Hessian is singular. We develop two Newton-based methods beyond lower-level strong convexity by exploiting curvature shared by the lower-level and multiplier updates. Newton-based GALET handles potentially singular lower-level curvature under a positive spectral-gap condition, exploiting its range structure to avoid the squared-Hessian multiplier system used in GALET. Newton-based BAMM uses a strongly convex aggregation model when the lower-level objective is convex and the upper-level objective is strongly convex in the lower-level variable. We establish convergence guarantees and improved complexity bounds relative to GALET and sl-BAMM under the corresponding assumptions. We further introduce GNBO, a generalized Newton-based framework that captures the shared-curvature structure of the proposed methods and several existing Newton and quasi-Newton bilevel methods. Experiments on synthetic problems, data hyper-cleaning, and GPT-2 data reweighting evaluate the proposed methods across different bilevel learning tasks.
cs.LG / 127 / 2610.25565
Concurrent two-scale topology optimization of printable bone scaffolds via rectified-flow generation and exact latent-space refinement
Abstract
Bone scaffolds must reconcile a macroscale mechanical response with microscale geometric feasibility, yet conventional approaches to their design restrict attainable constitutive properties. This work presents a concurrent two-scale topology optimization framework; a conditional rectified flow generates periodic lattice microstructures, which are then refined in latent space by exact finite-element homogenization sensitivities. The resulting unit-cell geometries and their spatial assignments are jointly optimized under mechanical and feature-size constraints. In a distal rabbit femur benchmark, all eight flow-started refinements produce printable cells with maximum relative errors under 1\% across six prescribed engineering moduli. Constructing the implant from a single tiled cell reproduces intact whole-femur compliance within 0.1\% and regional strain energy within 0.35\% under four loading conditions. In human femoral reconstruction, multiscale concurrent optimization produces a cortical mid-shaft scaffold with porosity 63\% while satisfying the compliance constraint within 2\% under walking and stair-climbing loads. It achieves an 84\% porosity on a proximal cancellous scaffold with strain-energy deviations below 0.25\% under the same loads. Analysis of the learned representations reveals distinct architectural families in latent space. Further, the attention heads of the rectified-flow transformer specialize for spatial shifts -- learning the analogue of a finite-difference stencil -- and coordinate symmetries, with ablations showing their influence on generated topologies. These results establish a route to printable, mechanically constrained scaffold designs. The framework is general enough for extending to many other multiscale structural design applications including topology-microstructure co-design in metal additive manufacturing.
cs.LG / 128 / 2610.26138
Reinforcement Learning Controllers for Fusion Scenario Development
Abstract
Scenario development, the design of the waveforms that take a tokamak discharge from breakdown to flat-top, is still an iterative, labour-intensive process. Every candidate scenario has to be simulated with a controller in the loop, and the classical shape and current controllers need low-level actuator references. So the process is iterative and includes a number of trial-and-error steps. We propose to replace the per-scenario controller tuning by a single goal-conditioned reinforcement-learning controller trained once, in advance, against a free-boundary equilibrium and transport simulator. A scenario is then specified only by the target waveforms of the physical quantities of interest (plasma current, line-averaged electron density, LCFS shape, x-point positions), and the controller produces the coil-current and fueling commands that realize them. We demonstrate the approach on an EAST plasma-current ramp-up: the controller takes over a small limited plasma, ramps the current and the density, grows the cross-section, forms two x-points and holds a double-null diverted configuration against sampled shape targets for the rest of a 1.2 s episode. The neural network controller is trained using TQC algorithm. On 1024 test scenarios sampled from a low-discrepancy scrambled Sobol sequence in 11-dimensional target space, the controller forms the divertor in 86% of the cases, holds it for more than 0.5 s in 67% and reaches the end of the episode in a diverted state in 58%, with flat-top tracking errors of a few kA in plasma current, $\sim 0.15\times10^{19}$ m$^{-3}$ in density, $\sim 2$ cm in position and $\sim 8$ cm in x-point location; the failures concentrate at the high-current density ($J$) region which is believed to be physically or from engineering perspective not feasible within ohmic mode.
cs.LG / 129 / 2610.25671
Arbitrage-Free Implied Volatility Surface Reconstruction via Additive Martingales
Abstract
This work focuses on reconstructing a continuous implied volatility (IV) surface from a small number of option quotes, a fundamental task for option pricing and risk management. We propose Additive Martingale (AddMart), a few-shot IV surface reconstruction framework that satisfies static no-arbitrage conditions by construction. Rather than predicting IVs directly, AddMart maps sparse quotes to non-negative increment densities of an exponential additive martingale, from which valid risk-neutral marginals and IVs are recovered. By incorporating the no-arbitrage structure directly into its architecture, rather than enforcing it through regularization in the learning objective, AddMart produces IV surfaces that satisfy static no-arbitrage conditions throughout the domain. Extensive experiments on ten real-option datasets demonstrate that AddMart achieves competitive reconstruction accuracy with no material static arbitrage violations, while directly enabling the estimation of VaR, CVaR, and crash probabilities.
cs.LG / 130 / 2610.26456
Regime-Aware Interpretable Ranking for Sector Rotation
Abstract
Sector rotation is a common decision in benchmark-relative portfolio management, where portfolio managers rank sectors for overweight and underweight positions. We formulate this task as ranking sector actions by expected future risk-adjusted active performance, measured with forward sector-benchmark Sharpe ratios. Such rankings should preserve temporal information, express model outputs through economically familiar signal characteristics, account for changes in signal relevance across macroeconomic environments, and capture dependencies among sectors. We propose RegimeRank, an interpretable sector-ranking model that represents fundamental and technical signals through expert-defined temporal characteristics, conditions them on learned macro factors via cross-attention, and models cross-sector dependencies through self-attention. Expected-gradient attributions identify the contributions of individual signal-transformation pairs, while macroeconomic and cross-sector attention maps provide diagnostics of the learned conditioning and relational structure. We evaluate RegimeRank on weekly S&P 500 data from 2002 to 2024 under combinatorial purged cross-validation, where it outperforms representative ranking and machine-learning baselines across multiple ranking metrics. Ablation studies support the contributions of latent macro states, regime conditioning, and cross-sector modelling, while attribution and attention analyses illustrate the temporal characteristics, macro states, and sector relationships underlying its rankings.
cs.LG / 131 / 2610.26833
Supervised Learning with Quantum Hamiltonian Embeddings
Abstract
Machine learning is widely regarded as a promising application in which quantum computing could offer a computational advantage. However, general quantum machine learning models with provable guarantees on performance, trainability, and efficient inference remain limited. In this work, we introduce quantum Hamiltonian embeddings, a supervised learning framework in which a classical training dataset is encoded into a single effective, parameterized Hamiltonian. Training optimizes the parameters of this data-dependent Hamiltonian, whose approximate ground state defines the classification rule. Once the ground state has been approximated using a quantum algorithm, inference can be performed either classically using a classical representation of the learned state, or with a quantum representation of the state. We establish efficient methods for gradient estimation and prove favorable generalization guarantees. Furthermore, under standard cryptographic assumptions, we show that Hamiltonian embeddings can realize hypothesis functions beyond the reach of efficient classical algorithms. We then demonstrate the framework at scale on IBM Heron processors using a standard benchmark dataset. Together, these results establish Hamiltonian embeddings as a new framework for quantum machine learning, combining theoretical guarantees with a path for experimental investigations at scale on quantum hardware.
cs.LG / 132 / 2610.26267
A Pitfall in the Evaluation of Gene Perturbation Models
Abstract
Gene perturbation models aim to predict transcriptional responses to unseen genetic interventions, but comparisons rely on heterogeneous evaluation metrics and differentially-expressed (DE) gene subsets. We investigate the consequences of these choices using eight methods across six single-gene perturbation datasets under a unified evaluation protocol. Varying the number of top DE genes reveals substantial instability in Precision-based rankings, with the winning method changing $86.7\%$ of the time, while rankings based on the rest of the metrics remain comparatively stable. Investigating this discrepancy exposes that repeating a single predicted expression profile across cell-level rows, a standard practice for many methods, can inflate significance estimates and produce excessively large predicted significant-DE sets. Specifically we found that $67.1\%$ of tested perturbation--gene pairs pass the significance threshold despite some of them having minuscule effects. We derive an inflation factor showing that replicated predictions can amplify the standardized Mann--Whitney statistic even after correction for tied ranks, without adding independent information. This artifact changes Precision mainly as the number of DE genes increases because for smaller cutoffs, most of the genes are already significant due to the effect-based ordering. Taking this pitfall into account in our benchmark comparison, we find that a random forest using Gene Ontology features achieves the best overall results.
cs.LG / 133 / 2610.25796
Fair Regression under Approximate Demographic Parity
Abstract
Exact demographic parity requires the same prediction distribution across sensitive groups. In regression, this can be too restrictive and reduce predictive accuracy. We instead bound every pairwise Wasserstein distance between group-specific prediction distributions by a chosen tolerance, allowing the groups to differ while keeping fairness tunable. On the real line, the prediction and constrained transport problems have the same infimum. If the transport problem attains this value and suitable Monge maps exist, a deterministic transport-form predictor realizes it. Under a standard one-dimensional condition, the problem becomes a bounded quantile optimization that is convex when the conditional expected loss is convex in prediction. Under squared loss with two positively weighted groups, the $W_1$, $W_2$, and $W_\infty$ constraints yield soft-thresholding, shrinkage, and clipping. Under the stated conditions, we establish rates for the known-cost plug-in estimator. When the cost is unavailable, we describe an observed-loss computational proxy. Synthetic and California 2018 ACS studies illustrate the tunable fairness-accuracy trade-off and distinguish calibration feasibility from held-out fairness and accuracy.
cs.LG / 134 / 2610.25997
Federated Community Detection in Bipartite Networks
Abstract
We study federated spectral estimation when a large matrix is horizontally partitioned across servers that cannot share raw data. We propose FLICKER, a target-specific gradient-based power iteration that recovers a singular direction associated with a prescribed spectral target using only matrix-vector products on local blocks. Different target directions can be recovered independently and in parallel. We establish non-asymptotic convergence guarantees and provide a recursive grid-search procedure for spectral target selection. Building on FLICKER, we develop BiFLICKER for federated community detection in bipartite networks, together with a sequential projected variant for nearby or repeated singular directions. Under a sparse bipartite stochastic block model, we show that properly rescaled leading singular values from a reference server provide valid global targets, and that Sequential BiFLICKER achieves a vanishing misclustering rate and exact recovery under stronger sparsity conditions. Simulations show that BiFLICKER remains close to centralized spectral clustering and is more robust than existing federated spectral baselines when local sample sizes are small. On a LastFM user-artist network naturally partitioned by country, Sequential BiFLICKER attains an overall ARI of 0.94 against centralized clustering while revealing cross-country communities with distinct listening profiles.
cs.LG / 135 / 2610.26029
Towards instance-dependent regret optimality in Episodic MDPs with Posterior Sampling
Abstract
We study regret minimization in finite-horizon episodic Markov Decision Processes (MDPs). While minimax-optimal algorithms are known, tractable approaches to instance-dependent optimality are still lacking. Motivated by this gap, we introduce $π_0$-PSRL, a variant of PSRL that uses posterior samples to decide when exploration is needed, while following a fixed reference policy $π_0$ during exploration episodes. This decouples the test for exploration, triggered when the sampled and empirical MDPs have different optimal policies, from the choice of the policy used to gather information. The resulting design addresses a limitation of standard PSRL, where the sampled optimal policy may not be the optimal choice for exploration. We prove instance-dependent regret bounds for $π_0$-PSRL, identifying the logarithmic exploration cost induced by $π_0$ and taking a step toward matching asymptotic lower bounds for episodic RL. Our proof techniques showcase a novel proof structure to derive problem-dependent regret bounds in episodic MDPs, and concentration results for Dirichlet random variables, that may be of independent interest.
cs.LG / 136 / 2610.26380
Selecting the Best Few with Human Evaluators: Adaptively Combining Ratings and Pairwise Comparisons
Abstract
This paper considers a decision-maker tasked with identifying the $k$ best items from a large pool of candidates, such as job applicants, LLMs, award nominees, or new product ideas. The decision-maker learns about the items' quality by soliciting human evaluations. Each evaluation can be costly, so the decision-maker aims to use the limited evaluation budget efficiently by allocating evaluations adaptively. A key consideration is the form of evaluation to solicit, which should be something evaluators are comfortable providing. Two popular forms are ratings and pairwise comparisons, and we allow for both, as they may be especially useful in combination given the limited budget. Ratings can be used early on to get a coarse picture of which items are good and which are not. As that picture forms, comparisons can refine the ordering of the good items, narrowing in on the ultimate top-$k$ selection. To solve the allocation problem, we develop an adaptive algorithm built from two allocation policies. The first, our primary contribution, is based on a surrogate objective that is naturally forward-looking, valuing evaluations that help inform what to evaluate in future rounds, not just those that directly improve the selection decision. The decision rule is interpretable, and we prove that it is asymptotically consistent. Late in the evaluation process, when being forward-looking is less productive, we construct a myopic policy and a criterion for switching to it. In simulation, the method gives insight into how to combine these two types of evaluations, and beats the (limited) existing benchmarks.
cs.LG / 137 / 2610.26437
R2-GP: Variance-Calibrated Priors for Additive Gaussian Processes
Abstract
We develop R2-GP, an R2D2-type prior for Gaussian process (GP) generalised additive models (GAMs) that calibrates explained variance ($R^2$) for Matérn covariance kernels. The function variance for standard Matérn kernels depends jointly on the effect magnitude and length-scale. As such, smooth and rough components with identical amplitudes contribute very different amounts of variation. With independent priors on component amplitudes, this variation results in variance overestimation, especially in higher dimensions. We overcome this by taking direct control over each component's expected contribution to prior explained variance. We do this by rescaling each component based on its length scale and observed inputs, and modelling each rescaled variance as a share of the function variance induced by a global $R^2$ prior. This allows R2-GP to scale well to many components. To also scale to large sample sizes, we extend R2-GP to Hilbert space approximation methods, a framework for kernels exploiting their spectral decompositions for fast and efficient inference. Our simulations and real-world case studies demonstrate superior generalisation and parameter recovery over independent variance priors.
cs.LG / 138 / 2610.26789
Cooperative Multi-Agent Reinforcement Learning for Market Making on Prediction-Market Order Books
Abstract
A market maker provides liquidity by quoting prices to buy and sell an asset, earning the spread while managing inventory risk. We study market makers trained with multi-agent reinforcement learning on prediction markets, specifically Polymarket, whose thin order books pose distinct challenges. Three challenges underpin this research: adverse selection may cause a policy trained on real data to stop quoting; the consequences of a quote can arrive many steps after the action; and prediction-market L2 data does not identify individual orders needed for simulation. We address these challenges by reconstructing order-level LOBSTER-format data from Polymarket and training agents in JaxMARL-HFT on 417 real markets. We enrich the baseline market maker's observation space with the book and recent flow, redesign its actions and inventory controls, and propose a cooperative multi-agent decomposition into a quoter and a risk manager trained using recurrent Independent Proximal Policy Optimisation (IPPO) with RNN policies. We investigate reward shaping, particularly reward blending and post-fill markout, to mitigate credit-assignment difficulties between the agents. Our redesigned single-agent market maker quotes more frequently and incurs smaller held-out losses than the retrained JaxMARL-HFT baseline, although the contributions of jointly changed observations, actions and risk constraints are not separately identified. Decomposition alone performs poorly; careful reward design substantially improves the desk relative to the unshaped split and yields our best result. Models are evaluated on realised end-of-episode portfolio value on held-out markets. None establishes positive average portfolio value in periodic held-out evaluation, and each configuration was trained with one seed. The environment, data and protocol lay a foundation for makers that quote many linked markets at once.
神经与进化计算 (cs.NE)
3
cs.NE / 1 / 2610.25561
Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents
Abstract
Agents that run for a long time have a memory problem, and it is not the one most systems optimize for. Storing every trace and retrieving the nearest neighbors works early and quietly stops working later: the store fills up with failures the agent then repeats, and with facts that were true fifty turns ago. RSPM (Recursive Self-Pruning Memory) borrows the fix biology already found. At intervals the agent sleeps. It clusters its recent failures, distills each cluster into a short rule, and deletes the raw episodes it just learned from. Rules decay unless something reinforces them, and rules that recur across many contexts get promoted into general ones. We tested this on three settings with little in common: a difficulty-calibrated SWE-bench simulation, temporal consistency across 12 conversational benchmarks, and long-horizon QA on LoCoMo-10. All three improve, by 9.1, 11.9, and 7.6 points respectively. More interestingly, the two studies that instrument the store itself converge on nearly the same operating point without being tuned toward it: roughly 93% fewer retry loops and roughly 30% fewer stored entries. That hints that knowing what to forget is a general property of agent memory rather than a benchmark-specific trick.
cs.NE / 2 / 2610.26202
Hybrid metaheuristics for architectural design optimization under small evaluation and wall-clock budgets
Abstract
Architectural and structural design optimization in parametric software affords about a thousand evaluations of an expensive model for problems with tens of variables, a regime for which most metaheuristics were not designed. On the side of artificial intelligence, two hybrid metaheuristics are proposed: a covariance matrix adaptation evolution strategy (CMA-ES) whose sampling is enriched with differential-evolution trial vectors, and a particle swarm with success-history parameter adaptation, differential-evolution injection and local refinement. A budget-aware delay of population-size reduction in adaptive differential evolution is also proposed. Under a hard cap of 1,200 evaluations, all counted, the CMA-ES hybrid ranks first on seven shifted and rotated functions at 10 and 30 variables and beats CMA-ES on six of seven functions at 30 variables. Ablations attribute this gain to the injection and show that a local search pays off only when its result re-enters the population. The delayed schedule improves three adaptive differential-evolution variants in 40 of 51 comparisons, with no loss. On the side of engineering application, in Grasshopper the CMA-ES hybrid and plain CMA-ES are never significantly worse than the optimizers of Galapagos, SilverEye, Opossum and Tunny on travelling-salesman instances and on a built 85-variable free-form shell, where they end 48 % below the best external tool in 15 minutes. On a 22-variable building-performance problem with binary choices, differential evolution leads instead. Shifting and rotating the functions reverses the rank of Harris hawks optimization. Solvers should therefore be selected by problem structure and stopping time, not by a single ranking.
cs.NE / 3 / 2610.26027
A Spiking Ring-Attractor Architecture for Representation and Control of Revolute Robot Joints
Abstract
Robotic systems with revolute joints evolve on circular state spaces, yet most control architectures represent these states using Euclidean variables that ignore this underlying geometry. We propose a fully spiking neural control architecture in which ring attractors encode joint states, enforcing representations consistent with the topology of revolute systems. Homeostatic regulation computes goal-state mismatch directly over these distributed representations, and opponent gain-modulated populations convert this mismatch into controlled phase shifts of the activity bump, implementing set-point regulation as motion on the ring. We show that the bump dynamics follow an SO(2) geometry, and that gain asymmetries induce a signed phase velocity on the ring. To interface distributed activity with scalar regulation signals, we derive an analytic Fourier phase decoder and implement its signed components using push-pull spiking populations. For multi-joint systems, we demonstrate that local ring-based Fourier features compose into product representations that linearly decode SO(3) orientations through forward kinematics. Experiments in NEST validate the full SO(2) regulation mechanism and the spiking product-feature decoder for lift, pitch, and yaw. Robot experiments on a Tiago arm show that decoded ring dynamics generate joint velocity commands, producing directed motion with joint-dependent accuracy. Together, these results establish a neuromorphic control framework in which representation, regulation, and action are unified within a ring-attractor substrate aligned with the geometry of revolute systems.
