Daily Research Digest
arXiv Papers
2026-10-19
450
Papers
8
Categories
108
Translated
收藏清单 0
精选 · Favorites
108
cs.AI / 1 / 2610.22568
Repairing LLM-Generated Specifications with Deterministic Diagnosis
用确定性诊断修复LLM生成的规范
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used as formalizers: given a reasoning problem expressed in natural language, they produce a formal specification that a symbolic solver can then solve, with guarantees relative to that specification. Many specifications are runnable yet semantically wrong, encoding incorrect constraints, domains, objectives, predicates, or goals, and solver-error feedback dries up once execution succeeds. Purely LLM-based diagnosis, lacking a formal ground truth, often fails to spot these errors, and its ungrounded critiques cause regressions in 11.8% of initially correct specifications in our experiments. We propose DRAGO, a repair framework that replaces this ungrounded critique with deterministic, formalism-specific diagnosis, and retains a rewrite only when the diagnosed error weight strictly decreases under a reference-free acceptance predicate. Across five benchmarks spanning logical reasoning (SMT-LIB2), constraint programming (MiniZinc), and planning (PDDL), DRAGO cuts regressions by nearly an order of magnitude, to 1.3%, while improving answer accuracy by 12.9 percentage points over zero-shot formalization and 8.7 percentage points over LLM-diagnostic repair, averaged across benchmarks. We also uncover two rendering defects in Planetarium that cause 50% (3,086 instances) of its floor-tile split to be inconsistent with the reference formalizations, and release a corrected split.
Chinese Translation
大语言模型（LLM）正日益被用作形式化器：给定一个用自然语言表述的推理问题，它们会生成一个形式规范，随后符号求解器可以求解该规范，并相对于该规范给出保证。许多规范虽然可运行，但在语义上是错误的，它们编码了错误的约束、域、目标、谓词或目的，而且一旦执行成功，求解器错误反馈就会枯竭。纯粹基于LLM的诊断缺乏形式化真值，往往无法发现这些错误，并且其缺乏依据的批评在我们的实验中导致11.8%的初始正确规范出现退化。我们提出DRAGO，一个修复框架，它用确定性的、针对特定形式体系的诊断取代这种缺乏依据的批评，并且仅当在无参考接受谓词下诊断出的错误权重严格下降时才保留重写。在涵盖逻辑推理（SMT-LIB2）、约束编程（MiniZinc）和规划（PDDL）的五个基准上，DRAGO将退化减少了近一个数量级，降至1.3%，同时在基准上平均而言，相比零样本形式化将答案准确率提高了12.9个百分点，相比LLM诊断式修复提高了8.7个百分点。我们还发现了Planetarium中的两个渲染缺陷，这些缺陷导致其floor-tile划分中有50%（3,086个实例）与参考形式化不一致，并发布了一个修正后的划分。
cs.AI / 2 / 2610.22706
Brain-Aligned Multimodal LLMs for Social and Emotional Understanding of Cartoon Movies
面向卡通电影社会与情感理解的脑对齐多模态大语言模型
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Multimodal large language models (MLLMs) offer scalable annotations of social and affective content, but their temporal agreement with human judgments and neural predictive value require further evaluation. We compared diffusion-based and autoregressive MLLMs with continuous human Theory of Mind (ToM) and affective valence ratings across three short cartoon movies. Behavioral analyses assessed temporally aligned Pearson correlations, rating distributions, and the effects of preceding scene descriptions. Ratings from each source were separately used in subject-specific regression models to predict regional fMRI responses. Model ratings captured aspects of human temporal variation despite systematic distributional differences, with generally stronger alignment for valence than ToM. Preceding context improved alignment in some conditions, depending on the model and movie. In \emph{The Present}, more regions showed significant encoding for MLLM ratings than for human ratings, whereas effects were more limited in the other movies. Spatial overlap across rating sources occurred mainly among MLLMs; overlap between human and model ratings was limited and involved primarily visual, dorsal-attention, and posterior default-mode regions. These findings support MLLMs as scalable tools for generating brain-predictive annotations of social and emotional content, while highlighting systematic rating differences and model- and stimulus-dependent alignment with human judgments.
Chinese Translation
多模态大语言模型（MLLMs）能够为社会与情感内容提供可扩展的标注，但其与人类判断的时间一致性以及神经预测价值仍需进一步评估。我们在三部短篇卡通电影中，将基于扩散模型和自回归的 MLLMs 与连续的人类心理理论（Theory of Mind, ToM）评分和情感效价评分进行了比较。行为分析评估了时间对齐的皮尔逊相关、评分分布以及前置场景描述的影响。来自每个来源的评分分别用于受试者特异性回归模型，以预测区域 fMRI 反应。尽管存在系统性的分布差异，模型评分仍捕捉到了人类时间变化的一些方面，且对效价的对齐通常强于对 ToM 的对齐。前置上下文在某些条件下改善了对齐，具体取决于模型和电影。在 \emph{The Present} 中，对 MLLM 评分表现出显著编码的脑区多于对人类评分的脑区，而在其他电影中效应则更为有限。不同评分来源之间的空间重叠主要出现在 MLLMs 之间；人类评分与模型评分之间的重叠有限，且主要涉及视觉、背侧注意和后部默认模式区域。这些发现支持将 MLLMs 作为生成社会与情感内容脑预测标注的可扩展工具，同时凸显了系统性的评分差异以及与人类判断之间依赖于模型和刺激的对齐。
cs.AI / 3 / 2610.22723
LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs
LEMON：面向新大型语言模型零样本不当行为监测的注意力图等变学习
large language model
大语言模型相关
Abstract
Large language models exhibit misbehaviors that hinder deployment in high-stakes settings. One approach to monitor such failures is to use an external LLM as a monitor, but this is prohibitively expensive. A well-known and efficient class of monitors can detect such failures from internal representations (such as activation probes), but are tied to the model they were trained on, requiring costly re-annotation as new LLMs appear. A natural way to avoid this repeated annotation cost is to develop internal-representation based monitors that transfer zero-shot across LLMs. However, such transfer is challenging because internal representations are not naturally aligned across models; in particular, attention maps, known to contain signals of misbehavior, may differ in head function, ordering, and model depth. In this work, we characterize those cross-model attention mismatches as algebraic symmetries and use geometric deep learning principles to build LEMON (Language-model Equivariant MOnitoring Network) an online graph neural network-based monitor, whose predictions are invariant under these symmetries. LEMON processes attention maps token-by-token, as they are produced and outputs a score after every token (or every window of $W$ tokens) at negligible cost. Across 16 LLMs and six datasets, in the zero-shot cross-model setting, LEMON outperforms all baselines on both hallucination and jailbreak monitoring tasks, improving over the strongest by up to 14 ROC-AUC points.
Chinese Translation
大型语言模型会表现出不当行为，这阻碍了它们在高风险场景中的部署。监测此类失败的一种方法是使用外部 LLM 作为监测器，但这种方式成本高得令人望而却步。一类众所周知且高效的监测器可以从内部表示（例如激活探针）中检测此类失败，但它们受限于其所训练的模型，随着新 LLM 的出现需要昂贵的重新标注。避免这种重复标注成本的一种自然方法是开发能够跨 LLM 零样本迁移的基于内部表示的监测器。然而，这种迁移具有挑战性，因为内部表示在不同模型之间并非天然对齐；特别是，已知包含不当行为信号的注意力图可能在头功能、排序和模型深度方面存在差异。在这项工作中，我们将这些跨模型注意力不匹配刻画为代数对称性，并利用几何深度学习原理构建 LEMON（Language-model Equivariant MOnitoring Network），一种基于在线图神经网络的监测器，其预测在这些对称性下保持不变。LEMON 在注意力图生成时逐 token 处理它们，并在每个 token（或每 $W$ 个 token 的窗口）之后以可忽略的成本输出一个分数。在 16 个 LLM 和六个数据集上，在零样本跨模型设置中，LEMON 在幻觉和越狱监测任务上均优于所有基线，相比最强基线最多提升 14 个 ROC-AUC 点。
cs.AI / 4 / 2610.22924
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning
Sep-Latent：分隔符词元作为隐式推理的自然压缩锚点
large language model
大语言模型相关
Abstract
Large Language Models (LLMs) have demonstrated remarkable success in complex reasoning through explicit Chain-of-Thought (CoT), but the associated computational overhead and token verbosity remain significant barriers to efficiency. While implicit reasoning in continuous latent spaces offers a promising alternative, existing methods like Coconut and CODI often rely on fixed-length latent sequences and suffer from representational collapse. In this paper, we propose Sep-Latent, an adaptive implicit reasoning framework that utilizes separator tokens as structural semantic anchors. By introducing an Attention-Decay mechanism, we force the model to aggregate and compress reasoning logic into these anchors during training. We then align explicit and implicit pathways through multi-level distillation across all layers to ensure logical consistency. Crucially, Sep-Latent enables dynamic reasoning depth by adaptively terminating the latent thinking phase based on query complexity. Experimental results on the GSM8k-Aug benchmark show that Sep-Latent achieves an accuracy of 56.8\% with an average of only 5.4 tokens, significantly outperforming prior fixed-length implicit models in efficiency while narrowing the performance gap with explicit CoT. Further analysis via PCA and symbolic variant testing confirms that our method ensures semantic grounding and exhibits superior robustness to perturbations.