计算语言学 (cs.CL)
67
cs.CL / 1 / 2610.25440
Design Considerations in Low-Resource Archival OCR: A Case Study in Lakota
Abstract
Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization. Designing effective tools to support these efforts requires taking into account the community-, language-, and document-specific context. We investigate how small ($\leq$3B) vision-language models (VLMs) can be adapted as optical character recognition (OCR) tools for Lakota, a North American Indigenous language. In addition to benchmarking modeling approaches, we inspect four practical facets of low-resource archival OCR that arise in previous work and conversations with community members: constrained training budgets, segmentation noise, scan degradation, and orthographic variance. We demonstrate that fine-tuning VLMs for OCR post-correction on 35 pages of transcribed data decreases character error rate (CER) from 17.1 to 3.9\% on a bilingual corpus, and find that training only with nine pages increases CER by only 1.1\% over this. However, evaluating under realistic automatic segmentation nearly doubles the CER of the best-performing models, stressing the importance of taking realistic use cases into account. We hope this work encourages future research on developing language technologies grounded in the workflows and constraints of ongoing revitalization efforts.
cs.CL / 2 / 2610.25458
Quality-Controlled Reply and Summary Data for Post-Training Small Language Models in AI Call-Screening Agents
Abstract
Small language models have improved to the point where a phone can plausibly screen its own calls, answering on an absent owner's behalf without sending call content off the device, a privacy-respecting alternative to cloud call handling. Realizing that promise takes a model that can do two things well: reply within the owner's authority during the call, and afterward tell the owner what happened and what needs attention. Released small models do neither reliably - they raise unwarranted alarms about ordinary callers and write summaries that drop material conditions or report a caller's claims as fact - and real calls supply no reference replies or summaries to learn from. We construct supervision for both tasks from synthetic calls under an information-boundary constraint: a caller writer works from a private scenario brief, a reply writer sees only the conversation up to the current turn, and a summary writer sees the completed transcript. Separate model-based reviews of the full dialogue and the summary decide which calls enter training, yielding a corpus of 3,395 calls with 15,102 targets (11,707 reply, 3,395 summary). Training a single shared LoRA adapter on both tasks, we evaluate six models (0.6-4B parameters) before and after adaptation on 192 held-out calls. Reply qualification rises from 21-54% to 94-97% and summary qualification from 0-6% to 43-72%; all twelve paired gains (six models x two tasks) have bootstrap intervals above zero, and the summary gains hold in every source family. Because one model both screened the training data and scored the outputs, these rates measure agreement with that evaluator's standard rather than human acceptance, which a planned human study will test. After adaptation, coherence and naturalness failures nearly vanish, leaving grounding and usefulness - faithfully tracking who said what, under which conditions - as the remaining challenge.
cs.CL / 3 / 2610.25497
Self-Distillation for Supervised Fine-Tuning
Abstract
Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert demonstrations into model-native supervision. GCD uses one model as both teacher and student: the teacher conditions on a prompt containing the gold response in context, the student conditions on the original prompt as at inference, and the student is trained to minimize the forward KL divergence to the teacher's next-token distribution at each gold response token. This lets the model use its prior knowledge to teach itself the new task: correctness is anchored by the gold response, while the soft target preserves plausible alternatives. Across various training tasks on Qwen and Llama models, GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.
cs.CL / 4 / 2610.25505
Where Are Language-Specific Features Represented in Multilingual LLMs? Everywhere
Abstract
Previous work has argued that multilingual LLMs encode language-specific features most strongly in their initial and final layers, implying that intermediate layers support language-agnostic representations. This paper reconsiders this finding by performing four different experiments: three that undertake a fine-grained analysis of representations of Portuguese and Spanish, two closely related languages, and one large-scale language identification experiment that covers over 200 languages. The results of these experiments challenge the strict separation between language-specific and language-agnostic layers in multilingual LLMs, using Llama 3.2 1B as a test case. By analyzing and manipulating contextual sentence embeddings, we demonstrate that language identity remains strongly recoverable throughout the model rather than being restricted to initial and final layers. Moreover, we find that the model tends to classify ambiguous strings as closer to Spanish than to Portuguese, revealing a skew in model behavior that differs from human judgments.
cs.CL / 5 / 2610.25549
EmotiQ: A Quantum Residual Specialist Framework for Text Emotion Classification
Abstract
Emotion classification from text is challenging because many affective states are distinguished by subtle linguistic signals and semantically overlapping boundaries. We investigate whether compact quantum modules can serve as targeted residual specialists that assist traditional classifiers on difficult emotion decisions. We introduce EmotiQ, a residual specialist framework for text emotion classification using a frozen transformer classifier as the primary decision maker, identifies challenging emotion boundaries from validation behavior, and trains boundary specialists, including quantum encodings, data re-uploading specialists, prototype and fidelity specialists, and kernel-based specialists. Specialists are permitted to intervene only when they satisfy validation-based admission criteria. We evaluate the framework across six text emotion classification datasets and establish traditional model baselines. Experimental results show that the admitted EmotiQ policy improves mean macro-F1 metrics over the baseline, and achieves positive gains on five of six datasets. Among the quantum components, ZZ interaction feature maps are the most consistently effective, yielding a mean macro-F1 improvement of +0.0030 and positive gains on four datasets. These findings suggest that current quantum NLP methods can be deployed as residual specialists that provide localized improvements under strong classical supervision.
cs.CL / 6 / 2610.25566
Flattening Fantasies: Analyzing Stereotypes in Pornographic Discourse
Abstract
This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world's fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we investigate how nationalities, ethnic groups, and gender identities are portrayed in pornographic discourse, focusing on stereotypes and deviations from implicit norms of pornography. The results show the dominance of American pornography over other groups, which acts as the de facto norm for the adult industry. We identify persistent ethnic stereotypes reflecting colonial power dynamics, religious fetishization, and the sexualization of cultural taboos. The analysis reveals gendered representational differences across cultures, with a prevalent Madonna-Whore dichotomy in the portrayal of women and stereotypical depictions of men in physically dominant roles.
cs.CL / 7 / 2610.25574
Representations in Motion: Layer-wise Rotation and Trajectory Divergence in LLMs
Abstract
Recent work suggests that activation dynamics--how representations change during inference--might be more revealing about language model behavior than static readouts from individual layers. In this work, we investigate the layer-wise evolution of representations of factuality, using true and false statements when the model is prompted to either lie or to be honest. We report a geometric phenomenon in which the factuality directions extracted from the lying and honest conditions appear to rotate away from each other across layers. We find that such layer-wise rotation also occurs in a task that requires the model to take different visual perspectives in the 3D world, suggesting that the rotation phenomenon might be indicative of manipulating conceptual spaces in general, rather than unique to factuality. We then investigate the "real time" transformation of conceptual representations on a trial-to-trial basis. We show that it is possible to make strong predictions about the model's target task and behavior from trajectories alone. In addition, diverging points in single-trial trajectories correspond to the onset of rotation in aggregated-trial analysis. Finally, we show that both rotation and divergent trajectories occur in more naturalistic settings in which models are not overtly prompted to "lie" but rather report false statements in the course of completing a different task.
cs.CL / 8 / 2610.25586
Assertion and Belief: Moore's Paradox in Language Models
Abstract
Moore's paradox \citep{moore1942} concerns sentences such as ``It is raining, but I don't believe that it is,'' which may well be true yet cannot be sensibly asserted. The absurdity lies not in the content but in the act, since to assert is to vouch for one's own belief. This makes the paradox a test of whether language models take the speech act into account when computing meaning. We present the first test of Moore's paradox in LLMs, evaluating 12 models on a curated set of Moorean sentences in varied contexts. We find that models largely recognize the infelicity of Moorean sentences, and that the judgement is not tied to a memorized form and tracks the speech act. Using the Rational Speech Act framework (RSA; \citealt{goodmanfrank2016}), we then test whether these judgements arise from a listener reasoning about a speaker. We find the models' own listeners do not reason about their own speakers, and their reasoning traces suggest that they instead recall knowledge about the paradox and apply rules therein. The models thus appear to have learned the paradox as a fact about sentences, not as a requirement on speakers.
cs.CL / 9 / 2610.25650
Consistency Training Suppresses Cue Following Without Teaching Cue Selectivity
Abstract
Language models often change answers because of a cue, a hint such as "I believe the answer is B" that gives no valid evidence. Consistency training (CT) trains a model to answer the same with and without the cue. Four CT methods apply this idea to output tokens (BCT), residual activations (ACT), attention patterns (AttCT) and sampled-answer rates (RMCT). Prior work suggests some learn different solutions. We train all four on Gemma-2-2B-it and Llama-3.1-8B-Instruct with the same data. We test predictions written in advance and inspect the models with a known circuit that carries cues to the answer. All four make the trained cue less effective, and their fixes share three properties. First, correct cues lose about as much influence as wrong ones. Second, the fix is usually large for cues before the answer options, as in training, and much smaller after them. Llama RMCT depends least on position. Third, the circuit heads still carry most remaining cue-induced errors. Making BCT targets depend on the cue's correctness mostly weakens the fix. CT thus weakens the trained cue without teaching models when a cue is right. Evaluations with only wrong cues at the trained position miss the first two properties.
cs.CL / 10 / 2610.25731
Does Unlearning Survive Collaboration? Robust Unlearning for Shared-Model LLM Teams
Abstract
LLM unlearning is essential for removing sensitive or unwanted knowledge while preserving useful capabilities. As LLM teams are increasingly used to solve complex problems, forgetting must remain effective during collaboration. Despite advances in forgetting effectiveness, efficiency, and robustness to relearning, team deployment remains underexplored. We find that collaboration weakens forgetting: final-answer evaluations miss recoverable evidence in team messages, including complementary clues supporting joint recovery. We establish theoretically and confirm empirically that controlling final answers or individual messages does not ensure low recovery from their combination. To address this challenge, we propose Aggregation-Conditioned Unlearning (ACU). ACU trains a shared model using reconstruction feedback from individual and pooled messages, prioritizes views with higher recovery risk, and uses replacement-based credit assignment while protecting direct and collaborative utility. TOFU evaluates trajectory-level recovery, while WMDP-bio tests native forgetting and retained utility. On TOFU, ACU lowers full-trajectory recovery from 41.8% to 19.3% in ROUGE-L recall and factual recovery from 34.2% to 18.7% versus output-supervised training, with comparable retained-task performance. These results demonstrate that ACU curbs information leakage in LLM teams and generalizes across team sizes and readers.
cs.CL / 11 / 2610.25735
A Calibrated Instrument for Detecting Disposition Drift in Language Models
Abstract
Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy. The safe behavior these tests certify may degrade as models are modified, by topical fine-tuning, new training data, or changes to settings such as the depth of their reasoning. Most work reports such drift as single summary rates scored by another AI model, without the error analysis needed to interpret them. We adapt a calibrated measurement methodology recently developed for assessing how inference optimizations affect the quality of model outputs. Applying it to several misalignment dimensions in two open-weight models, we first establish the instrument's precision, showing that shifts of a few percentage points can be reliably detected. We then induce drift deliberately by fine-tuning on correct and on subtly harmful medical and financial advice; harmful advice induces a far greater shift than correct advice. Most strikingly, a model that reasons before answering can conceal the shift almost entirely in its answers, even though a direct reading of its internal state shows that its disposition has moved. With reasoning switched off, the same model shifts steadily as the share of harmful advice in its training grows. These results demonstrate the importance of testing safety with reasoning turned both on and off.
cs.CL / 12 / 2610.25790
CoMem: Learning to Anticipate Failures from Retrieved Experience
Abstract
Reflection lets LLM agents improve from their own attempts, but the memory it builds is local to a single task, so an agent reflecting on its own tends to revisit the same explanation for failure. Cross-task reflective memory addresses this in two forms: retrieved memory keeps past experience explicit and looks up related cases, which is specific but bounded by the bank, and parametric memory internalizes experience into parameters, which generalizes but is grounded in nothing beyond the task description. When both are used, existing agents combine them additively, feeding their outputs to the solver side by side, so retrieved experience never changes what the learned memory produces. We introduce CoMem (Compounding Memory), which makes retrieved experience an input to the learned memory itself, at training and at inference. The agent builds a bank of its own successful and failed reflections, learns from retrieved cases to write failure hypotheses, candidate explanations of how an attempt is likely to go wrong, and at test time reasons from newly retrieved cases to guide each attempt. Varying the retrieved cases gives the memory different perspectives, yielding reflective diversity that is not merely sampled but also grounded in retrieved experience. Across two backbones and five benchmarks in math, code, and multi-hop QA, CoMem achieves strong overall performance. Additional evaluations show that it generalizes to out-of-distribution tasks and enables smaller models to provide effective guidance to stronger solvers.
cs.CL / 13 / 2610.25816
Modal Sense Disambiguation in Language Models: The Role of Context, Training Domain, and Input Frequency
Abstract
Modals such as may and must can take on different senses depending on discourse context, alternating between root interpretations (e.g., permission, obligation) and epistemic interpretations (e.g., knowledge, belief). While modals are known to be notoriously difficult for children to acquire, limited research exists on whether language models (LMs) can reliably distinguish between different modal senses. We evaluate seven LMs that vary in architecture, size, and training data type using vector-based similarity and surprisal measures across four datasets. Specifically, we investigate whether LMs can leverage contextual information to disambiguate modal senses and whether their performance reflects the distribution of modal senses in their training data. We show that larger LMs trained on diverse text successfully leverage additional contextual information when available. In the absence of context, they tend to favor the dominant sense based on corpus statistics. Smaller developmentally plausible LMs, including those trained on child-directed language (CDL), fail to show consistent sensitivity to modal sense distinctions and do not benefit from context to the same degree.
cs.CL / 14 / 2610.25836
Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining
Abstract
Reinforcement learning with verifiable rewards (RLVR) is widely used to improve reasoning in language models, yet growing evidence suggests that it mostly sharpens what the base model can already do. This raises a practical question: when RL fails, what is the base model missing? We look inside the RL policy's own rollouts and find that failures are localized: whether a rollout fails is strongly predicted by a small set of characteristic n-grams, ranging from reflection phrases such as "let's recheck" to specific operations such as "the inverse of A". These failures persist throughout RL training, although they are not beyond the model's reach: other sampled answers to the same question are often correct. Building on this observation, we use the failure $n$-grams to retrieve text from the pretraining corpus and continue pretraining the base model on it before RL. The resulting corpus has only 35M tokens, 3--4 orders of magnitude smaller than typical mid-training. We evaluate this approach on OLMo2 and Qwen3 models trained with RL on (1) GSM8K; (2) a mixture of math, commonsense, and knowledge questions; and (3) general math and science reasoning. Despite its size, the corpus yields consistent gains: our approach outperforms continued pretraining on the same amount of text sampled from the pretraining corpus across 10 out of 11 settings, including by $2.0$ points on GSM8K when inserted before mid-training. The retrieved text also transfers across model families: text retrieved for OLMo2-1B's failures improves Qwen3-4B on the mixture by $1.9$ points over the same baseline. Together, our results suggest that where RL fails is a cheap and effective signal for selecting pretraining data.
cs.CL / 15 / 2610.25854
Arabic Financial NLP
Abstract
Financial communication connects numerical accounts of economic activity with explanations of performance, responsibility and future expectations. Arabic financial natural language processing offers methods for studying these connections across corporate reports, market disclosures, journalism and banking interactions. Its contribution to digital humanities depends on retaining the relationship between a computational representation and the document from which it was made. This chapter follows that relationship from source collection and Arabic text processing to the analysis of vocabulary, evaluation, entities, causal attribution and multilingual reporting. It then examines what financial summaries and retrieval-grounded answers preserve or lose, and how dialectal intent detection changes the perspective from institutions speaking to institutions listening. Published Arabic resources are discussed in relation to the questions they were designed to answer, with a distinction between financial pretraining corpora, annotated datasets and evaluation benchmarks. A running constructed example connects the methodological discussion without presenting illustrative analysis as empirical findings. The chapter argues for an account of Arabic financial NLP in which linguistic context, documentary provenance and interpretive judgement remain part of the evidence throughout the analytical process.
cs.CL / 16 / 2610.25876
Learning to Reason Efficiently with Interventional Step Credit
Abstract
Large reasoning models often produce lengthy chain-of-thought (CoT) traces mixing useful intermediate steps with unnecessary ones. Recent methods explore step-level signals such as attention mass and step entropy to identify and suppress unnecessary steps. Despite their effectiveness, such signals reflect the model's behavior at a step rather than that step's contribution to the answer. To address this, we propose the interventional step credit (ISC), which takes a causal-intervention view and assesses step importance to the reasoning through targeted parameter-level interventions. Building on ISC, we further propose IGPO, which uses ISC to guide fine-grained advantage rescaling in reinforcement learning, reducing redundant steps while preserving the important ones. Experiments across two model scales and nine benchmarks show that IGPO achieves up to 62% average token reduction and better overall performance among compared methods.
cs.CL / 17 / 2610.25908
CALM: Consistency-Aware Logit Masking in Decoding Space for Mitigating Overthinking of Large Reasoning Models
Abstract
Large reasoning models (LRMs) have achieved strong performance on challenging mathematical and scientific tasks through long chain-of-thought reasoning, but they often suffer from overthinking with unnecessary reasoning steps. Existing truncation-based methods typically detect consecutive trial-answer agreement and force early exit, yet such local stopping signals do not necessarily indicate final-answer stability and can disrupt useful subsequent reflection or verification, potentially leading to performance degradation. We propose consistency-aware logit masking (CALM), a training-free decoding intervention for mitigating overthinking in LRMs. CALM first monitors trial answers at step boundaries and then applies logit masking to action transition point (ATP) tokens such as "Wait", "Alternatively", and "Hmm", suppressing redundant transitional reasoning while allowing the model to continue its thought process more naturally. Experiments across four advanced LRMs and five benchmarks show that CALM reduces token consumption by 18.6%-25.0% compared with full CoT reasoning while maintaining performance, and can even improve accuracy by 0.07%-1.48% by reducing redundant reasoning. LLM-as-a-Judge results also demonstrate that CALM provides more coherent and concise results.