Chinese Translation
大型语言模型（LLMs）通过显式思维链（CoT）在复杂推理中取得了显著成功，但相关的计算开销和词元冗长仍然是效率的重要障碍。尽管连续潜在空间中的隐式推理提供了一种有前景的替代方案，但 Coconut 和 CODI 等现有方法通常依赖固定长度的潜在序列，并遭受表征崩溃问题。本文提出 Sep-Latent，一种利用分隔符词元作为结构化语义锚点的自适应隐式推理框架。通过引入注意力衰减（Attention-Decay）机制，我们迫使模型在训练期间将推理逻辑聚合并压缩到这些锚点中。随后，我们通过跨所有层的多级蒸馏来对齐显式和隐式路径，以确保逻辑一致性。关键是，Sep-Latent 通过基于查询复杂度自适应地终止潜在思考阶段，实现了动态推理深度。在 GSM8k-Aug 基准上的实验结果表明，Sep-Latent 以平均仅 5.4 个词元实现了 56.8\% 的准确率，在效率上显著优于先前的固定长度隐式模型，同时缩小了与显式 CoT 的性能差距。通过 PCA 和符号变体测试进行的进一步分析证实，我们的方法确保了语义基础，并对扰动表现出更优的鲁棒性。
cs.AI / 5 / 2610.22928
Multiple Turns, Short Horizons: Measuring Effective-Horizon Compression in Jailbreaks
多轮、短视界：度量越狱中的有效视界压缩
large language model
大语言模型相关
Abstract
Multi-turn jailbreaks use feedback from a victim large language model to adapt their queries over several turns. However, once an effective trajectory has been found, it is unclear how much of the earlier query history is still needed. In this work, we study this by removing early attacker queries and running the remaining queries again in a new conversation while regenerating the victim's responses. We define the effective horizon as the shortest remaining query sequence that recovers the same attack progress. Empirically, we find that roughly half of the original query sequence is sufficient on average to recover the same attack progress. We call this phenomenon effective-horizon compression. We show that this shorter history can recover the best point reached during an unsuccessful attack and provide a better starting point for continuing the attack. It also helps identify which earlier queries should receive less learning signal during attacker training, and downweighting those queries improves performance. We further find that short reward-propagation windows preserve calibrated defense performance close to full-trajectory returns. These results show that multi-turn attacks can contain more queries than are needed to recover the progress reached later in the interaction.
Chinese Translation
多轮越狱利用来自受害大型语言模型的反馈，在若干轮中调整其查询。然而，一旦找到有效的轨迹，尚不清楚此前的查询历史中有多少仍然是必需的。在本工作中，我们通过移除早期的攻击者查询，并在新的对话中重新运行剩余查询，同时重新生成受害模型的响应，来研究这一问题。我们将有效视界定义为能够恢复相同攻击进展的最短剩余查询序列。经验上，我们发现平均而言，原始查询序列的大约一半就足以恢复相同的攻击进展。我们将这一现象称为有效视界压缩。我们表明，这种更短的历史可以恢复一次不成功攻击期间达到的最佳点，并为继续攻击提供更好的起点。它还有助于识别在攻击者训练期间哪些早期查询应获得较少的学习信号，并且对这些查询降权可以提升性能。我们进一步发现，较短的奖励传播窗口能够保持校准后的防御性能接近全轨迹回报。这些结果表明，多轮攻击可能包含比恢复交互后期所达到进展所需的更多查询。
cs.AI / 6 / 2610.22973
DiffusionRVGE: Reinforcement Learning via Regression-Based Value Gradient Estimation
DiffusionRVGE：通过基于回归的价值梯度估计进行强化学习
diffusion
扩散模型相关
Abstract
Reinforcement learning improves diffusion models using rewards on generated images, but repeated denoising and gradient estimation make training expensive. We introduce \method, an efficient method that constructs useful update directions by regressing centered rewards on centered final denoised images. It requires neither reward derivatives nor adjoint backpropagation through the sampling trajectory. A unified first-variation formulation identifies the value-gradient target with or without a trajectory-KL penalty. A population analysis under Gaussian noising explains the regression geometry, while a subspace analysis motivates clean-image coordinates. The implementation adapts rollout bandwidth and query weights to reward continuity: local rollouts suit smooth rewards, whereas broader groups provide informative contrasts for discrete rewards. Trained jointly on five rewards, one SD3.5-Medium model matches or exceeds DiffusionNFT's joint model on seven of eight metrics, reaching GenEval $0.94$, OCR $0.92$, and PickScore $23.82$ at about one fifth of its estimated training compute. In single-reward training, \method reaches matched reward thresholds $2.1$--$6.7\times$ faster than reproduced DiffusionNFT on identical hardware. Ablations support the clean-image estimator, direct direction transfer, and reward-dependent rollout choices. Code is available at https://github.com/jy-wang-tsinghua/DIFFUSIONRVGE.
Chinese Translation
强化学习利用生成图像上的奖励来改进扩散模型，但重复的去噪与梯度估计使训练代价高昂。我们引入 \method，这是一种高效方法，它通过将中心化奖励回归到中心化最终去噪图像上来构造有用的更新方向。它既不需要奖励导数，也不需要沿采样轨迹进行伴随反向传播。一个统一的一阶变分表述在有或没有轨迹-KL 惩罚的情况下都能识别价值梯度目标。高斯加噪下的群体分析解释了回归几何，而子空间分析为干净图像坐标提供了动机。该实现根据奖励连续性调整 rollout 带宽和查询权重：局部 rollout 适合平滑奖励，而更广的组为离散奖励提供有信息量的对比。在五个奖励上联合训练时，一个 SD3.5-Medium 模型在八项指标中的七项上达到或超过 DiffusionNFT 的联合模型，在约为其估计训练计算量五分之一的情况下达到 GenEval $0.94$、OCR $0.92$ 和 PickScore $23.82$。在单奖励训练中，\method 在相同硬件上达到相同奖励阈值比复现的 DiffusionNFT 快 $2.1$--$6.7\times$。消融实验支持了干净图像估计器、直接方向迁移以及依赖奖励的 rollout 选择。代码可在 https://github.com/jy-wang-tsinghua/DIFFUSIONRVGE 获取。
cs.AI / 7 / 2610.22992
GOLD: Geometric Optimized Latent Diffusion for Structure-Aware RNA Inverse Folding
GOLD：面向结构感知 RNA 逆折叠的几何优化潜在扩散
diffusion
扩散模型相关
Abstract
Designing RNA sequences that fold into specific 3D structures is a fundamental challenge in therapeutics. Current RNA inverse folding methods typically condition on 3D backbone coordinates but the generation is limited to 1D discrete sequences. This ignores the continuous side-chain orientations, which are essential for overall structural stability. To address this issue, we explicitly introduce side-chain geometry into the RNA generative process. Unlike co-design frameworks that diffuse across separate raw spaces, we propose Geometric Optimized Latent Diffusion (\textbf{GOLD}), which project sequence and structure into a unified latent manifold. GOLD employs a novel GeoVAE to elegantly integrate discrete RNA sequences with continuous side-chain angles. By mapping these modalities into a single, smooth latent manifold, GOLD naturally facilitates a synergistic co-diffusion process. To enhance structural validity, we further introduce a training-free, gradient-guided sampling mechanism. By decoding the generated sequence probabilities and side-chain geometries into structures during the sampling process, we compute analytical gradients from physical constraints. This dynamic feedback actively guides the generated sequence to a physically stable state. Extensive experiments validate the superiority of GOLD, achieving state-of-the-art performance across sequence recovery, secondary structure validity, and 3D foldability metrics. Our code and data are publicly available at https://github.com/darkflash03/GOLD.