cs.CL / 18 / 2610.25919
LEAPS: Long-context Evaluation of Adaptive Pattern Learning and State Tracking
Abstract
Long-context language models (LCLMs) power agents that act over long trajectories, from improving a codebase to running research experiments to forecasters that condition on lengthy histories. These applications draw on three distinct long-context capabilities: recall, where the answer is a fact present in the context; state tracking, where the outcome can be computed by performing a series of transformations to a specified state; and pattern learning, where the goal is to make a prediction by sampling from a distribution inferred from many observations. Existing benchmarks primarily focus on the first capability. In this work, we introduce LEAPS (Long-context Evaluation of Adaptive Pattern learning and State tracking), a collection of 12 novel datasets, spanning agentic workflows, behavioral modeling, and forecasting of market and world events. By construction, LEAPS is resistant to data contamination, extends to millions of input tokens, and uses reliable metrics. We evaluate 30 API and open models from 8K to 512K input tokens. LEAPS distinguishes models that perform near perfectly on recall tasks, finding that such models' state tracking and pattern learning abilities can vary significantly and degrade much more dramatically with increasing input lengths--even for pattern learning, where more observations should make learning a theoretically easier task. Moreover, thinking models often underperform their base counterparts, exhausting their generation budget or oversimplifying the task. We hope LEAPS can guide the development of the next generation of LCLMs.
cs.CL / 19 / 2610.25942
Speech LLMs That Breathe: Modeling Non-Verbal Vocalizations in Dialogue
Abstract
The user experience of speech interaction models has improved, benefiting from the modeling of paralinguistic attributes and advances in speech reconstruction paradigms. However, human communication follows extensive norms regarding non-verbal vocalizations (NVVs) spanning respiratory, affective, and discourse. The omission of these events imposes inherent limitations on the naturalness of current approaches. In this work, we cover 13 major types of NVVs and construct a natural dialogue dataset that conforms to human conversational norms through an automated synthesis pipeline. We enhance the ability of speech interaction models to generate contextually appropriate NVV events through multi-stage fine-tuning and NVV-aware GRPO (Group Relative Policy Optimization). Based on distinct modeling paradigms, we substantially improve the expressiveness of VocalNet and VITA-Qinyu in empathy interactions and role-play dialogues, and further discuss the architectural necessity for achieving high-quality natural speech generation.
cs.CL / 20 / 2610.25999
TempoMoE: Identifying and Steering Fast-Slow Thinking Experts for Efficient Multimodal Reasoning
Abstract
Multimodal Mixture-of-Experts (MoE) reasoning models rely on long chain-of-thought (CoT) for complex tasks, yet often overthink easy inputs, increasing token cost and latency, and sometimes degrading accuracy. Existing methods optimize reasoning length during training or impose external control at inference, but leave the underlying reasoning mechanism a black box. We find that fast and slow thinking exhibit distinct expert-routing patterns: holding the input fixed and varying only the prompted reasoning tempo produces stable routing shifts. Suppressing slow-thinking experts shortens the reasoning chain, whereas suppressing fast-thinking experts lengthens it, providing causal evidence that these experts influence reasoning depth. Based on this finding, we propose TempoMoE, a training-free, two-stage framework. It identifies fast- and slow-thinking experts via tempo-contrastive routing analysis and redistributes routing weights within the original top-k set according to sample difficulty. It amplifies fast-thinking experts and suppresses slow-thinking experts on easy inputs and reverses this adjustment on hard ones, enabling difficulty-adaptive control over reasoning depth. We evaluate TempoMoE on Kimi-VL, Qwen3-VL, and ERNIE-4.5-VL across six benchmarks. Compared with Vanilla, TempoMoE consistently improves accuracy while reducing output tokens, with accuracy gains of up to 2.6 points and token reductions of up to 35.4%.
cs.CL / 21 / 2610.26006
Scaling Synthetic Pre-training Data and Evaluation for Indic LLMs
Abstract
Building capable language models for India requires scalable multilingual data and robust evaluation. We introduce Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of approximately 2.83 trillion tokens covering all 22 scheduled Indian languages in native script and twelve in romanized form. It is the largest openly released Indic pretraining corpus. Using our GPU-accelerated pipeline, we translate curated English documents, assess translation quality, and repair low-quality outputs while preserving technical content and document structure and maintaining source traceability. Through continued-pretraining experiments across twelve languages, we show that combining our synthetic corpus with natural Indic text improves knowledge and reasoning over a public-corpus baseline. Additional experiments show how data volume, script coverage, and Indic training share affect English performance. We also introduce the Indic-LLM-Eval-Suite, a multilingual benchmark suite for evaluating Indic language models across chat, instruction following, reasoning, knowledge, code, and translation. The suite combines existing Indic benchmarks with task-aware translations of English benchmarks, using automated quality assessment, targeted correction, and native-speaker review in a human-in-the-loop pipeline. We use the suite to comprehensively benchmark contemporary open-weight Indic and multilingual models across capabilities and languages, revealing substantial variation across tasks and languages. Together, these resources provide a scalable foundation for developing and evaluating more capable Indic language models.
cs.CL / 22 / 2610.26008
Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis
Abstract
Every scientific text is embedded in the writing norms of its time: it reflects earlier publications while also shaping the texts that follow. The growing use of LLM-based writing tools introduces a new factor in this evolving system, but the linguistic changes coinciding with their emergence remain insufficiently understood. We investigate whether the trajectories of scientific-preprint writing changed in rate or direction after conversational LLMs became publicly accessible, using the release of ChatGPT in November 2022 as the observational breakpoint. We compare pre- and post-2022 trajectories across vocabulary, punctuation, readability, and syntax in abstracts from arXiv's artificial intelligence, statistics, and quantitative biology categories and from medRxiv. Using interrupted time-series regression, we find substantial post-2022 changes across multiple feature families and research domains, including increases in average word length, reduced formula-based readability, and declines in maximum dependency-tree depth and sentence-level depth variation. Individual features follow different dynamics: some lexical markers accelerate persistently, while others rise and then recede. Our findings show that isolated marker frequencies capture only part of the shift: changes in the rate, direction, and structure of linguistic trajectories provide complementary evidence of broader post-2022 change in scientific-preprint writing. More broadly, these findings motivate monitoring scientific writing as a multidimensional, evolving system rather than through a fixed set of LLM-associated markers.
cs.CL / 23 / 2610.26059
Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness
Abstract
The Chinese idiom "kou mi fu jian" ("honeyed words, hidden daggers") names speech that is polite in form but hostile in function. We present ZhenhuanBench, a 1,592-item benchmark built from expert commentary on Empresses in the Palace, covering speaker intent, emotion-language mismatch, and per-listener uptake. Text-only evaluation covers 34 LLMs. GPT-6 Astra leads the leaderboard at 73.48% overall, while its listener-reaction score is 58.37%. The annotations also supply supervised traces and string-verifiable training rewards. With 338 gold traces, an 8B model outperforms teacher distillation on identical items. Subsequent RL saturates, and chance correction reduces the estimated recoverable performance gap by a factor of 7.5. Our CPR-GRPO loop (Compress-Probe-Replenish) uses this estimate to weight gold supervision, reaching 67.0 on public multiple-choice tasks and 74.5 under the training-matched hidden protocol. Without target-task training, CPR outperforms Base, SFT, and GRPO across six Chinese and English pragmatic and sarcasm evaluations. On three external datasets, training with CPR yields higher mean Macro-F1 than SFT and GRPO.
cs.CL / 24 / 2610.26155
DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia
Abstract
While language models have demonstrated remarkable capabilities, their performance remains disproportionately skewed towards high-resource languages. This disparity is particularly acute for the 700 million inhabitants of Southeast Asia, where vast linguistic diversity remains under-represented in existing benchmarks, often suffering from limited language coverage or narrow task scopes. To address this, we introduce DeepSEA, a comprehensive benchmark constructed from established datasets through professional human translation into 12 Southeast Asian languages, spanning both high-resource (e.g., Indonesian, Thai, Vietnamese) and low-resource (e.g., Burmese, Khmer, Lao) tiers across five distinct tasks: mathematical reasoning, scientific reasoning, question answering, abstractive summarization, and machine translation. Extensive evaluation of frontier proprietary and open-weights models reveals that while frontier models like Gemini-3.7-Flash maintain robustness across resource tiers, smaller models face severe performance degradation on low-resource languages. Crucially, models perform better in reasoning and summarization when pivoted through English rather than target languages, exposing a critical dependency on English representations, especially for low-resource languages. Beyond these observations, we provide a cross-lingual error analysis and a cross-script tokenizer fertility analysis. We hope these findings shed light on the development of next-generation inclusive multilingual models.
cs.CL / 25 / 2610.26168
VISAGENT: An Agent-Based End-to-End Voice Interaction System
Abstract
Voice-based public services have evolved from manual switchboards to intelligent platforms to meet modern user needs. However, current systems remain constrained by rigid menu-driven interaction, fragmented vertical services, and limited extensibility. To address these issues, we present VISAGENT, an agent-based end-to-end voice interaction system for large-scale public services that integrates end-to-end speech interaction, hierarchical agent routing, a scalable voice-agent marketplace, and user memory support. The system replaces menu-driven workflows with natural speech and dynamically dispatches user requests to either a general conversational model or specialized domain agents via a two-level routing paradigm that achieves sub-second intent-to-agent selection over hundreds of agents. User memory informs preference-aware agent selection and supplies relevant preferences and context to selected agents, supporting personalized task execution and reducing the need for users to repeat previously provided information. A voice-agent registration mechanism converts task agents into conversational entities, enabling dynamic onboarding and ecosystem-level extensibility. We validate VISAGENT through a real-world deployment in China's 114 hotline across multiple service domains. Experimental results demonstrate low-latency full-duplex interaction, robust routing scalability, and competitive speech naturalness. More broadly, this work advocates a shift from isolated voice assistants toward agent-centric conversational infrastructures, providing a general blueprint for building extensible voice-agent systems in public services and beyond.
cs.CL / 26 / 2610.26173
Conditional information Retrieval for Longitudinal Survey Harmonisation
Abstract
Identifying conceptually equivalent questions across different surveys is a challenging information retrieval task where surface-level textual similarity is insufficient. Two questions may be lexically similar yet assess entirely different sociological sub-concepts, or lexically different yet conceptually equivalent. We introduce a conditional information retrieval framework that addresses this by decomposing the task into two steps: concept extraction, which identifies the underlying sociological concept of a survey question using a fine-tuned generative model, and concept-aware retrieval, which augments question representations with the extracted concept as an explicit conditioning signal for finer-grained alignment. To reflect the realistic data scarcity conditions of specialised social science annotation, we use CLOSER as a case study. Our framework consistently outperforms baselines by 2-3 points in both MAP and Recall. We further analyse fine-tuning strategies including SFT, DPO, and dynamic DPO, finding that SFT delivers the most stable gains while dynamic DPO interacts differently with instruction-tuned versus distilled reasoning backbones. Finally, we demonstrate that a model fine-tuned on a general social science thesaurus (ELSST) transfers robustly to task-specific survey data, offering a practical solution under data scarcity.
cs.CL / 27 / 2610.26181
Reading Time Modeling with Pixel-Based Language Models
Abstract
Tokenization-free, pixel-based language models (PLMs) have recently been proposed as an alternative to conventional token-based language models (TLMs), typically motivated by practical considerations, e.g., the flexibility to support new languages or vocabularies. Nevertheless, PLMs also have the potential to serve as infrastructure for computational (psycho)linguistics, where tokenization has been a source of bias in estimating information-theoretic values, such as surprisal, and has pushed the bottom-up visual orthographic features outside the scope. In this study, we first integrate PLMs into the modeling of naturalistic human reading and, in particular, examine their use to estimate next-word predictability, which can also integrate orthographic and parafoveal information. Our results demonstrate that next-word prediction difficulty, computed in pixel space, has strong predictive power for gaze duration and word skipping behavior across various languages. In particular, their advantage is complementary to existing token-level surprisal and entropy estimates. Cross-lingual analyses with script variations also clarify their distinctive visual processing biases, facilitating research into how bottom-up visual information is incrementally integrated into expectation-based reading.
cs.CL / 28 / 2610.26190
VEILED-SH: Vulnerability Expression through Indirect Language Evaluation Dataset (Self-Harm)
Abstract
AI systems may miss indirect, culturally situated distress or provide support mismatched to the application context. Existing benchmarks do not jointly address cultural variation, application context, and practitioner disagreement. We introduce VEILED-SH, a benchmark and construction methodology for evaluating single-turn responses to indirect expressions of suicide, self-harm, and distress. Its 10,003 prompts span counselor chat, forum posts, platform search, and triage, drawing on Singapore-grounded sources and public datasets. An expert-validated multi-label taxonomy distinguishes intent, communicative function, and response behavior. Six automated profiles, individually calibrated to three peer-support practitioners and three clinicians, retain separate reference policies for proxy-alignment scoring. Evaluating five systems on a coverage-enriched sample reveals variation in proxy alignment across profile-generated reference policies, even for the same system responses.
cs.CL / 29 / 2610.26225
Rethinking BPE: Merge Selection as Model Selection
Abstract
Subword tokenization methods such as Byte-Pair Encoding (BPE) are a core component of modern language models; yet, tokenizer capacity remains typically determined by a fixed merge budget, chosen heuristically. Since larger vocabularies increase memory usage and computational cost, selecting compact and effective token inventories is essential for efficient language modeling. We study the optimization of a BPE merge budget as a statistical model selection problem and investigate information-theoretic criteria for \emph{computing optimal merge sets} directly from corpus statistics. To further improve the merge selection in limited data conditions, we augment the BPE training corpora with synthetic texts generated from the original data, providing estimates for the unigram and bigram distributions. Experiments across multiple corpora and downstream tasks show that using a principled vocabulary selection technique reduces model size, with little degradation in performance. We further observe that synthetic data can yield better tokenizers, especially in low-resource settings.
cs.CL / 30 / 2610.26233
Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents
Abstract
Tool-using voice agents often leave users waiting after speech ends, even when actionable information arrives earlier in the request. Speculative approaches begin work during speech, but chunk-bounded or interrupt-on-update generation can truncate a tool call before its arguments are complete. This couples computation time to speech updates, limiting useful work completed in advance. We introduce Heartbeat Thinking, which handles updated input through new Thinker invocations while allowing ongoing invocations and dispatched tools to finish. Periodic invocations read the latest partial transcript and a shared record of completed reasoning and tool results. These results help resolve missing arguments and initiate dependent calls, allowing preparation to progress across speech updates. Across SpeechFCEval, FDBench v3, and AudioCRAG, Heartbeat improves aggregate speculative tool coverage or retrieval reuse over a SHANKS-style windowed baseline, with benchmark-dependent effects on final task quality. On FDBench v3, it increases tool coverage by 16 percentage points and reduces mean final-response latency by 36.4\%. The results show that heartbeat thinking paradigm preserving ongoing computation during speech can turn listening time into reusable preparation for faster and more capable voice-agent responses.
cs.CL / 31 / 2610.26236
Shaping the Document, Not the Model: Optimized Suffixes for Context Following
Abstract
When a user supplies a document and wants the answer drawn from it, a language model is expected to follow it even when it contradicts its parametric knowledge. Prior work has focused on making the model rely on context by steering activations, contrasting logits during decoding, or instructing it in plain language. We ask whether the document alone, with no instruction written into it and no access to the model at inference time, is enough. We optimize a document suffix by discrete token search, starting from the answer to a single question and extending the objective to a set of passages, to multi-hop questions that follow from the planted fact, and to later turns of a conversation. On counterfactual passages that contradict parametric knowledge of the model, the suffix flips nearly all items the model resists across three instruction-tuned model families, in most settings outperforms the decoding baseline and the strongest instruction we could write, and matches activation steering. Controls show that the suffix largely acts through the passage rather than injecting the answer, and a single suffix per model transfers to unseen passages. Each extended objective carries the effect to its target, so the objective, not only the document, sets how far the effect reaches.
cs.CL / 32 / 2610.26339
Large-Scale Biomedical RAG: Corpus Scaling, Document Representation, and Efficiency Trade-offs
Abstract
We present a controlled, large-scale evaluation of how corpus scale, document representation, and reranking affect the effectiveness and efficiency of biomedical Retrieval-Augmented Generation (RAG) systems. Retrieval collections were progressively expanded while retaining the relevant source documents, enabling consistent comparisons across corpus sizes. We compared abstract, flat full-text, and hierarchical full-text representations across first-stage retrieval, reranking, and end-to-end multiple-choice question answering. Configurations selected on development data were evaluated on five external biomedical benchmarks and a held-out test set, while retrieval latency and index storage were measured to quantify operational cost. Performance patterns observed with small or controlled collections did not consistently persist as collections grew, and the usefulness of reranking depended on the corpus and retrieval setting. RAG improved three of four model configurations across the external benchmarks, with mean gains of up to 9.6 percentage points, and improved all four configurations on the held-out test set by 10.2-30.3 points. Abstract retrieval performed best when reference evidence was abstract-aligned, whereas full-text representations were generally more effective when richer evidence was needed. Hierarchical representation matched or outperformed flat full-text retrieval in most end-to-end comparisons while requiring 67% fewer indexed units and 44-56% less storage. These findings show that corpus scale should be treated as an explicit evaluation variable in biomedical RAG, and that the most appropriate design depends on evidence format, model behavior, and operational constraints.
cs.CL / 33 / 2610.26397
Bridged Self-Distillation for On-Policy Learning
Abstract
On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities may be incompatible with, or even unavailable to, the student. Existing methods handle these settings through different forms of supervision. To this end, we propose Bridged Self-Distillation (BSD), a unified framework for distilling from any teacher that provides reasoning traces. The model to be distilled from, termed the external teacher, first provides a reasoning trace as privileged context, which can be generated in an offline fashion. The student generates rollouts without this context, while a copy of the student conditioned on it, serving as the self-teacher, scores the same rollouts in the student's own token space. BSD uses the resulting token-level guidance to adjust updates on unsuccessful rollouts, with a verifier providing the outcome signal. As the external teacher is queried only to produce text, the same training procedure applies to all teacher scenarios: shared-tokenizer, cross-tokenizer, and closed API teachers. Across these scenarios, BSD improves over the base model and outperforms baselines tailored to each scenario, offering a novel and unified pathway for distillation. Code is available: https://github.com/TUMLegalTech/Bridged-Self-Distillation
cs.CL / 34 / 2610.26399
How Benchmark Graph Construction Shapes Multi-Hop KG Navigation: A Controlled Audit of PathQuestion and PathQuestion-Large
Abstract
Multi-hop knowledge-graph question answering (KGQA) benchmarks are often treated as fixed language-and-reasoning tasks, but for navigation-based methods, difficulty also depends on the graph in which those tasks are instantiated. This makes graph construction a consequential but often entangled part of the navigation task. To isolate this effect, we audit PathQuestion (PQ) and PathQuestion-Large (PQL) as graph-navigation benchmarks under a controlled intervention: we hold questions, topic entities, released answers, released path annotations, and family-level splits fixed while replacing each released task-filtered graph with the larger source KG distributed with the benchmark. Source restoration expands the effective navigation search space and graph-consistent answer and path multiplicity while sharply reducing unguided reachability to released answers. We evaluate whether these structural changes persist across navigation policies using structural references, frozen small-language-model navigation policies, and trained MINERVA policies. Across both frozen and trained policies, restoration substantially reduces released-answer accuracy and trajectory fidelity, while graph-consistent evaluation recovers only part of the loss. These results show that benchmark graph construction is part of the navigation task specification rather than merely an implementation detail and can materially shape measured behavior even when the released QA task is held fixed.