Chinese Translation
设计能够折叠成特定三维结构的 RNA 序列是治疗领域的一项基础性挑战。当前的 RNA 逆折叠方法通常以三维主链坐标为条件，但生成仅限于一维离散序列。这忽略了连续的侧链取向，而侧链取向对整体结构稳定性至关重要。为解决这一问题，我们明确地将侧链几何引入 RNA 生成过程。与在分离的原始空间中扩散的协同设计框架不同，我们提出几何优化潜在扩散（\textbf{GOLD}），它将序列和结构投影到一个统一的潜在流形中。GOLD 采用一种新颖的 GeoVAE，优雅地将离散 RNA 序列与连续侧链角度整合在一起。通过将这些模态映射到单一、平滑的潜在流形中，GOLD 自然地促进了协同的联合扩散过程。为增强结构有效性，我们进一步引入一种无需训练、梯度引导的采样机制。通过在采样过程中将生成的序列概率和侧链几何解码为结构，我们计算来自物理约束的解析梯度。这种动态反馈主动引导生成的序列达到物理上稳定的状态。大量实验验证了 GOLD 的优越性，在序列恢复、二级结构有效性和三维可折叠性指标上均取得了最先进的性能。我们的代码和数据已在 https://github.com/darkflash03/GOLD 公开。
cs.AI / 8 / 2610.23038
Where Research Ideation Begins: Forward-Citation Supervision for Actionable Limitation Generation
研究构思始于何处：面向可操作局限性生成的前向引用监督
large language model
大语言模型相关
Abstract
Automated research ideation has advanced rapidly, with multi-agent workflows that generate whole research ideas end to end. These frameworks leave largely unexamined the step that seeds a good idea: identifying an actionable limitation from literature. In this paper, we focus on this first step in the research ideation pipeline. Starting from a pilot study, we find that as large language models (LLMs) scale from 0.6B to 32B parameters, they become far better at recognizing which limitation later work addressed, but no better at generating one, suggesting that the bottleneck lies in the supervision signal rather than model capacity. We therefore construct ActLim-Bench from forward citations, yielding 18,279 evidence-linked supervision triples with human verification. On this signal, we propose a three-stage post-training pipeline: supervised fine-tuning, verifier-filtered rejection-sampling fine-tuning, and reinforcement learning with the verifier as an online dense reward. Experimental results show that our method achieves the highest recall score compared to existing methods among all frontier models. We further show that our generated limitations are preferred in 64.9% of non-tied human comparisons and improve downstream ideation systems. We will release our data and code upon publication.
Chinese Translation
自动化研究构思已快速发展，多智能体工作流能够端到端地生成完整的研究想法。这些框架在很大程度上仍未审视孕育一个好想法的步骤：从文献中识别出可操作的局限性。在本文中，我们聚焦于研究构思流程中的这一第一步。从一项初步研究出发，我们发现，随着大型语言模型（LLMs）从 0.6B 参数扩展到 32B 参数，它们在识别后续工作解决了哪一局限性方面变得好得多，但在生成局限性方面并没有变得更好，这表明瓶颈在于监督信号而非模型容量。因此，我们从前向引用构建了 ActLim-Bench，产生了 18,279 个经人工验证的、与证据关联的监督三元组。基于这一信号，我们提出一个三阶段后训练流程：监督微调、经验证器过滤的拒绝采样微调，以及以验证器作为在线稠密奖励的强化学习。实验结果表明，在所有前沿模型中，与现有方法相比，我们的方法取得了最高的召回率得分。我们进一步表明，在非平局的人工比较中，我们生成的局限性在 64.9% 的情况下更受青睐，并且能够改进下游构思系统。我们将在论文发表时发布数据和代码。
cs.AI / 9 / 2610.23141
LexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation
LexPilot：面向中国民事案件咨询中法律大语言模型与普通民众主动交互能力的基准测试
large language model
大语言模型相关
Abstract
Large Language Models (LLMs) have greatly assisted legal professionals. However, our empirical studies reveal that, when faced with underspecified consultations from non-experts, current LLMs often produce irresponsible responses with limited utility despite lacking sufficient case details for definitive conclusions. To better assist laypeople, legal LLMs must identify focal points of dispute and proactively elicit crucial context from users. Unfortunately, existing data resources often focus on interaction and response quality, offering limited insight into how proactive interaction affects LLMs' downstream claim-level judgment prediction in actual civil cases. To address this gap, we introduce \textbf{LexPilot}, a benchmark linking proactive legal interaction to claim-level verdict prediction using authentic Chinese civil judgments. Constructed from 432 real-world judgments across 36 common causes of action, LexPilot comprises expert-reviewed pairs of complete and underspecified consultation histories with corresponding court verdicts. Experiments reveal that current LLMs struggle with proactive legal interactions in underspecified consultations, with downstream judgment prediction substantially underperforming that in the fully specified consultation setup (-23.84% on average). Further analysis identifies critical bottlenecks hindering these capabilities. Our work highlights the urgent need for more user-centered legal AI to better serve the public.
Chinese Translation
大语言模型（LLMs）极大帮助了法律专业人士。然而，我们的实证研究表明，当面对非专家提出的信息不充分的咨询时，当前LLMs尽管缺乏足以得出确定结论的案件细节，却往往会产生效用有限且不负责任的回答。为了更好地帮助普通民众，法律LLMs必须识别争议焦点，并主动从用户处引出关键背景信息。遗憾的是，现有数据资源往往关注交互和响应质量，对于主动交互如何影响LLMs在实际民事案件中的下游请求级判决预测，所提供的洞见有限。为弥补这一空白，我们提出\textbf{LexPilot}，一个利用真实中国民事判决将主动法律交互与请求级判决预测联系起来的基准。LexPilot构建自36种常见案由的432份真实判决，包含经专家审阅的完整咨询历史与信息不充分的咨询历史配对，以及相应的法院判决。实验表明，当前LLMs在信息不充分的咨询中难以进行主动法律交互，其下游判决预测显著低于完全明确咨询设置下的表现（平均-23.84%）。进一步分析识别出阻碍这些能力的关键瓶颈。我们的工作凸显了迫切需要更加以用户为中心的法律AI，以更好地服务公众。
cs.AI / 10 / 2610.23226
Characterising and Analysing Retrieval-Augmented Generation Techniques
检索增强生成技术的刻画与分析
large language model
大语言模型相关
Abstract
The advent of Large Language Models (LLMs) across many application domains has increased the need for Retrieval-Augmented Generation (RAG) to reduce hallucinations and improve LLM performance. Thus far, several RAG techniques with impressive features and capabilities have been proposed in the literature, making it a herculean task to select the most appropriate RAG techniques for specific problem scenarios. Thus, we propose a conceptual framework comprising main dimensions and sub-dimensions for characterising (defining "what it is") and analysing (defining "how well it does") existing RAG techniques. The conceptual framework was developed through a qualitative analysis of the attributes of RAG techniques reported in the literature. It can serve as a pre-empirical filtering mechanism for researchers and practitioners in selecting the highest-fit RAG candidates that meet specific user and task requirements before empirical benchmarking of RAG techniques, which can be computationally and financially expensive.