cs.CL / 35 / 2610.26406
Expert L2 English Speech Evaluation on a Gaming PC: Accuracy, Fluency, and Prosody with XLS-R, XEUS, and PhoneticXeus
Abstract
oollim is a locally executable system for evaluating L2 English pronunciation accuracy, fluency, and prosody on a consumer computer. The system combines pretrained XLS-R, XEUS, and PhoneticXeus representations with reference-text features and ridge regressions trained on SpeechOcean762, the Korean Learners' L2 Speech Corpus (English), J-AESOP, and their combinations. On 305 independently held-out recordings assessed by three expert raters, several ridge configurations substantially outperform archived Azure API scores obtained in January 2026 in both Spearman correlation and quadratic weighted kappa across all three constructs. These results support expert speech evaluation on consumer hardware using pretrained representations and construct-specific regressions, without neural-model fine-tuning or calibration on the private evaluation data.
cs.CL / 36 / 2610.26416
Co-Evolving Policy and Reward for Code Generation: A Mechanistic Study of Single-Adapter Drift
Abstract
Reinforcement learning with verifiable rewards (RLVR) is particularly well suited to code generation because generated programs can be objectively evaluated through execution. Execution also provides behavioral information that can be reused as additional supervision. The Synergistic Policy and Reward Co-Evolving Framework (SPARK), originally proposed for reasoning tasks, exploits this idea by recycling training rollouts and their evaluation results to train a model not only to generate solutions but also to evaluate and improve them. We investigate whether this policy-reward co-evolution approach transfers effectively to code generation and whether execution-grounded reflection provides an additional source of self-improvement. We compare SPARK-style rollout recycling with execution-only Group Relative Policy Optimization (GRPO) using Qwen2.5-Coder-3B-Instruct. Our results do not support either hypothesis under the configuration tested. Although co-evolution improves training performance, these gains do not generalize to held-out problems. Compared at the same stage of training, rollout recycling produces roughly 34x greater KL divergence from the reference policy than execution-only GRPO. Over training, trained tokens per GRPO sequence fall from 185 to 86 as training pass rate increases, while held-out performance declines or fails to sustain early gains. We characterize these patterns as single-adapter policy drift: the shared trainable parameters increasingly depart from pretrained code-generation behavior without corresponding gains in generalization. The results show that higher training reward alone is not evidence of generalizable code self-improvement and highlight the importance of held-out evaluation and drift diagnostics.
cs.CL / 37 / 2610.26421
Self-Consistency of Entity Sets and Citations in a Deep Research System under Repeated Prompting
Abstract
Deep research agents (DRA) are commonly used for open-ended synthesis tasks where closed-world QA benchmarks are poorly suited to characterize relevant aspects of performance. We study the self-consistency of one commercial deep research system under repeated identical prompting, representing each report with the set of named entities it mentions, categorised by a custom schema, and the set of citations it retrieves. Across repeated runs of a single open-ended cybersecurity related prompt, we quantify pairwise overlap to characterize how stable the retrieved and synthesized output structure is under repeated identical prompts. In ten runs, the pairwise entity Jaccard results in a median of 0.12 and the pairwise citation Jaccard is below 0.02, with discovery curves that do not visibly saturate. The magnitude and structure of (dis)agreement between entity sets varies by entity class, which suggests that self-consistency in a DRA is better understood as a function of the underlying entity space than as a stable property of the system.
cs.CL / 38 / 2610.26423
Governing the Rule Space. Ruliology, Computational Irreducibility and Anticipatory Governance of Quantum Technologies
Abstract
The regulation of emerging technologies is often conceived as a strategy of catching up, because legislators arrive too late. Yet gathering signals, constructing scenarios and adapting rules to an anticipated future are fundamental tasks for those entrusted with governance, including the governance of rules themselves. The literature on so-called future law has already shown that anticipation means building institutional capabilities rather than predicting outcomes precisely (anticipatory governance). Taking this premise as its starting point, this article proposes to understand regulation as an exploration of the transformations that may occur. The Ruliological Framework for Anticipatory Regulation distinguishes rules, transformations and trajectories, assessing alternatives and possible developments. Within this approach, computational irreducibility is adopted conditionally, not as a demonstrated property of legal systems. The Quantum Europe Strategy provides a case study and further grounds for reflection.
cs.CL / 39 / 2610.26425
SNO-PI: Label-Scarce SNOMED CT Entity Linking with Pre-trained Encoders and Curated Terminology Knowledge
Abstract
We present SNO-PI, the top-performing submission on SNOMED International's SNOMED CT Entity Linking Benchmark as of 18 October 2026. SNO-PI achieves a macro character IoU of 0.5387, outperforming the organisers' reference system (0.4582; based on the 2024 challenge winner's approach) by 18% relative. Linking clinical mentions to SNOMED CT is a zero- and few-shot problem over a very large label space: the 272 annotated MIMIC-IV discharge summaries cover only 6,595 of the roughly 219,000 eligible concepts, and a third of those occur only once. This motivates a design built on transfer from pre-trained models and human-curated knowledge. SNO-PI combines high-recall span proposal, SapBERT candidate retrieval and listwise ensemble reranking with three BERT-family encoders, each further pre-trained with masked language modelling on unlabelled MIMIC-IV notes and then on SNOMED CT synonyms and hierarchy before task fine-tuning. We find that curated knowledge helps most on concepts never seen in training, yet these concepts remain the main source of error.
cs.CL / 40 / 2610.26426
HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
Abstract
Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses. We identify two failure modes. Under non-discriminative rewards, scores collapse to uniformly low or high values due to insufficient exploration or a mismatch between task difficulty and policy capability. Under spurious rewards, higher scores do not correspond to better responses. Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves. We introduce HarnessRL, a training-time harness that co-evolves with the policy through three interfaces for task sampling, rollout guidance, and rubric criteria. HarnessRL maintains a training-time memory from which a harness evolver retrieves within-task history and cross-task evidence. Using this evidence, the evolver attributes each reward failure to the interface whose update can address it. It then runs a proposer and just-in-time critic loop in which the critic accepts a candidate update only after measuring its effect on the current frozen policy, returning rejected candidates for revision. We also curate Rubric-Mix-10k, a five-domain training and evaluation suite for open-ended RL across medical, science, writing, role-play, and instruction-following tasks. On Rubric-Mix-10k, HarnessRL outperforms the strongest baselines by 2.8 to 3.1 points across both 4B and 9B model scales and nearly halves the training steps to its best checkpoint. Code and data are available at https://github.com/Tongyi-ConvAI/Qwen-Character/tree/main/HarnessRL.
cs.CL / 41 / 2610.26433
The Geometry of Agent Memory: Forecasting Retrieval-Bound Memory Failures, and Characterizing Its Limits
Abstract
We study whether geometric relationships between conversation turns, memory texts, and queries can predict answer correctness before generation. We evaluate four memory implementations with dense or hybrid retrieval and one observation-summary comparator. We embed these texts in a shared space. A logistic classifier uses 25 features describing query proximity, memory-conversation correspondence, and conversation structure. On LongMemEval-small, pooled AUROC is 0.640 over 2,367 question-system pairs, below a question-difficulty baseline of 0.655. Performance varies by question type: geometry reaches 0.841 on single-session-assistant questions, but system identity alone reaches 0.826. After controlling for difficulty and identity, geometry adds no statistically established pooled improvement (delta = +0.002, 95% CI [-0.010, 0.014]). A fixed-store context-selection experiment finds higher accuracy with proximity-ranked memories than with random subsets, consistent with ordinary retrieval relevance. On LoCoMo, nearest-turn similarity has modest predictive value for single-hop questions and near-chance estimates for several other categories. Training uses correctness labels; prediction uses pre-answer features and no additional LLM generation call. The results characterize associations between memory representations and correctness, rather than establish a reliable deployment diagnostic. In particular, embedding proximity does not verify fact coverage or distinguish memory-construction, retrieval, and generation errors.
cs.CL / 42 / 2610.26444
WenSight-Decision-1 for Direct Decisions under E-commerce Rules
Abstract
E-commerce decisions require applying natural-language rules to the available case evidence. We introduce WenSight-Decision-1, a shared model that scores supplied options directly from rules and case information for six operational tasks, with examples spanning six languages. We fine-tune Qwen3.5-9B with reference options as supervision, using the same option scores for training and inference. On 1,440 synthetic operational cases reserved for final evaluation, accuracy rises from 61.60% for Qwen3.5-9B to 92.92% for WenSight-Decision-1 under identical inputs and scoring. Joint fine-tuning improves all six tasks, bringing heterogeneous operational judgments under a common decision interface.
cs.CL / 43 / 2610.26448
TeleASR-Amharic: A Benchmark for Robust Amharic ASR in Realistic Telecommunication Conditions
Abstract
The real-world usability of ASR models remains challenging, particularly for low-resource languages, where existing speech resources are often collected in clean studio environments. In this work, we introduce TeleASR-Amharic, a benchmark for evaluating Amharic ASR under various acoustic and telecommunication conditions. We analyze the major degradations observed in real telecommunications recordings and develop a benchmark that covers multiple channel conditions, environmental disturbances, and telecom-domain variations. Our results show that telecommunication degradation can substantially reduce recognition performance, making even high-performing ASR models impractical under severe conditions. For example, the WER of the best-performing model increases from 0.23 to 0.79 under severe telecommunication degradation, corresponding to a relative increase of 243% in recognition errors. We further investigate the effects of different noise and channel distortions and show that scaling training data and applying carefully designed telecommunication-aware augmentation substantially improve robustness. In particular, our best training configuration achieves a 23.6% relative WER reduction under extreme noise conditions. Together with the release of our telecommunications-domain data, these findings provide a reproducible benchmark for measuring progress in Amharic ASR while highlighting open challenges in developing robust speech technologies for real-world deployment.
cs.CL / 44 / 2610.26461
Training Negatives Still Matter After Calibrating Answerability Readouts
Abstract
Answerability readouts predict whether a passage supplies an answer, but whether calibration can compensate for removing negative passages during training remains unclear. We present a controlled study of how training-negative construction affects calibrated prediction, ranking, and downstream answer retention. We compare native training, which retains negative passages, with empty training, which removes them, holding questions, labels, positive examples, and model weights fixed. Both are calibrated on the same labeled native examples. On SQuAD 2.0 and RepLiQA, native-trained logistic readouts at selected depths achieve lower Brier loss for Qwen2.5-7B-Instruct and Gemma-2-9B-it, with or without output scores. Native fusion also outperforms calibrated output alone. Post-hoc AUROC analysis shows better ranking under native training, beyond what threshold adjustment alone can explain. On the official SQuAD 2.0 development set, native readout policies improve F1 over empty policies but retain more answers to unanswerable questions. The advantage varies by estimator and depth, while RepLiQA's text shortcuts limit claims about contextual understanding. Unfinished generations prevent complete answer-level evaluation, although worst-case bounds preserve the native-empty F1 gains. Our findings show that training negatives affect ranking and downstream decisions even after common calibration, without establishing a universal advantage for native training.
cs.CL / 45 / 2610.26496
Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties
Abstract
Reinforcement learning for language models has moved from preference reward models to verifiable rewards, which sustain much longer optimization. Preference models are prone to reward hacking: the policy raises their score without improving response quality, or even while degrading it. Yet for open-ended tasks, where quality can neither be verified nor fully written down as rubrics, they remain valuable. We introduce Dynalize, which mitigates reward hacking online, during policy training. Periodically and asynchronously, a language model analyzes training rollouts and writes penalties for undesirable behaviors that the reward model appears to reward; a judge applies them to every rollout. In clinical counseling and creative writing, training on the reward model alone raises its score continuously but ultimately leads to a decline in external benchmarks. Dynalize sustains longer training and results in improved downstream performance. For example, we find that our method suppresses patterns incentivized by preference models that, without penalties, spread to most responses, and even to the model's reasoning, which the reward model does not explicitly score. Furthermore, we show that its gains hold in multi-turn training, and even after we retrain the reward model on the discovered behaviors.
cs.CL / 46 / 2610.26614
Conducting Reliable Keyword Estimation with Unreliable ASR in Police Audio Data
Abstract
Extracting insights from large audio data sets using human transcription is expensive and time-consuming. Off-the-shelf Automatic Speech Recognition (ASR) systems offer an easy-to-use scalable alternative, but systematic transcription errors, especially in domain-specific settings, can bias downstream analyses and lead to unreliable conclusions. We investigate whether imperfect ASR transcripts can still be used to obtain statistically valid keyword frequency estimates by combining model outputs with a small set of human-labeled examples, following Prediction Powered Inference and Design-based Supervised Learning frameworks. We evaluate this approach on police radio communications from Baltimore and Chicago, a domain in which keyword distributions could provide valuable insights into police activity, but off-the-shelf ASR systems exhibit systematic errors on operationally important terminology. Using bootstrap-based confidence interval coverage as our evaluation metric, we find that our approach substantially increases the number of domain-specific keywords that can be analyzed with statistically valid confidence intervals, from 43.0% using ASR alone and 27.2% using only a small set of human-labeled transcriptions to 56.2% in the Baltimore dataset, with similar improvements observed in Chicago data. Our results demonstrate that post-hoc error correction can improve statistical validity of downstream analyses even for poor-quality transcriptions, offering an avenue for better leveraging off-the-shelf ASR models for large-scale speech analysis.
cs.CL / 47 / 2610.26655
The Development of Sycophancy in Language Models: Tracing Vulnerabilities from Pretraining to Post-Training
Abstract
Language models can change a committed answer to accommodate a user's opinion or pushback without new evidence. However, when this behaviour develops during training and what a released model inherits from earlier stages remain unclear. We follow 741 checkpoints of 30 open model lineages, from early pretraining to released models. We plant initial verdicts on moral scenarios and measure how far a model follows a user's stated opinion or abandons its verdict under pressure. In most lineages the behaviour is present before post-training, and its average strength changes in either direction, at different stages across lineages. A clearer pattern emerges at the level of individual questions. Beyond responses without pressure, pretraining behaviour predicts little of which questions a released model yields on. That pattern forms during post-training: on routes with supervised fine-tuning, the fine-tuning checkpoint predicts part of it and the preference-optimization checkpoint most of it. A sibling trained from the same base along another route predicts far less: for Tülu-3 8B, its own preference-stage checkpoint removes 85% of the remaining prediction error, a sibling 19%. Internal states add predictive information without changing this ranking. Similar average scores conceal different vulnerabilities, and training history matters for auditing a released model.
cs.CL / 48 / 2610.26695
Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction
Abstract
Hierarchical abstraction is a fundamental principle across many data modalities, enabling complex sequences to be represented compositionally at multiple levels. Learning such hierarchies directly from data, without explicit structural supervision, remains a central challenge in representation learning. We investigate whether meaningful hierarchical structure can emerge from compression alone. While autoencoders learn compressed representations through reconstruction, this objective does not necessarily favour meaningful compositional structure. We study this challenge through unsupervised grammar induction, where the underlying hierarchy can be evaluated explicitly. We introduce a hierarchical token-merging autoencoder that recursively combines adjacent tokens into increasingly abstract representations, compressing entire sequences through a variational information bottleneck. The model jointly learns representations and hierarchical structure by marginalising over possible binary trees using dynamic programming. To address training instability, we combine independently trained models through unsupervised consensus and distillation. Experiments on a synthetic, ambiguous grammar demonstrate the emergence of meaningful hierarchical structure without structural supervision. The resulting model achieves 93% grammatical parse-validity and 93% exact reconstruction on held-out sequences, compared with 42% parse-validity for the strongest trivial baseline. These findings establish a proof of concept for hierarchical token merging as a mechanism for unsupervised structure discovery and abstraction learning.
cs.CL / 49 / 2610.26714
HoloCosm: Beyond Single-Metric Benchmark Compression Across Multiple Metrics
Abstract
Benchmark compression reduces evaluation cost by selecting a small set of samples. Existing methods typically guide selection using a single metric, although benchmarks often assess models along multiple dimensions. We find that model rankings are often weakly correlated across metrics, indicating that different metrics capture distinct aspects of model performance. Samples selected for one metric do not reliably preserve scores and rankings on others when used for compressed evaluation. At 80\% compression, each of the five single-metric baselines yields higher reconstruction error than equal-size random sampling in over 70\% of cross-metric comparisons. To address this challenge, we present HoloCosm, which jointly minimizes normalized score-reconstruction error across source models and metrics while constraining ranking distortion. We evaluate HoloCosm across multiple compression rates on 15 language, multimodal, VLA, and agent benchmark configurations with 85 component metrics. With 80\% compression, HoloCosm outperforms Random, the strongest baseline overall: average NMAE drops by 35.3\% to 0.236, worst-component NMAE drops by 39.5\% to 0.330, and average Spearman correlation rises from 0.884 to 0.915. In addition, HoloCosm remains competitive with existing baselines in single-metric evaluation. We will release our code and experimental results.
cs.CL / 50 / 2610.26754
TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning
Abstract
Expert human readers process large inputs (e.g., code repositories and long documents) by skimming to closely read only what matters. On the other hand, language models process all input tokens through each model layer, resulting in significant processing costs for long-context inputs. We introduce TokenGate, a small, drop-in component that extends existing trained models to enable end-to-end learnable token pruning over long-context inputs for fast and memory-efficient prefill. To enable end-to-end learning despite the non-differentiable top-K operator for token pruning, we introduce token-gated attention, a mechanism inspired by Mixture-of-Experts that re-weights attention logits using token importance scores. Across diverse tasks, TokenGate achieves $1.4$-$11.3\times$ faster time-to-first token (TTFT) and $1.5$-$3.4\times$ lower peak memory at no accuracy degradation for prefill in transformers. Compared to the best-performing prior methods for token-pruning, TokenGate gives up to $6.4\times$ faster prefill and $2\times$ lower peak memory. TokenGate natively extends both standard transformers (e.g., Qwen3) and hybrid-attention models (e.g., Qwen3.5) with simple lightweight adaptation comparable to standard parameter-efficient fine-tuning. We make our code available at: https://github.com/tokengate-models/TokenGate.
cs.CL / 51 / 2610.26776
Monsoon: Scalable Spontaneous Speech Collection for Low-Resource ASR in 50 Languages
Abstract
For most languages, automatic speech recognition (ASR) is limited less by model architecture than by a lack of accurately transcribed spontaneous speech. We present Monsoon, a framework for collecting such data at scale. It elicits unscripted conversations between paired contributors, screens recordings with automated classifiers, corrects transcripts over five levels of human review, and selects segments for lexical and acoustic coverage. It has collected speech in 50 languages; we analyse seven 3,500-hour corpora. At equal volume, its quality stages reduce semantic word error rate (WER) by 10-39% relative to a conventional protocol. Compared with an architecture-matched baseline trained on public corpora, Whisper fine-tuned on Monsoon performs comparably on the baseline's own test sets and reduces semantic WER by 20-44 points on conversational benchmarks. The gains transfer to a second architecture and grow with data volume. We release the framework, its quality-control models, and the fine-tuned models.
cs.CL / 52 / 2610.26782
RESIN: Multi-Chain Reasoning with Entity-Structured Indexing
Abstract
Multi-chain reasoning is a challenging task not only because the multi-chain evidence is dispersed, but also because different parts of reasoning may require distinct, interacting chains of support. Flat retrieval can flood a model with loosely related text, whereas early selection of a single reasoning path can discard evidence that becomes useful only after other facts are considered. We introduce RESIN (Reasoning with Entity Structured Indexing), which uses structure to locate evidence while keeping the original source as the basis for reasoning. RESIN decomposes a question into focused one-hop needs, retrieves a small set of source-grounded units for each hop, and places only this relevant evidence in a compact reasoning buffer. The buffer preserves alternative supports until final composition, making intermediate decisions explicit without forcing a brittle path. On the full HolisQA benchmark, RESIN achieves 82.92 context recall and 76.70 answer relevancy over all questions, with higher reported CR and AR across five domains and higher F1 in four domains than the published baselines. Ablations show that both typed evidence diversity and hop-wise decomposition are essential, supporting a clear conclusion: Precise evidence organization and delayed commitment make multi-chain reasoning effective and auditable. Our code is available at https://github.com/RamonMeng/RESIN.
cs.CL / 53 / 2610.26821
AfriBED: a Speech Recognition Bias Evaluation Dataset for Low Resource African Languages
Abstract
Despite advancements in automatic speech recognition systems (ASR), current systems exhibit speaker-attribute bias, i.e., their speech recognition performance is biased against certain speakers based on their demographic attributes such as gender, accents, and speaker type. However, accurately assessing speaker attribute bias in a given ASR system requires the availability of a bias evaluation dataset with comparable speech across subgroups of a speaker attribute. In this paper, we introduce African speech Bias Evaluation Dataset (AfriBED), a novel speech recognition bias evaluation dataset for two low resource African languages, Bemba and Nyanja, to detect and estimate gender and speaker-type bias for ASR systems. The dataset comprises 9 hours of speech with 3928 utterances recorded by 80 speakers for Bemba, while the Nyanja set includes 6260 utterances with approximately 18 hours of speech recorded by 127 speakers. The results of our baseline experiments demonstrate that the datasets are useful for examining gender and speaker-type attribute bias in the target languages.
cs.CL / 54 / 2610.26824
SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum
Abstract
Matryoshka Representation Learning (MRL) is designed to support flexible representation sizes by jointly optimizing nested prefixes. While effective, this design lacks an explicit constraint to ensure that the leading portion provides a suitable compact encoding. To address this, we propose SpecMRL, which reorganizes the feature space to enable more effective prefix readout. Specifically, we extend MRL with a parameter-derived spectral basis organized with feature importance. SpecMRL otherwise follows standard MRL, with prefix readout performed in the learned spectral space. Experiments on MTEB show that SpecMRL outperforms MRL at all evaluated reduced dimensions, with larger gains under more aggressive compression. These results suggest that the learned spectral basis better prioritizes informative directions, enabling more effective representation compression. Notably, SpecMRL extends the performance frontier of MRL, surpassing full-dimensional MRL with only half the embedding dimensions.
cs.CL / 55 / 2610.26825
MemT-1.0: Towards Memory-Native Evolution
Abstract
Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation models that internalize memory as part of model computation. Building on Metis, MemT-1.0 employs Gated DeltaNet-2 with a stable token-level memory update mechanism within each turn. We further introduce Mixture-of-Storage to scale native memory capacity. Following a deliberate backbone preparation, we develop a two-stage training pipeline consisting of continued pre-training and post-training, supported by 34B tokens of training data and efficient training and inference infrastructure. We release a model family comprising MemT-1.0-Nano and MemT-1.0-Flash, and conduct a comprehensive evaluation of memory and general capabilities. The results show that they outperform the parametric memory baselines on average and remain competitive with retrieval-augmented methods, while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano achieves approximately $1.86\times$ end-to-end inference speedup and a $66\times$ reduction in persistent storage relative to full context KV cache. By internalizing retrieved procedural experience, MemT-1.0 improves task success on agentic tasks, providing a practical step toward memory-native evolution.
cs.CL / 56 / 2610.26826
RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development
Abstract
Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage development, a practical and readily evaluable setting in which existing benchmarks remain largely idealized: they typically focus on reconstructing visual targets or generating webpages from textual instructions alone, overlooking the more common workflow of developing from visual examples. Specifically, we introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. We evaluate the resulting open-ended artifacts using independently verifiable keypoints and complementary programmatic and VLM judges. Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
cs.CL / 57 / 2610.26827
KEEP-Bench: Evaluating Evidence Preservation in Context Compression
Abstract
Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, so these metrics underestimate such loss. To bridge this gap, we introduce KEEP-Bench, a diagnostic benchmark that directly assesses compression faithfulness with 4,102 cases across three challenging settings that require keeping evidence for plausible reference resolutions, explicit qualifiers, and minority perspectives. Experiments show that even with tailored prompting, the strongest models remain unreliable: despite context windows exceeding 100K tokens and inputs averaging only 3.1K tokens, the best open-weight model (31B) preserves all required evidence in only 66.8% of cases, and the best proprietary model in only 61.6%. Performance also varies sharply across evidence structures, so conclusions from limited structures may be biased. We further introduce a metric, Evidence Preservation Return (EPR), to relate evidence recall to inference FLOPs, and find that it peaks at moderate scales (4B-8B), where recall remains low. Additional output budget mainly benefits already strong models and widens capability gaps, yet the stronger preservation of these models comes at a steep computational cost. These findings highlight the importance for future work to investigate compressors with strong information preservation under a small computation budget, rather than relying on a large output budget or model scaling.
cs.CL / 58 / 2610.26830
Competing Learning in Hybrid Architectures
Abstract
Many modern language models have adopted hybrid architectures that incorporate both linear-time recurrence and softmax attention. Modern recurrence layers provide a highly expressive mixing mechanism while ensuring high throughput by compressing contextual information into a fixed-size state, but underperform softmax attention in contextual information retrieval. In this paper, we reveal a Competing Learning phenomenon where different kinds of mixers compete to learn the same function, and when any layer learns a function, it suppresses the corresponding learning signal for other layers. Layer attribution analysis reveals that under typical data distributions, modern recurrence learns retrieval functions faster than softmax attention, causing attention to fail to learn retrieval in most cases. Stopping weight updates in recurrence in the middle of training helps attention learn retrieval. Based on these results, we propose a staged training procedure for pre-training hybrid architectures on natural language. Empirical results on models up to 1B parameters show that our method results in substantial improvement in language modeling loss, common-sense reasoning, as well as a wide range of retrieval-related tasks across different context lengths.
cs.CL / 59 / 2610.25503
From Seeing to Reporting: A Theory of Reliable Visual Set Generation in Vision-Language Models
Abstract
Vision-language models (VLMs) are asked to report what they see: to count objects, box them, parse a screen or pick a click target (GUI grounding). We argue that reporting is part of the visual task, not merely its formatting. In a dense scene with 173 shapes, GPT-6.1 lists 173 records that match every shape, yet the counts it states in the same response sum to 188. From Seeing to Reporting (FSR) is a theory of visual reporting and a set of inference-time procedures that improve the reports of frozen VLMs. It models reporting as a chain from responses to decoded records, constructed sets and task answers, linked by task-compatible matching. The chain gives exact relations for deleting records, averaging their coordinates and switching readouts, and it splits the error of one answer chosen among several calls into missing candidates, lost representatives and wrong choices. These relations guide procedures for counting, detection, document layout, screen parsing and GUI grounding that use the task specification and a shipped configuration, without requiring labels from the target task. On 50 dense synthetic scenes, a one-call reporting-and-decoding protocol cuts per-type count error by 85.2% for GPT-6.1 and 98.8% for Claude Sonnet 5.5. On 600 ScreenSpot-Pro items, a GUI-grounding tier that combines three models lowers the miss rate from 13.67% for GPT-6.1's two-stage zoom to 9.33%, against 12.67% for three such zooms combined by a cluster vote at about the same cost. A plain vote over the same three models misses 12.33%: the other models supply the candidates, and selecting among them turns those candidates into fewer misses. We release the protocols, the stored responses and the analysis code.
cs.CL / 60 / 2610.25964
Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation
Abstract
We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We looked at the evaluation code and found a readout choice that moves F1 on its own. A two-token logit readout decides yes versus no by comparing two vocabulary IDs, 3582 ("yes") and 1217 ("no"). Across 9,000 greedy-decode questions spanning all three splits, neither ID is generated even once. The model emits " Yes" (token 3869) and " No" (token 1939), because SentencePiece prepends a space byte after the prompt suffix "ASSISTANT:". Read with an eight-token lookup that covers the space-prefixed surface forms, LLaVA-1.5-7B scores F1 0.8221, 0.8498, and 0.8713 on the adversarial, popular, and random splits. That 6.13-point adversarial shift exceeds the adversarial gain reported by five of the six papers we surveyed. None of them states its readout, and our reference ran under 4-bit NF4 quantization, so we cannot say how much of any reported gain a matched evaluation would keep. That comparison stays unresolved until readout and precision are matched. On the fixed baseline we run nine inference-time corrections of our own (none beat it) and dissect why the adversarial split resists logit-level intervention. In a 3,000-question probe, wrong predictions draw more image attention than correct ones (Cohen's $d=-0.46$, $p=1.4\times10^{-24}$), the opposite of what grounding-based corrections assume. A four-readout audit across four further VLMs from three tokenizer families confirms the confound is protocol-level, not model-specific: porting the LLaVA-1.5 IDs verbatim drops F1 to 0.00 on the two disjoint-vocabulary models, whereas a tokenizer-derived readout holds between 0.80 and 0.92.
cs.CL / 61 / 2610.26404
LaMER: Language-Conditioned Multi-Expert Routing for Weakly Supervised Referring Expression Comprehension
Abstract
Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and medical image analysis. Under weak supervision, REC is challenged by the visual-linguistic semantic gap and the difficulty of inferring region-text correspondences from coarse image-level signals alone. Moreover, no single visual encoder captures all complementary properties---semantic alignment, fine-grained texture, and boundary awareness---while naively fusing multiple pretrained encoders yields heterogeneous, hard-to-reconcile representations. We propose LaMER, a weakly supervised REC framework that dynamically aggregates four frozen expert encoders through a language-conditioned Cross-Modal Routing Fusion module, adapting expert selection to each referring expression without instance-level bounding-box supervision. A Cell-Text Matching Head with Negative Anchor Augmentation, trained with a multi-objective loss combining contrastive, reconstruction, alignment, and load-balancing terms, further stabilizes optimization. LaMER achieves 77.67\%, 62.81\%, and 69.49\% Acc@IoU\(\geq 0.5\) on RefCOCO, RefCOCO+, and RefCOCOg, surpassing DViN and WeakMCN by up to 16.49 and 14.49 points, with ablations validating each component.
cs.CL / 62 / 2610.26659
SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?
Abstract
Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.
cs.CL / 63 / 2610.26861
Can It Actually Work? Synthesizing Functional CAD Mechanisms
Abstract
While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret underspecified requests to construct and iteratively revise mechanisms that satisfy functional requirements in physical simulation. Verifying these requirements is nontrivial: different candidate mechanisms may realize the same function through entirely different components, controls, and operating sequences. To bridge this gap, we propose FunCAD-Harness, a framework that translates abstract requirements into candidate-specific physical tests and uses simulation feedback to guide design revisions. To systematically evaluate this capability, we introduce FunCADBench, a benchmark spanning 46 mechanism families with interdependent and multi-step operations. Experiments across frontier LLM agents show that FunCAD-Harness exposes subtle mechanical failures invisible to visual inspection, substantially improving functional success rates over standard simulation-augmented baselines. Ultimately, our work marks a step toward CAD agents capable of translating human intent into working mechanisms.
cs.CL / 64 / 2610.26258
AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models
Abstract
Large Audio Language Models (LALMs) reason over speech, environmental sounds, music, and other acoustic information. However, their audio encoders produce dense token sequences that make long-form audio costly to process. Existing audio pruning methods reduce this cost using heuristics such as attention, representation similarity, or temporal sampling, but do not explicitly learn which tokens are most useful for a given question. We propose AudioScorer, a lightweight model that scores audio tokens according to the user question. It is trained by distilling the ranking of audio tokens under the frozen LALM's answer-gradient saliency, together with downstream answer supervision. Compared with other pruning methods, AudioScorer achieves the best or tied-best result at 10% token retention on 7 of 10 semantic and acoustic tasks, including speech recognition, emotion recognition, and audio-scene classification.
cs.CL / 65 / 2610.26446
Label Authority in Speech Emotion Recognition: Intended and Perceived Ground Truth in Emozionalmente
Abstract
Ground truth in Speech Emotion Recognition (SER) is often presented as a single categorical fact, yet an acted emotion and the emotion heard by listeners describe different aspects of the same recording. We examine this distinction in Emozionalmente, an acted Italian corpus that provides both an elicited production label and five listener judgements for every clip. Using frozen XLS-R, HuBERT, and WavLM representations with matched probes, we compare three label authorities: the expressed condition, the listeners' hard plurality, and their complete vote distribution. The results tell a consistent story. Recovering the expressed condition is easier than predicting a tie-broken listener plurality, but that gap compares two different questions rather than revealing a general loss of SER ability. As listener agreement weakens, the plurality becomes an increasingly brittle hard target. Training on the full vote distribution does not substantially change which class wins the model's hard decision, but it produces probability assignments that more closely follow the alternatives selected by listeners, especially for ambiguous clips. No authority therefore replaces the others. Expressed labels describe how speech was produced, pluralities provide an operational single-class decision, and vote distributions preserve perceptual uncertainty. For human-facing Internet of Sounds applications, evaluations should state which authority is being predicted, retain its aggregation rule and agreement information, and use distributional scores when disagreement is part of the target.
cs.CL / 66 / 2610.26412
Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content
Abstract
Much of the manipulation in online influencer content is not false but disproportionate: an accurate fact is framed as historic, urgent, or catastrophic. Fact-checking and clickbait detection are not designed to capture such framing, and little is known about where it originates: with the creator, the topic, the platform, or the audience that replies. We operationalize rhetorical distortion through five literature-grounded dimensions (Significance Inflation, Anxiety Manufacturing, Novelty Claims, Loaded Language, and Temporal Distortion) and measure it with a three-tier pipeline in which bilingual lexical rules propose candidates, negative filters cancel known false-positive contexts, and an LLM verifies or relabels the candidates. Applying the pipeline to 13,060 posts from 330 accounts on six platforms (RSS/newsletters, YouTube, Bluesky, Reddit, Weibo, and X/Twitter), we find that platform is far more strongly associated with account-level distortion than content category (Kruskal-Wallis $\varepsilon^2=0.33$; two-way ANOVA $η_p^2=0.49$ for platform vs. $0.09$ for category). X/Twitter scores highest (DI 7.74 vs. 0.78 for RSS). Classifying tweets and their replies separately shows where this comes from: the reply layer readers see under a post (its top-5 replies) carries 3.5x the distortion of the tweet it answers (7.65 vs. 2.17, $p<10^{-6}$, 36 of 43 accounts), driven by Loaded Language. At equal text length the gap shrinks to 1.5x but remains significant, and creator tweets alone are statistically indistinguishable from YouTube and Bluesky. Two independent human annotators confirm the reply gap, and the platform ordering and the reply gap replicate with GPT-4o as a second classifier. On threaded platforms, measured distortion thus lives largely in the audience layer, which motivates measurement that separates creator and reply text.
cs.CL / 67 / 2610.26797
Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models
Abstract
Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger models, a frozen pretrained backbone and temperature calibration. We analyze each ingredient in Gaussian surrogates where the finite-sample risk can be computed exactly. A student trained on teacher targets inherits the teacher's bias, which more student data does not remove and agreement with the teacher does not reveal. A fixed representation lowers the risk only below an explicit sample-size threshold. In a probit model the exact recalibration map is a temperature scaling, which removes all excess risk only when the teacher's coefficient has the correct direction and otherwise leaves a residual that calibration diagnostics do not detect. Under Luce's choice axiom, candidate scores are coherent across refinements of an event if and only if they aggregate by log-sum-exp. On 231,521 credit card complaints from the Consumer Financial Protection Bureau, where the decision is whether a complaint ends in monetary relief, a student trained on a logistic teacher's probabilities ends at the teacher's decision cost while a student trained on outcomes goes 8% below it, a 30B model read zero-shot remains worse than the teacher after recalibration, and that model departs from the coherence condition by 17.5 decibans. Pretrained representations reduce the labeled data required and calibration corrects the scale of the probabilities, but neither corrects a biased teacher, so the probabilities have to be checked against realized outcomes.
多智能体系统 (cs.MA)
6
cs.MA / 1 / 2610.26272
Bidirectionality is Necessary in Cellular Machines for Exact-Arrival Multi-Agent Path Finding
Abstract
On ranked directed acyclic graphs coupled by vertex conflicts we study exact-arrival multi-agent path finding (departure and arrival pinned in place and time) posed as a distributed output task: every cell must emit its own bit of a valid joint path-support mask. The machines are cellular: one automaton per cell, state factored over cells, communication only along a fixed local fabric, synchronous rounds, a query-independent output round, no central decoder, and a per-topology transition code. "Necessary" here means: on an explicit hard topology, over all its feasible queries, no universally correct local-only architecture can renounce either temporal orientation; with global flags, impossibility becomes a bandwidth-round trade-off. The class is not vacuous: a flooding machine solves every instance in as many rounds as the diameter of the local fabric. The bound survives two concessions: we supply every agent's exact individual cavity, which alone solves the single-agent problem in zero rounds and makes the single-agent direction theorem vacuous; and we allow $k$ one-bit global aggregate flags. On a twelve-cell, two-agent, depth-three topology no forward-only machine with $k=0$ is correct on two feasible queries, for any number of rounds; the time-reversed topology does the same to backward-only machines. Placing $p$ copies side by side gives, for forward-only machines, $J_R+kR\ge p$ at constant depth, where $J_R$ measures how much of the query reaches the targets' cones and is $0$ here. A backward-only machine solves that family in constant rounds and a flag-streaming one in $\lceil p/k\rceil+O(1)$, so the bound is tight in order; on the composite family neither orientation can be renounced. The finite lemmas of the gadget and the pair-invariance contradictions are checked in Lean 4; the semantic bridge, the cone lemma and the counting are paper proofs. We report no benchmarks.
cs.MA / 2 / 2610.25748
MAPS: A Multi-Agent Personalized Script-Driven Video Advertising System Enhanced by Reinforcement Learning
Abstract
With the rapid growth of online video consumption, video advertising has become a dominant component of digital advertising. Most personalized video advertising systems follow a retrieval-based paradigm, which matches ad content with user preferences in a feature space and recommends the most similar candidates. However, since the candidate creatives are pre-produced, this paradigm limits the granularity of personalization and leads to performance disparities across diverse user groups. To address this limitation, we propose a multi-agent personalized script-driven video advertising system (MAPS). We design four collaborative agents that form a closed-loop pipeline of user profile modeling, script generation, video synthesis, and feedback-driven iterative refinement. To train the Scripter Agent, we develop a reinforcement learning framework equipped with a multi-dimensional reward function, which steers the agent toward producing scripts that simultaneously align with user content preferences and target product information. Extensive offline experiments on representative datasets demonstrate that our approach achieves encouraging performance in generating and integrating personalized creatives. Online A/B experiments in a specific scenario in the Kuaishou advertising system show that MAPS improves CTR by 2.87%, CVR by 2.74%, and ARPU by 9.85% over the generative baseline.
cs.MA / 3 / 2610.25508
How Far Can an Agent Push? Trust-Safety Margins in Live Language-Based Economic Games
Abstract
The previous version contained Unicode em dashes. Paste this ASCII-only version: Game theory identifies when a strategic action should be accepted, but not how long a live, language-based counterpart will continue to accept it. We entered the GLEE competition in its final week with one protected agent, Milton, and four instrument agents that ran deterministic, per-game randomized trials of policy changes; Milton received changes only through pre-specified gates or registered, rollback-guarded production rules. In negotiation, learned pricing improves adjusted percentile by 0.071 (n=21,168); in bargaining, a continuation model conditioned on the rejected offer did not improve on a recency-weighted incumbent and was retired. In persuasion, a preregistered five-dose experiment (7,391 assigned games, blinded until a fixed stop) shows that recommending every unit raises sales by 0.194 of the maximum (95% CI [0.147, 0.241]) where the receiver's slack makes obedience rational (52% of the Bayesian benchmark), lowers them where it does not (-0.067), and produces a larger early-to-late purchase decline than honest recommendations (-0.123). A rule fixed before unblinding now routes Milton's dose per cell. Milton stood 11th of 460 agents in bargaining at the fixed stop; we also report the close, where the final day's pool shift moved all five agents' bargaining ratings together while their play was unchanged, and three dispatch defects were found by checking delivered behavior against the design.
cs.MA / 4 / 2610.25966
Behavioral Probes for Information Flow in LLM Swarms
Abstract
Most existing approaches to LLM swarm optimization are utility-driven, focusing on final task utility while leaving the processing, integration, and reliance of information within individual interactions largely implicit. To address this, we study LLM swarm interactions from an information-flow-based view and introduce behavioral probes to quantify information flow patterns. Specifically, we repurpose External Context Score (ECS) and Parametric Knowledge Score (PKS)---originally designed for hallucination analysis---to measure a node's external-context dependence and parametric-knowledge dependence, respectively. Across controlled and realistic swarm settings, we observe consistent patterns: ECS increases with input relevance and distinguishes useful multi-source contributions, while PKS decreases as informative context accumulates. Building on this heuristic interpretation, we demonstrate how these probes can complement utility-driven structure optimization in two controlled applications. Probe-guided node pruning reduces redundant communication and token consumption, and probe-guided reasoning-path search accelerates convergence compared to utility-only baselines. Our empirical results suggest that behavioral probes can make the internal information flow of LLM swarms observable and serve as effective auxiliary signals for structure optimization.
cs.MA / 5 / 2610.26359
Discovery-Probability Graph Planning for Embodied Agents in Tasks with Hidden Cooperation Requirements
Abstract
Embodied agents are increasingly capable of perceiving, planning, and acting autonomously in dynamic environments, demonstrating strong performance on complex cooperative tasks. Recent advances typically assume that cooperative tasks are known in advance. In real-world applications, however, hidden-cooperation tasks are more common: the need for cooperation only becomes apparent after an agent observes and interacts with task-related targets. In this paper, we first present HiddenCoop, a benchmark for effectively evaluating agent cooperation in hidden-cooperation tasks. We further propose Discovery-Probability Graph Planning (DPGP), a method for enabling effective agent cooperation in such tasks. At the core of DPGP is a discovery-probability graph that encodes the probability distribution over the number and locations of task-related targets, facilitating effective and efficient information exchange among agents. A search algorithm is then used to derive action plans that minimize expected costs. Experiments show that DPGP achieves cooperative-task success rates of 99.3% on C-WAH and 58.3% on TDW-MAT, surpassing all compared baselines, while also reducing overall embodied execution time and substantially lowering LLM token usage. Our code is available at https://github.com/hnu-vis/DPGP.
cs.MA / 6 / 2610.26106
A GPU-Resident Pipeline for Multi-Robot 3D Inspection with Limited Perception
Abstract
We present a complete, GPU-resident pipeline for multi-robot 3D inspection of underwater structures. Given a triangle mesh and a fleet of $K$ robots, each carrying a short-range, limited field-of-view camera, the pipeline computes 6-DOF trajectories that jointly cover a target fraction of the surface while minimizing a linearly interpolated makespan/total-cost objective. The joint problem is NP-hard, so we decompose it into five core stages - candidate viewpoint generation, visibility computation, set cover, min-max vehicle routing solved as a mixed integer program (MIP) with lifted Miller-Tucker-Zemlin subtour elimination, and priority-based multi-agent path finding with continuous-time inter-robot conflict checking - flanked by mesh loading and surface sampling as preprocessing and by trajectory smoothing as postprocessing. Our contributions are: an end-to-end GPU-accelerated integration of the five stages, with its kernel design documented; a GPU batch adaptation of Lien's $\varepsilon$-visibility with four custom CUDA kernels; a $β$-aware per-vehicle tour upper bound derived from the interpolated objective, supplementing the Desrochers-Laporte coefficient lift with forbidden-pair cuts; a 4D space-time extension of Zhou and Zeng's GPU parallel-frontier A*; and an experimental evaluation on 14 meshes (up to $7.2$M faces) that ablates the routing cuts and the parallel-frontier search, audits collision freedom and waypoint attainment \emph{after} trajectory smoothing, and measures the end-to-end gap against analytical and LP-relaxation lower bounds. Code and raw results are publicly available.
软件工程 (cs.SE)
6
cs.SE / 1 / 2610.25601
Step-Aligned Script Generation with Verification and Recovery for Web Automation Testing
Abstract
Automated generation of end-to-end web tests remains costly: developers must identify stable selectors, encode multi-step workflows, attach assertions, and maintain scripts as interfaces evolve. Recent LLM-based web agents can complete natural-language tasks in a live browser session, but task completion is not the same as producing a test-ware artifact - code that is executable, replay-able from a reset state, inspectable by developers, and tied to an oracle. We present TESSARA, an adaptive test-script generation framework that converts natural-language test intent into Playwright scripts consisting of executable interaction code, goal-state assertions, and step-level provenance linking each step to its generated code fragment, selector evidence, execution outcome, and repair history. TESSARA grounds each planned step via Step-Aligned Scripting (SAS), enabling fine-grained failure localization, selective regeneration of failing suffixes, and reversibility-aware backtracking that avoids undoing likely commit points (e.g., form submissions). A three-level verification pipeline (syntactic, execution, and semantic/oracle), detects and repairs malformed code, interaction failures, and incorrect landing states. We evaluate TESSARA on three public web benchmarks, measuring script executability, oracle satisfaction, replay stability, and repair efficiency, and show substantial improvements over baselines and underlying VLMs used directly as script generators.
cs.SE / 2 / 2610.25629
Selecting and Composing Agent Design Patterns: Industrial Case Evidence and Controlled Reference-System Experiments
Abstract
Agent pattern catalogs describe architectural alternatives, but engineers also need evidence for selecting and composing them. We extend a cognitive-function--execution-topology framework with an evidence-driven selection procedure and explicit composition contracts. The study combines two industrial case accounts, a contributor's collaboration design, seven practitioner workshops, and executable fault-injection experiments. Three local reference systems use SQLite state, shared artifact files, and a file-backed map-publication surrogate. A 2 x 2 ablation separates function-specific contracts from execution controls across 36 scenario families, with ten input variants per family. Under equal family weighting, composition correctly completes or safely refuses 11 of 12 families in each system, compared with 2, 2, and 1 for the simple baselines. These are constructed challenge-suite results, not production success rates. Recovery without consumer-side verification increases false-success cases in the map surrogate. All composed systems still fail a boundary case in which trusted evidence is wrong. An audit of a frozen six-arm language-model pilot reproduces its scores but confirms an evaluator-dependent ranking reversal. The contribution connects industrial failure, pattern selection, executable treatment, external acceptance, and revision. Selection is demonstrated rather than evaluated against an alternative procedure; catalog admission rests on separate criteria. The results support bounded composition mechanisms, not a claim that the framework improves engineers' design decisions over an equally informative conventional method.
cs.SE / 3 / 2610.25774
What Is Under the Hood? A Cross-Ecosystem Study of Vulnerable Dependencies in Construction Software
Abstract
The construction industry increasingly relies on software built from third-party components, yet the dependency composition and known-vulnerability exposure of construction-related software remain poorly analyzed. In this study, we conduct an empirical, cross-ecosystem investigation of 904 construction-related GitHub repositories spanning six construction functions and a multifunction category. Using commit-pinned repository snapshots and OSV-Scanner, we examine dependency assessability, known-vulnerability prevalence, severity, weakness classifications, and shared upstream exposure. Our results reveal that only 494 repositories (54.6%) completed vulnerability assessment, revealing substantial gaps in dependency visibility. Among successfully scanned repositories, 395 (80.0%) contained at least one dependency version associated with a published vulnerability advisory, and 97.2% of affected repositories contained a High or Critical severity risk. Additionally, Common Weakness Enumeration (CWE) mapping identified recurring resource-consumption, path-traversal, information-exposure, and code-injection weaknesses. Vulnerability burden was uneven, with a median of 50 distinct advisory clusters per affected repository, while just two affected upstream package families collectively reached ~70% of affected repositories. These findings establish an empirical baseline for dependency assessability and known-vulnerability exposure in construction-related open-source software. Our findings underscore the importance of documenting assessment coverage, interpreting advisory matches in their deployment context, and examining shared upstream dependencies when evaluating construction software supply-chain security.
cs.SE / 4 / 2610.25899
SWE-Residual: Evaluating Starting-State Robustness of Coding Agents
Abstract
Software engineering (SWE) benchmarks are the primary way to measure the capabilities of coding agents: an agent is placed in a code repository or machine environment, given a request, and judged by hidden tests. These benchmarks almost always start the agent from a clean state that contains no earlier work on the task. In practice, however, agents often continue work that a developer or another agent has already done and reported as complete, with no indication of what remains incorrect or missing. We introduce SWE-Residual, a benchmark for Starting-State Robustness: the ability to maintain performance on a task when the starting state already contains earlier work on that task. SWE-Residual contains 198 Residual States drawn from 111 tasks in DeepSWE and Terminal-Bench, where the earlier work modifies a code repository or a whole machine environment. Each Residual State is the state a predecessor agent left behind after submitting its work as complete, although the hidden tests still fail. The evaluated agent receives only the original request and this state, without the predecessor's trajectory or any sign that the attempt failed; because the request and hidden tests are unchanged, the starting state is the only difference from the source benchmark. Across five frontier agents, success drops by 20.6--53.5 percentage points on DeepSWE and 20.6--34.9 points on Terminal-Bench relative to a Clean Start, and agent rankings on DeepSWE reverse. Failed continuations show a pattern we call acceptance bias: agents treat the inherited work as the definition of what remains to be done instead of checking it against the request. On DeepSWE, in 58.9--88.2% of failed runs, they submit with exactly the same tests failing, although they pass most of these tests from a Clean Start. SWE-Residual thus exposes a capability gap that clean-start evaluation does not capture.
cs.SE / 5 / 2610.26089
Function Contracts for Bounded Model Checking of C Programs
Abstract
Bounded model checking unrolls a program into a single formula, so a function holding a loop costs the solver a fresh copy of that loop at every call site. Function contracts break that coupling. Verify the body once against its contract, then replace each call by a stub derived from it. CBMC supports this through dynamic frame condition checking, applied at industrial scale, but with no formal account of why the replacement is sound and no characterisation of what the frame check costs. We present the function-contract framework for ESBMC. The clause syntax, both phases, and the first assigns enforcement the tool has had are all introduced here. Its technical core is an observational frame rule: snapshot the locations the contract does not permit the function to write, then compare them at exit. The encoding emits a number of verification conditions set by how many globals the program has, not by how many writes the body performs, while dynamic frame condition checking grows with the write count. That is a property of the encoding rather than a timing, so it moves with neither the machine nor the solver. The rule itself is the classical one; what we add is its bounded form inside a BMC encoding, with the pipeline's algebraic skeleton machine-checked in Lean 4 and the limits of that model stated. We evaluate on synthetic sweeps, SV-COMP tasks, an ANSSI X.509 parser, the FreeRTOS heap_1 allocator, and the ML-KEM reference implementation whose CBMC contracts we reuse unchanged. At the coefficient count FIPS-203 mandates, neither tool finishes on the monolithic route inside ten minutes, and both finish on the modular one. The framework is merged into ESBMC master. Contracts are what is left when unrolling runs out of room.
cs.SE / 6 / 2610.26394
When Coding Agents Outpace Review: Evaluating PRD-Driven Development in Mature Enterprise Repositories
Abstract
Coding agents increasingly undertake development tasks in mature enterprise repositories, yet existing benchmarks offer limited insight into whether they fulfill complete Product Requirement Documents (PRDs) and where delivery breaks down. Issue resolution and development from scratch leave incremental PRD fulfillment underexplored, while isolated tests provide incomplete evidence of business and integration requirements. We introduce PACE, a benchmark of 43 enterprise tasks with 750 expert-audited Code Points. It combines requirement reconstruction, criterion construction, and evidence-based evaluation to measure PRD fulfillment and connect delivery gaps to agent behavior. Across 11 model--scaffold configurations, most scores fall between 52.1 and 78.1; GPT-5.6-Sol achieves the highest complete-delivery rate of 14.0\%. Incomplete implementation and missing integration are the principal bottlenecks: producing individual components does not ensure a complete delivery. Evaluation achieves 95.2\% agreement with expert consensus, with lower case-level variation across judging records than repeated executions. Beyond model comparison, the benchmark provides actionable feedback for harness refinement: diagnosis guides targeted improvements, and successive harness releases raise mean performance from 59.4 to 63.0 using the same model and task suite. Code is available at https://github.com/TianqiLiu1/PACE-Bench.
操作系统 (cs.OS)
1
cs.OS / 1 / 2610.26274
Comments Paper: Counterexample for Schedulability Analysis for Self-Suspending Tasks under EDF-Like Scheduling
Abstract
This paper provides a counterexample of the sufficient schedulability test for self-suspending real-time tasks under EDF-Like scheduling given by Wang et al. in Theorem 11 of IEEE Transactions on Computers, 74(7):2364-2375, 2025. This paper also shows that their necessary condition is unsound.
硬件架构 (cs.AR)
7
cs.AR / 1 / 2610.25498
Cryogenic Recharacterization of a Commercial 22 nm FD-SOI Foundry Standard-Cell Library
Abstract
Wiring and bandwidth bottlenecks between a quantum processor and room-temperature electronics can be reduced by placing digital control at the cryogenic stage, but commercial foundry standard-cell libraries are not specified for operation at 4 K. A cryogenic recharacterization flow is therefore developed for a commercial 22 nm FD-SOI foundry standard-cell library using commercially available cryogenic compact-device models, while retaining the original cell circuits, layouts, logic functions, and implementation semantics and remaining compatible with the commercial synthesis, place-and-route, and signoff methodology used at room temperature. Body-bias dependence is quantified over 38,640 matched slew/load points at zero, +/-0.9 V, and +/-1.8 V body bias. Relative to full bias, zero-bias operation increases the median rising and falling delays by 38.6x and 33.0x, while approximately 90% of transitions exceed the foundry-defined maximum-transition envelope. The 0.8 V-to-0.4 V level shifter is even more sensitive to body bias: its worst-case falling delay decreases from 30.6 ns at zero bias to 0.56 ns at +/-1.8 V. For the outbound 0.4 V-to-0.8 V crossing, driving the 0.8 V receiver without a level shifter dissipates 57.7 uW of static power per bit at full bias, against 0.196 nW when level-shifted. In post-route analysis of the unchanged extracted physical database of a fabricated 32-bit ALU, the register-to-register setup boundary increases from 1.3 MHz at zero bias to 37.9 MHz at +/-0.9 V and 132.9 MHz at +/-1.8 V.
cs.AR / 2 / 2610.25528
Heuristic-Driven Resource Optimization for Decoupled Memory Channels on FPGAs
Abstract
Determining OpenCL channel depths for Decoupled Access-Execute (DAE) architectures in High-Level Synthesis (HLS) is critical for balancing performance and hardware efficiency. Conventional static analysis tools often over-provision these memory channels to ensure performance safety, leading to significant resource stranding. This paper presents an expansion of an LLVM-based static analysis framework, originally proposed for automated DAE generation, by introducing a hardware-aware refinement layer. We identify a counter-intuitive Performance Inversion Paradox linked to baseline memory stalls, where high-stall, grid-based kernels reach peak performance at minimal channel depths and suffer up to a 61% throughput degradation when depths are increased. We propose a Heuristic Refinement Layer that optimizes OpenCL channel depths based on this correlation between baseline stalls and performance saturation. Experimental results demonstrate that our expanded framework successfully identifies the Minimal Peak Performance point across various benchmarks. By avoiding the performance traps of standard over-provisioning, we achieve a 2.52x improvement in Resource Efficiency for an illustrative HotSpot application. This approach maintains peak execution speedup while ensuring the leanest possible hardware profile, outperforming conservative tool-based static predictions.
cs.AR / 3 / 2610.26083
MX-Like Tensor Core for LLM Inference with NPU Integration via Hardware-Software Co-Design Exploration
Abstract
Block-scaled quantization formats such as OCP Microscaling (MX) and NVFP4 have been widely adopted as efficient data formats for LLM inference. However, the design space of these formats spanned by shared scale, weight, and activation format combinations has not been systematically explored, and the joint accuracy-silicon tradeoff across this space remains uncharacterized, leaving the true Pareto-optimal configurations for NPU-based edge LLM deployment an open question. To bridge this gap, we conduct a joint hardware-software (HW-SW) co-design exploration of the above underexplored MX-like design space with three innovations. 1) We define and systematically characterize the full 8-bit scale format family within the MX-like design space, jointly with asymmetric weight-activation precision, identifying configurations that outperform the MX and NVFP4-like (without per-tensor scaling, PTS) baselines. 2) We build a per-format specialized MAC datapath template that enables fair hardware cost evaluation across the space. 3) We develop an automated co-design flow, integrated into the open-source SNAX NPU platform, that enables joint HW-SW analysis across the design space. Overall, our co-design exploration reveals that scale format matters only under aggressive FP4 quantization, yielding a case-oriented design guideline for both scale format and asymmetric weight-activation precision. We further identify UE5M3 as a favorable accuracy-hardware tradeoff for the FP4 regime, reducing perplexity degradation by up to 25.3% relative to the NVFP4-like baseline, at a cost of only +0.06% area and +0.49% power overhead.
cs.AR / 4 / 2610.26432
E-Layout: Layout-Stage Logic Restructuring via RL-Optimized Equality-Saturated Graph Retrieval-Augmented Generation
Abstract
Achieving high sign-off performance in advanced ICs increasingly depends on layout-stage optimization under realistic physical conditions. While recent efforts have moved toward synthesis-layout fusion, most existing works still focus on bringing estimated physical information earlier into synthesis, whose optimization does not fully correlate with real layout-stage behavior. Meanwhile, layout-stage optimization remains largely restricted to a fixed logic structure, relying mainly on buffering and sizing. This fundamentally constrains the optimization space under realistic physical conditions. To address this, we present E-Layout, an ML-driven framework for post-placement layout restructuring. E-Layout jointly optimizes logic transformation and physical implementation across multiple subcircuits under design-level timing feedback. This enables the selection of the most timing-beneficial set of disjoint restructured subcircuits to form the optimized circuit. Specifically, it first constructs equivalence-preserving logic alternatives via e-graph equality saturation, then couples logic selection with physical realization through a graph-based retrieval-augmented generation (RAG) model, and finally refines these local logic-physical decisions with reinforcement learning under design-level timing feedback. Evaluated on 9 diverse designs, it achieves on average 6% TNS improvement and 2% WNS improvement, while even reducing area by 1%, over an industry-leading commercial physical design flow.
cs.AR / 5 / 2610.26552
A 105 MS/s 4.18 nJ/inference 0.765 $\text{mm}^\text{2}$ Mixed-Signal Streaming ML Engine for Wireless Baseband in 12LP+ FinFET
Abstract
We present SIXPPAC, the first streaming machine-learning engine for real-time integrated sensing and communication (ISAC) in wireless systems. SIXPPAC integrates two SAR ADCs and a digital binary-weight six-layer neural network (NN). The 0.765 mm$^\text{2}$ 12 nm FinFET engine stores all NN parameters on chip, enabling sustained 105 MS/s processing of 8b I/Q baseband samples at 0.8 V. By deploying bipolar inner products using AND instead of XNOR gates, SIXPPAC consumes 4.18 nJ/inference (0.42 POPS/W) at 0.8 V. We evaluate SIXPPAC on three example applications to demonstrate its versatility.
cs.AR / 6 / 2610.26724
Injection of Space Radiation Errors on Low-power NVM using Generative Deep Learning
Abstract
We develop an error model for a Non-Volatile Memory (NVM) device that represents the effects of physical faults experienced in a CubeSat mission in Low Earth Orbit (LEO). The CubeSat carries an off-the-shelf low-power Ferroelectric RAM (FRAM) chip and runs repeated read/write tests according to a predefined pattern. We compare every memory snapshot received from the satellite with a reference bitmap to identify bit errors. We visually analyze erroneous values distributions in the memory snapshots and we understand that they are frequently distributed according to regular spatial patterns. We define a generative deep learning technique to automatically learn these spatial patterns and to synthetically generate additional error patterns mimicking the observed ones. The model is capable at producing memory errors for general application-level experimentation. Our preliminary validation indicates that when comparing the model output with the data we collect, 0.92 of the generated bit flips correspond to bit flips in the memory snapshots and the model captures 0.91 of the observed bit flips. A comparison of the spatial distribution of bit flips in 10,000 generated error patterns also provides a Wasserstein distance of 0.073, indicating that the two distributions differ by roughly 7% of the maximum possible displacement of their probability masses.
cs.AR / 7 / 2610.26599
A Type System for Communication-Safe Hardware Design
Abstract
Hardware components communicate through clock-cycle-sensitive protocols. Hardware Description Languages (HDLs), however, provide limited support for expressing these protocols or for checking that independently developed components agree on them. When components disagree, the composed design can perform unsynchronized exchanges or stop making progress. We call these failures communication-safety violations. Communication safety is therefore a global property, determined by interactions between multiple components. This makes it difficult to check modularly. This paper presents Pact, a typed extension of Anvil for modular verification of communication safety in synchronous hardware. Pact types describe message-passing, control-flow, and clock-cycle behaviour. Each component is checked against its type. Compatible types then guarantee that the components compose into a communication-safe system. Pact also supports incremental design. It can check a subsystem before the full system context is available. The check produces a summary type that captures the communication behaviour left for future compositions. To reduce annotation burden, Pact infers types by abstracting payload values while preserving communication, timing, control, and fairness. These inferred types are optimal for the checker: whenever some checker-accepted types can prove compatibility, the inferred types can prove it as well. Type inference and type checking are decidable, and compatibility is decided by a sound and complete algorithm. We implement Pact by extending the Anvil compiler and evaluate it along four dimensions: detecting communication errors during incremental design, expressiveness on ten clock-cycle-accurate translations of open-source hardware IPs, compatibility-checking time, and generated hardware overhead. Pact expresses realistic protocols and verifies all ten benchmarks in reasonable time.
密码学与安全 (cs.CR)
23
cs.CR / 1 / 2610.25636
Formalization of Rényi Differential Privacy in Isabelle/HOL (Draft version)
Abstract
Differential privacy (DP) is a gold-standard definition of privacy designed to evaluate the privacy guarantee of databases. Rényi differential privacy (RDP) is an important variant of DP that gives tight evaluations of machine learning algorithms. These formulations are easy to understand, but evaluating privacy for actual databases is complicated. Even a small change in the database can break its privacy, and evaluations of DP and RDP themselves are complicated. In particular, RDP evaluations in recent studies have become much harder. Therefore, formal verification of them has been studied for over a decade. In this paper, we propose an Isabelle/HOL library for formalizing Rényi differential privacy in a general setting. We first formalize RDP for continuous distributions. As a technical basis, we formalize several properties of convex functions and $f$-divergences for nonnegative convex functions. We formalize basic properties of RDP, conversion laws from RDP to DP, the moment accountants in DP, and the Gaussian mechanism and its RDP. Finally, we discuss the stochastic gradient descent mechanism with Gaussian noise and formalize a rough evaluation of its RDP and DP.
cs.CR / 2 / 2610.25714
Peer-Relative Trust-Guided Update Screening for Robust Semi-Decentralized Federated Learning under Label-Flipping Attacks
Abstract
Trust and reputation mechanisms can reduce the influence of malicious clients in federated learning (FL), but strict update filtering is difficult under non-independent and identically distributed (non-IID) data because honest local updates may temporarily reduce performance on a shared validation proxy. Building on our earlier trust-based semi-decentralized FL (SDFL) framework, this paper operationalizes the trust signals and introduces peer-relative update screening that combines a robust validation-utility score with direction- and norm-based update quality. The method is evaluated with 50 MNIST clients, 20% persistent label-flip attackers, Dirichlet heterogeneity levels of 0.5 and 0.1, 100 communication rounds, and five paired random seeds. Under moderate heterogeneity, the proposed method reaches 98.37% final accuracy while rejecting 95.82% of malicious updates and retaining 90.30% of benign updates; the final benign and malicious trust distributions are completely separated (KS=1.00). Under severe heterogeneity, it reaches 95.25% accuracy, rejects 88.55% of malicious updates, and retains 76.90% of benign updates (KS=0.96). Trust weighting alone preserves slightly higher utility in the most heterogeneous condition, but accepts all malicious submis-sions. The results therefore distinguish influence mitigation from explicit containment and quantify the utility-selectivity trade-off introduced by screening. The study provides an empirically validated off-chain governance core that can feed blockchain-based audit and incentive layers in SDFL.
cs.CR / 3 / 2610.25799
PQC Migration of IPsec: Reliable Composite Key Sharing Mechanisms in IKEv2
Abstract
The threat of quantum computers breaking classical cryptographic is accelerating the adoption of quantum- resistant cryptographic algorithms. Several government agencies have set a tight time-line for upgradation of all system to support post quantum cryptographs. We study post-quantum migration of IPSec protocols and their impact on network performance. The large key lengths in Post-Quantum Cryptography (PQC) cannot fit in a single IP packet leading to IP fragmentation. This can cause failure of security associations and degrade performance of IKEv2 under poor network conditions. To tackle this problem, we study a new key- sharing mechanism for IKEv2 and evaluate its performance. Unlike current hybrid key-sharing approaches, which send the classical and quantum keys sequentially over UDP or TCP (as in RFC 9370), the new approach concatenates them into a single composite key and transport them over TCP to ensure a reliable handshake. We implement and benchmark the new composite key-sharing handshake mechanism against existing hybrid algorithms that use either UDP or TCP. Extensive experiments under various network conditions demonstrate that the new approach significantly improves handshake time and energy consumption by avoiding the severe retransmission penalties of the Hybrid handshake mechanism. At 1% packet loss, the handshake time of composite key sharing mechanism is better by atleast 20% than hybrid mechanisms, and the improve- ment get better with higher packet loss reaching 30% at 10% packet loss. Energy consumption follows the same pattern achieving reduction of 29% at 10% packet loss
cs.CR / 4 / 2610.25916
Secure Agentic Coding through Counterexample-Grounded Feedback
Abstract
As coding models and agent harnesses become more capable, reliable security verification is becoming a central bottleneck. Improving secure code generation requires a signal whose reported progress remains aligned with actual program behavior under optimization; otherwise, an optimizer can improve the signal's verdict without removing the underlying vulnerability. Existing model-based approaches require models to either infer program behavior from source code or specify correct outcomes for adversarial tests, placing substantial reasoning demands on the security signal. We introduce \falcon{}, a counterexample-grounded verification framework that separates attack exploration from behavioral verification and connects them through execution traces of the program under test. In test-time repair, \falcon{} yields substantially stronger held-out functional-security gains than alternative security signals across secure code generation benchmarks, while preserving or improving functionality. As an RL reward, \falcon{} also outperforms static-analysis and learned rewards on held-out functional-security, even when all policies successfully optimize their respective training rewards. Combining \falcon{} with either proxy reward further improves held-out performance, suggesting that execution-grounded feedback can mitigate reward overoptimization while preserving useful guidance from existing security signals.
cs.CR / 5 / 2610.25982
Auditor-Assisted Summary-Channel Verification for Hosted LLM Identity Substitution
Abstract
Hosted LLM services increasingly expose model identity as a product claim, but external users and auditors often cannot inspect which model configuration served a request. We study an auditor-assisted regime in which a trusted auditor enrolls claimed identities and prepares verification material before deployment, while audit-time verification uses only the claimed identity and released output, without runtime access to model internals, inference logs, routing decisions, or provider metadata. We propose SUMMARK, a summary-channel watermarking framework that uses user-visible reasoning summaries as a keyed audit surface. During setup, SUMMARK constructs identity-specific vocabulary carriers, trains lightweight summary adapters, and calibrates per-identity thresholds. At deployment time, a keyed statistical test assesses whether the released summary is consistent with the claimed identity and, when needed, attributes the most likely enrolled identity. Across model families, scales, and task domains, SUMMARK reliably detects cross-family and cross-scale substitutions at low false-positive rates while preserving task-facing answer quality. We further study FP16->INT8/INT4 precision-consistency diagnostics and robustness under bounded post-processing, rewriting stress tests, and text-only spoofing, revealing both strong performance and clear failure boundaries. Our code is publicly available at https://github.com/trinitr0toluene/Summark.
cs.CR / 6 / 2610.25985
NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents
Abstract
Tool-use agents pay per decision: at each hop a language model re-reads the state and thinks for seconds, though its actions are finite and decisions recur. We present NIOCIS, a routing layer that folds execution logs into a loop-collapsed action graph (forward transitions plus recovery back-edges) answering one typed question: which declared option at a state, plus stop, does the fleet take next? A small model reads a calibrated distribution over them in one pass, emitting no text. On 1,615 production security-operations episodes we measure what the layer replaces, returns and is bounded by. Replaces: the typed encoder matches the logged next hop on 32.2% of the 3,875 held-out decisions (34.9% within the declared options) in 10 ms of encoder time (19.5 ms for the typed-choice form), against 2.2 s per LLM hop; two feature readouts reach 32.8% and 32.3%, an ordering the paired test does not resolve, so the typed layer is chosen for cost, not a separated lead. Tuning: jpt-4b reaches 9.9% zero-shot and 31.9% fine-tuned, a gap an exact paired test does not resolve (p=0.68), at 244 ms. Returns little: the realised hop falls outside the declared options for 7.9% of decisions and outside the model's top twenty for 9.3%; the most confident 15% holds 91.9% accuracy. Bounded: a cross-trace leave-one-out test puts the released state's ceiling at 25.6--44.9% as the state key grows more specific, i.e. repeated states disagree more often than they agree; the option-space ablation attributes the gain to per-node candidate restriction (+3.5 points over a flat catalog) rather than graph traversal (zero within its interval); a gated self-evolution loop lifts every arm above its anchor yet ends below full supervision; a confidence-gated cascade loses agreement when the escalation arm is the zero-shot stand-in; and agreement past a third of the fleet needs a state carrying the plan and the observation.
cs.CR / 7 / 2610.26013
SpanGuard: Decoupling Instruction Localization from Authorization for Lightweight IPI Defense
Abstract
LLM-based guardrails are effective against indirect prompt injection (IPI), but their performance depends heavily on model capability: as the guardrail model becomes smaller or weaker, detection degrades substantially. This raises a natural question: can we retain strong IPI defense while relying only on a small LLM? To understand why, we revisit what an LLM guardrail actually does. Detecting IPI requires simultaneously locating potentially injected instructions within heterogeneous tool outputs and judging whether they are authorized by the user's request. We find that the difficulty for small models arises largely from satisfying both requirements jointly, rather than from either subproblem alone. This suggests a different path to lightweight defense: explicitly decoupling localization from decision-making. In particular, localization needs no strong language-model reasoning, since injected instructions are often semantically discontinuous with their surrounding content and can be partitioned into local regions via lightweight unsupervised segmentation. Once the observation is partitioned this way, the remaining authorization check becomes a localized judgment that a small LLM can handle. Building on this insight, we design a lightweight IPI guardrail that performs attack-agnostic localization followed by block-level authorization judgment with a small language model. By separating where to look from what to decide, our approach substantially reduces the capability required of the guardrail model while preserving strong defense performance. Empirically, our guardrail achieves a 0.32% attack success rate with negligible loss in benign utility while remaining effective with substantially smaller models. These results show that strong IPI defense need not rely on strong LLMs when localization and decision-making are explicitly decoupled.
cs.CR / 8 / 2610.26088
LeakBench: Auditing Reward Hacking through Prediction-Time Leakage in Clinical Modeling Agents
Abstract
Autonomous clinical-modeling agents can optimize predictive performance while violating conditions that make a model usable. We introduce LeakBench, an executable, guideline-derived audit suite that tests agent decisions for prediction-time leakage, calibration, site-level evaluation, support awareness, and decision-curve net benefit. LeakBench combines a synthetic task with known failure structure and two real-data settings: MIMIC-IV (430k admissions) and governed multi-hospital inpatient record front-page (MHFP) data. Our primary analysis uses reward hacking descriptively, without attributing intent: an agent selects prediction-time-invalid inputs because they improve its validation reward. A deterministic optimizer and every evaluated unconstrained local-model configuration select a fatal leaking feature. In fully logged September 2026 A/B runs, DeepSeek Flash and GPT-6-astra each leak in 10/10 unconstrained runs and 0/10 runs with an explicit admission-time constraint. This contrast is version-sensitive: a blinded Claude/Fable-5.1 chat run does not leak in either condition, while explicit-constraint violations appear only in a historical August 2026 slice lacking raw trajectories. Replacing AUROC with AUPRC does not remove leakage selection. Renaming or disguising the leaking inputs changes neither selection nor provenance-based detection; a name-based filter fails. A deterministic provenance verifier-in-loop removes the leak, making the LLM repair loop defense-in-depth. Specific verifier feedback clears fatal leakage in a hosted agent, and a three-seed ablation finds principle-only feedback clears decision-level defects, though these mitigation results are limited. LeakBench treats trustworthy clinical prediction as a property of an agent's modeling process, not merely its score; it is an audit testbed with known-failure controls, not yet an expert-validated benchmark.
cs.CR / 9 / 2610.26174
PhantomTac: Hijacking Robotic Tactile Systems with Deploy-and-forget Tags
Abstract
Tactile perception is crucial for precise robotic manipulation, providing an important foundation for dexterous, closed-loop force control. Among various tactile sensing techniques, magnetic tactile sensors stand out thanks to their fine-grained, three-dimensional force perception capability and strong robustness against visual occlusion. In this study, we discover a physical loophole called PhantomTac in magnetic tactile sensors. The attacker can alter the magnetic field, thus spoofing the force interpreted by magnetic tactile sensors. The attacker can stealthily deploy such an attack by simply attaching a thin (the thickness is around 0.5 mm), battery-free magnetic tag to the surface of an object. Hence, the attack tag can be easily concealed by a barcode sticker. Such an attack scheme is "deploy-and-forget" -- the attacker can deploy the tag and leave. When a robotic hand is picking up an object, it may severely damage the object. Specifically, spoofing the normal force can crush the object, whereas spoofing the shear force can lead to a loose grip, thus dropping the object. Compared with directly sabotaging the robotic hand, PhantomTac is much more concealed as the attacker does not need access to the device. We performed extensive reverse-engineering efforts to characterize the sensing mechanism of commercial-off-the-shelf tactile sensors and proposed a practical tag design. End-to-end evaluations conducted on real robots and robotic arms demonstrate that our attack can bypass existing anti-interference mechanisms and effectively hijack their closed-loop control processes, achieving an attack success rate exceeding 95% when the grasping-position offset is bounded within 8 mm. Finally, based on the limitations of existing anti-magnetic interference mechanisms, we propose a multimodal-consistency defense mechanism to mitigate such physical-layer security threats.
cs.CR / 10 / 2610.26207
Signature Knowledge Shielding: Protecting Antivirus Detection Knowledge through Robust Embedding Matching
Abstract
Antivirus (AV) signatures encode valuable detection knowledge accumulated through expert analysis, yet repeated black-box queries can expose this knowledge for unauthorized reuse and targeted evasion. We propose Signature Knowledge Shielding (SKS), a design paradigm that protects existing AV rules, termed Protected Signatures (PS), through auxiliary Shielding Signatures (SS). SKS uses AV-supervised importance ranking to discover candidate byte blocks, filters and selects them through negative-sample calibration, and combines SS matches with original AV verdicts through logical OR. SS matching introduces additional positive responses and verdict changes during probing, making detection feedback less directly attributable to PS and increasing the difficulty of localizing signature-related content, while preserving original positive verdicts. Within SKS, we integrate five individually tuned non-learning matching methods and further propose Robust Embedding Matching (REM), a lightweight learned matcher trained with random context and byte perturbations. REM provides stronger and more consistent protection against content modifications and scanning-alignment changes. We develop and open-source a ClamAV-based prototype and evaluate it against five probing methods. With an observed additional test false-positive rate below 0.1%, REM raises the mean query count of state-of-the-art probing to as much as 282.57 times the unprotected baseline, while introducing only 0.67% GPU scanning overhead. These results demonstrate that SKS can substantially increase signature-extraction costs with limited impact on detection utility and scanning efficiency.
cs.CR / 11 / 2610.26240
Q-MALDetect: Quantum Image-Based Malware Detection and Adversarial Robustness Analysis
Abstract
Quantum machine-learning (QML) models are increasingly considered for security applications, yet their adversarial robustness in malware analysis remains insufficiently understood. This paper presents Q-MALDetect, a hybrid convolutional and variational-quantum detector for byte-image representations of x86-64 Executable and Linkable Format (ELF) binaries, and evaluates it under feature-space and a range of Grad-CAM-guided, executable-aware evasion attacks. The model achieves 97.2% training accuracy, but its robustness deteriorates as attack strength increases, with the strongest evaluated feature-space attack reaching a 78.7% attack success rate (ASR). Among executable-aware attacks, Grad-CAM-guided no-operation (NOP) insertion produces the highest conditional ASR of 32.9%. Dynamic adversarial training is the most effective evaluated defense, reducing feature-space ASR to as low as 5.0%, executable-aware ASR to 7.7%, and the sequential-attack ASR to 34.0%. In contrast, convolutional-autoencoder preprocessing increases executable-aware ASR to 39.7%, coinciding with a rise in Grad-CAM attribution mass within the mutable .text region from 35.2% to 45.7%. Ablation results reveal no monotonic relationship between qubit count and robustness, while the encoding interface modifies the learned attack surface without neutralizing the strongest attack. These findings demonstrate that the effectiveness of attacks and defenses depends strongly on the perturbation constraints and evaluation conditions. Reliable deployment of QML malware detectors therefore requires assessment against both feature-space and executable-aware attacks rather than reliance on clean accuracy or generic denoising alone.
cs.CR / 12 / 2610.26327
Fortress: Fork Resistant Transparency logs Relying on Social graph Structure
Abstract
Transparency systems -- such as Key Transparency (KT), Certificate Transparency, and binary transparency -- require that all honest participants see the same view of an untrusted server's directory, or, failing that, that the misbehaviour is publicly detected. Every existing design closes this gap with an external trust anchor: a publicly available bulletin board (BB), a blockchain, or an out-of-band (OOB) gossip network among auditors. We observe that transparency systems already have a built-in resource that has so far been overlooked: the social graph induced by the users themselves. We ask whether it can replace the bulletin board or the external gossip channel entirely, and answer in the affirmative (if occasional out-of-band interactions are allowed). We formalise a new primitive, Social-Graph based Forking-Resilient Transparency (sg-FORT), which captures the security guarantees of a BB-free transparency system and we construct Fortress Fork Resistant Transparency logs Relying on Social graph Structure, to our knowledge the first such protocol. In every epoch either all users output the same post, or at least one honest user produces a publicly verifiable Proof of Misbehaviour. Our guarantees hold against an arbitrary malicious server, including a snapshot adversary that may compromise a user and learn its secret state, as long as the server does not compromise close users. In honest executions, OOB interactions occur only when a user resets their secrets, so the total OOB cost approaches the natural lower bound of one call per reset -- dramatically below the full-gossip alternative. We prove security under standard cryptographic assumptions on signatures and dictionary commitment. Under simplifying assumptions (synchronous epochs, always-online users), we view this as a first feasibility result that opens up the design space of social-graph-based transparency systems.
cs.CR / 13 / 2610.26402
Temporal Dynamics and Security Assessment of SRT Protocol: A 96-Day Longitudinal Analysis of Production Streaming Infrastructure
Abstract
This study presents a 96-day longitudinal analysis of Secure Reliable Transport (SRT) protocol behavior in a production streaming environment between Iraq and Turkey, examining 262,672 transport metric samples and 5,033,104 system resource measurements collected between October 2025 and January 2026. The research addresses temporal patterns, system-resource relationships, attack vulnerability, and anomaly detection. Results reveal significant temporal variation: Negative Acknowledgment exhibits 7.6-fold hourly variation and 11.5-fold daily variation (p < 0.001), though with small effect sizes (eta^2 < 0.04), indicating that temporal period explains less than 4 percent of total variance. Central Processing Unit (CPU) utilization emerges as the strongest system-level correlate of transport variation (r = 0.617 with NAK rate). Controlled User Datagram Protocol (UDP) flood experiments at 500 packets per second cause CPU increases of 176 to 806 percent, while SRT transport metrics change by only plus or minus 13 percent demonstrating protocol resilience through the Automatic Repeat Request mechanism. An unexpected anomaly reaching 97.7 percent CPU enabled signature-based differentiation between attack types. A Z-score anomaly detection method outperforms fixed thresholds, detecting all 7 security events. This study contributes the longest temporal characterization of SRT metrics reported to date, the first system-SRT correlation analysis, and the first controlled security assessment of SRT infrastructure. These findings provide actionable guidance for capacity planning, security hardening, and multi-metric monitoring of production SRT deployments.
cs.CR / 14 / 2610.26442
Exact Compilation and Resource Laws for Hidden Integer Feedback
Abstract
Hidden randomized rounding couples the privacy loss of successive integer releases. We study a rational-feedback mechanism under predictable signed-unit participation and full observation of every physical output. A preparation's lifetime connects three results. First, preserving complete carries until an observed echo synchronizes the worlds gives exact finite-law point compilation, with support-width exponent $3C-2$ at every fixed budget $C\ge2$. Second, releasing commitments earlier gives a monotone hierarchy whose exact endpoint, strict gaps and numerical brackets quantify calibration loss. At one fixed discrete Gaussian, a loss allowance of 0.09 requires all four commitments. Third, legal regeneration yields an unrestricted KL rate with a square-root testing remainder and a complete asymptotic frontier under joint physical-time and participation constraints. At fixed rational parameters, finite-window occupation programs approximate that frontier to error $2^{-b}$ in polynomial bit time in $b$. Verified countable-transcript semantics and independently checked finite certificates support these analyses. A matched static workload illustrates the resulting design tradeoff: a structured finite integer matrix improves error and fresh randomness, while the scalar feedback instance retains fewer noise-state and source-table bits.
cs.CR / 15 / 2610.26460
Differential privacy with adjustable preservation of multivariate relationships
Abstract
This paper derives an exact attenuation formula for multivariate cross-moments under differential privacy achieved through randomized response. From a proposed alternative formulation of randomization matrices compliant with differential privacy guarantees we show that, under attribute-level randomization and where the randomization of the whole distribution is characterized by Kronecker products, any cross-moment of the anonymized dataset involving distinct attributes is equal to the original cross-moment multiplied by attenuation parameters, themselves determined by the part of the overall privacy budget allocated to each attribute. This characterization enables explicit and transparent controls of multivariate relationships preservations in differential privacy through privacy budget allocation. A simple numerical illustration demonstrates that identical total budgets can yield substantially different covariance-preservation profiles. The results also make the multiplicative loss of higher-order interactions explicit and thus can support mechanism design guided by analytical priorities.
cs.CR / 16 / 2610.26488
Mal-AAEGAN: Adversarial Autoencoder with Double-Critic WGAN-GP for Malware Generation
Abstract
Machine-learning (ML) malware detectors are known to be vulnerable to adversarial samples. Generative adversarial networks (GANs) such as MalGAN, Mal-LSGAN, and our previous Mal-D2GAN can craft adversarial malware to probe and harden these detectors, yet they operate directly on a high-dimensional, discrete binary feature space, which leads to unstable training, vanishing gradients, and adversarial samples of limited diversity. To address these problems we propose Mal-AAEGAN. First, an adversarial autoencoder (AAE) learns a low-dimensional, regularized and continuous latent representation of the $M$-dimensional binary API feature vector, and the generator performs an additive perturbation in this latent space rather than in the raw binary space. Second, training follows the Wasserstein GAN with gradient penalty (WGAN-GP) scheme using two critics: a mimic-black-box critic that inherits the role of the substitute detector and a naive critic that inherits the role of the additional detector. A decode-binarize-OR pipeline guarantees that the original API calls of the malware are preserved. The model is evaluated on two malware datasets of different input dimensionality ($M\in\{139,215\}$) against eight common ML detectors and compared, under an identical protocol, with MalGAN, Mal-LSGAN and Mal-D2GAN. Mal-AAEGAN trains stably in the continuous latent space and attains an evasion rate on par with or higher than these binary-space baselines across both datasets, while its generator stays responsive to the noise input rather than reducing to a single fixed perturbation per malware.
cs.CR / 17 / 2610.26501
Connection Escrow: Holding AI-Agent Pod Egress in the Kernel Until an Isolated Verifier Rules
Abstract
AI agents in containers act by opening network connections, and the text that drives them can be hostile. Controls placed between the agent and its tools are bypassed by a child process or a direct socket, and their approval is not tied to the connection the kernel later opens. We present connection escrow: every outbound TCP connection from a supervised pod is parked inside the kernel's connect() path, before any packet exists, by a 98-instruction eBPF program on the pod's cgroup. A node interceptor attributes the held socket to its kernel socket cookie, process, executable digest, cgroup and pod identity, parks the first bytes the agent writes, and asks a verifier in a separate container with its own key. The connection is released only on a signed verdict naming that one socket cookie and destination, unexpired and unused; no verdict by the deadline resets it, and a late approval is refused. The agent is unmodified. A 36-check adversarial suite with real cgroups passes on the eBPF hold point (33 of the checks also apply, and pass, on an iptables hold point): refused connections never reach the destination; forged, rebound, replayed, expired and late verdicts are refused; the agent's attempt to reach the verifier is itself escrowed; child-process and renamed-binary bypasses fail; UDP is pinned to the resolver. On a Kubernetes cluster, unmodified pods are held per connection and attributed to their SPIFFE identity, a raw socket that bypasses a sidecar proxy is held, and 600 connections from six pods are gated. The hold adds 4.5 ms median per released connection, of which 24 us is policy. Finally, we gate on request content without terminating TLS: an eBPF uprobe reads the plaintext at the agent's TLS-library boundary (shared OpenSSL, static OpenSSL, Go crypto/tls), a model scores it, and the connection is held on that verdict while TLS stays end-to-end and the enforcer holds no key.
cs.CR / 18 / 2610.26642
DM-RSA-OAEP: Formal Treatment of a Dual-Modulus Fault-Detection Countermeasure Based on Physical Diversity
Abstract
We revisit DM-RSA, an RSA variant that uses two independent moduli, whose original version was broken by Håstad's attack and contained an incorrect compromise-tolerance claim. We give a formal treatment of its repaired variant, DM-RSA-OAEP: (i) a complete IND-CCA2 security proof in the random-oracle model via a two-step hybrid argument, upper-bounding the security of the scheme by that of the weaker underlying RSA-OAEP instance, with an explicit simulation loss; (ii) a fault-detection theorem in an explicit adversary model that covers computation, key, and verification faults, including safe-error attacks, under a clearly stated physical-diversity assumption; (iii) an explicit analysis of why the countermeasure degrades to uselessness when the diversity assumption is violated (common-mode faults on a single chip). We position the scheme against the countermeasure landscape and against the broken dual-modulus literature. The contribution is deliberately framed as a fault countermeasure with a provable detection guarantee, not as a computationally stronger RSA. We conclude with a concrete instantiation and a detailed measurement protocol.
cs.CR / 19 / 2610.26781
Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams
Abstract
Phone scams impose substantial and growing costs on individuals and financial systems worldwide, yet existing countermeasures impose no resource cost on perpetrators. Strategic scam-baiting, in which automated agents engage scammers to deplete their operational capacity, offers a proactive alternative but requires dynamic strategy adaptation that static systems cannot provide. We introduce a game-theoretic framework enabling LLMs to learn effective counter-strategies through in-context experience without fine-tuning. The framework models scammer-baiter interactions as a two-player game with empirically grounded strategies and utility functions, using Nash equilibrium as a formal benchmark for evaluating learned behaviour. Across 300 simulated dialogues and three LLM architectures, adaptive strategy selection achieves 28.9-79.7% closer alignment to Nash equilibrium than non-adaptive baselines. The code and an annotated dataset of 300 dialogues with turn-level strategy labels and utility scores are released at https://github.com/nardinebasta/GameofPhones_/ .
cs.CR / 20 / 2610.26799
Accountable partial understanding: securing the epistemic attack surface of agentic systems
Abstract
Attackers profit from what defenders do not understand. Software security has treated the gap between system behavior and human comprehension as a condition to narrow through analysis and assurance, booking the remainder as accepted risk. Agentic AI systems change that gap: when software reads, writes, explains, and authorizes security-relevant behavior, partial understanding moves faster, grows more opaque, and couples to delegated authority. This paper characterizes the resulting epistemic attack surface and presents a taxonomy of four interacting attack categories, which target the interpretation of intent, the production and consumption of evidence, the scope of delegated authority, and the comprehension of human supervisors. It then defines accountable partial understanding (APU) and specifies six necessary conditions: bounded scope with explicit assumptions, inspectable evidence, assigned responsibility, visible residual unknowns, actionability, and temporal maintenance. It explains why delegating interpretation to software agents differs from distributing knowledge among people and distinguishes APU from risk acceptance, explainable AI, assurance cases, formal verification, and human-in-the-loop oversight. From the conditions, the paper derives five design principles: bounded delegation, decoupled evidence, assumption tracking with invalidation, role-specific accountability views, and intervention-preserving design with residual unknown registers. It introduces two measurable probe dimensions, evidence linkage and invalidation latency, for testing whether accountability mechanisms hold under change and adversarial pressure. Three running examples (an AI coding agent, an autonomous cyber-defense agent, and an AI-mediated insulin delivery update) ground the analysis. Security without full understanding is already here; the task is to make it accountable.
cs.CR / 21 / 2610.25773
When Care Expands and Control Declines: Understanding Digital Privacy in End-of-Life Care
Abstract
Terminally ill patients generate extensive digital health data through hospital portals, physiological monitoring devices, consumer health applications, and end-of-life care platforms. Yet their ability to control this data is systematically undermined by the structural conditions of terminal care. Through semi-structured in-depth interviews with 21 terminally ill patients, 11 family caregivers, and 8 physicians recruited from tertiary hospitals, this study examines the digital privacy circumstances and subjective attitudes of patients in end-of-life contexts. Thematic analysis reveals five structural dimensions of digital privacy challenge: continuous bodily digitization without meaningful informed consent; cognitive decline driving involuntary proxy control of accounts; inferential data generation exceeding patients' epistemic boundaries; dissolution of individual privacy boundaries amid relational network restructuring; and the absence of institutional governance over post-mortem data. Patient privacy attitudes span three profiles: fundamental absence of privacy awareness, sustained active management under severe tool constraints, and perception of problems followed by resigned abandonment. Interviews with caregivers and physicians reveal benevolent substitution in the name of care needs, expansion of access beyond the initial task, and unresolved boundaries under absent institutional guidance. These findings reveal a compounding privacy vulnerability unique to end-of-life contexts and call for privacy-by-design principles, granular permission mechanisms, and institutional governance frameworks tailored to the needs of dying patients.
cs.CR / 22 / 2610.25743
Fundamental Limits of Information-Theoretic Secure Aggregation with User Dropouts in Federated Learning
Abstract
Secure aggregation enables a server to compute the sum of users' gradients or model updates without revealing individual values. With user dropouts, correlated random keys protect local gradients while allowing recovery of the desired gradient sum. We consider three phases: key distribution, masked-gradient upload, and mask elimination. Users establish correlated keys, upload masked gradients, and provide additional messages to eliminate residual aggregate masks. Existing information-theoretic schemes typically design mask-elimination communication for a prescribed dropout threshold and evaluate its cost under worst-case dropouts across the masked-gradient upload and mask-elimination phases. Such designs cannot adjust communication to actual dropout scenarios. Moreover, the randomness and communication costs of establishing correlated keys remain largely unexplored. To address these limitations, we formulate an adaptive secure aggregation framework under the T-colluding security constraint and construct a single scheme that supports all feasible dropout scenarios by adapting its communication to the actual number of dropouts. We establish converse bounds on four resources: key randomness, key-distribution communication, masked-gradient communication, and mask-elimination communication. The latter two bounds are characterized for each realized dropout set in the masked-gradient upload phase. The proposed scheme achieves the optimal masked-gradient communication rate for every feasible dropout scenario and the optimal mask-elimination communication rate in almost all feasible regimes. Its key rate and key-distribution communication rate are each optimal or within a factor of two of their respective information-theoretic optima. These results characterize the fundamental randomness and communication costs of secure aggregation with user dropouts.
cs.CR / 23 / 2610.26054
Blame-aware Recomposition for Secure Compilation Proofs of Low-Level Compartment Isolation
Abstract
The security guarantees of existing formally verified compilers for compartmentalized programs in unsafe languages like C stop at a rather high level, assuming a built-in protection mechanism and a block-based memory model. Verifying a concrete, low-level enforcement mechanism for compartment isolation is still a major challenge for current proof techniques. In particular, Abate et al.'s proof technique fails because one of their proof steps, recomposition, essentially requires hiding all information about the code of each compartment, which is not practically achievable at a low level. We propose an extension of their proof technique to a low-level setting where compartments can fully observe each other's code. To do so, we introduce an overlay semantics that allows us to characterize these observations as a form of undefined behaviour, and we use a novel variant of recomposition that we call blame-aware recomposition, which essentially allows recomposing executions as needed in the original high-level proof technique, even if the low-level context observes the code, because this context is ultimately going to be discarded. We apply this new proof to a compartment isolation mechanism that uses a programmable tag-based reference monitor.