Chinese Translation
大型语言模型（LLM）在众多应用领域中的出现，增加了对检索增强生成（RAG）的需求，以减少幻觉并提升 LLM 的性能。迄今为止，文献中已提出了若干具有令人印象深刻的功能与能力的 RAG 技术，这使得针对特定问题场景选择最合适的 RAG 技术成为一项艰巨的任务。因此，我们提出了一个由主维度和子维度构成的概念框架，用于刻画（界定“它是什么”）和分析（界定“它做得有多好”）现有的 RAG 技术。该概念框架是通过对文献中报道的 RAG 技术属性进行定性分析而开发的。它可以作为一种先于实证的筛选机制，帮助研究人员和实践者在开展 RAG 技术的实证基准测试之前，选出最符合特定用户与任务要求的 RAG 候选方案，而此类实证基准测试在计算和资金上可能代价高昂。
cs.AI / 11 / 2610.23290
CoTA++: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective
CoTA++：从信息流视角理解 dMLLMs 中的重复诅咒
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Diffusion-based multimodal large language models (dMLLMs) decode by iterative denoising with bidirectional attention, which precludes the standard autoregressive key--value cache. Practical deployments therefore rely on approximate caching that reuses intermediate states across denoising steps. Such caching, however, often makes a response repeat one token at adjacent positions in a run that breaks the sentence, a failure we term the \emph{Repeat Curse}. Tracing the context-token information flow reveals three disruptions. Within a step, caching disperses the attention that anchors a position on nearby context tokens. Across steps, similarity-based refresh is not staleness-aware, so some positions are decoded against stale key--value states. Across layers, the deep-layer entropy of the context fails to converge where repetition arises. We accordingly propose CoTA++, a training-free, plug-and-play framework that counteracts each disruption where it originates: Context-Token Attention Re-anchoring re-forms the attention routing against the current context, Decode-Aware Refresh reserves part of the cache's refresh budget for the positions about to be committed, and Context-Token Entropy-Guided Voting discounts the confidence of a position by the deep-layer entropy of its context, which reorders the unmasking. The three components act at non-overlapping stages of a decoding step and compose freely with existing dMLLMs and caches. Across three dMLLMs and two caching backends, CoTA++ reduces the sample repetition rate of long responses from up to 99.6\% to at most 17.0\%, recovers most of the accuracy lost to caching on nine benchmarks, and keeps the speedup of caching at under 15\% overhead, making approximate caching reliable for dMLLMs.
Chinese Translation
基于扩散的多模态大语言模型（dMLLMs）通过双向注意力的迭代去噪进行解码，这排除了标准的自回归键--值缓存。因此，实际部署依赖于近似缓存，即在去噪步骤之间复用中间状态。然而，这种缓存常常使回复在一段连续文本中相邻位置重复同一个 token，从而破坏句子，我们将这一失败称为\emph{重复诅咒}。追踪上下文 token 的信息流揭示了三种扰乱。在单步内，缓存分散了将某个位置锚定在邻近上下文 token 上的注意力。在跨步之间，基于相似性的刷新不具备陈旧感知，因此某些位置是依据陈旧的键--值状态解码的。在跨层之间，上下文的深层熵在重复出现的位置无法收敛。据此，我们提出 CoTA++，一个无需训练、即插即用的框架，在每种扰乱产生的环节对其加以抵消：上下文 Token 注意力重锚定针对当前上下文重新形成注意力路由，解码感知刷新为即将提交的位置保留部分缓存的刷新预算，上下文 Token 熵引导投票根据某位置上下文的深层熵折减该位置的置信度，从而重新排序去掩码顺序。这三个组件在解码步骤中互不重叠的阶段起作用，并可与现有 dMLLMs 和缓存自由组合。在三个 dMLLMs 和两种缓存后端上，CoTA++ 将长回复的样本重复率从最高 99.6\% 降至最多 17.0\%，在九个基准上恢复了因缓存而损失的大部分准确率，并将缓存的加速保持在低于 15\% 的开销下，使近似缓存对 dMLLMs 而言变得可靠。
cs.AI / 12 / 2610.23311
Self-Reflective Chain-of-Thought (SRCoT): Evidence-Based Text Generation via a Closed-Loop Retrieve--Reflect--Revise Pipeline on Edge LLMs
自我反思式思维链（SRCoT）：在边缘 LLM 上通过闭环检索--反思--修订流水线进行基于证据的文本生成
large language model
大语言模型相关
Abstract
Large language models (LLMs) can generate fluent outputs that lack source support or conflict with local operational constraints. We introduce Self-Reflective Chain-of-Thought (SRCoT), a modular architecture that organizes generation as a closed Retrieve--Reflect--Revise loop. SRCoT separates environment state, persistent knowledge, accumulated experience, and short-term memory into provenance-labeled stores. BM25 retrieval supplies local context; subsequent passes expose intermediate drafts, missing-information markers, follow-up queries, and optional claim--evidence mappings. These artifacts inform retrieval, revision, clarification, or abstention decisions. We examine the architecture through qualitative, trace-based case studies in two fully synthetic settings: emergency medical services and Munich-Sim, an emergency-operations simulation. Local prototype runs on an Apple M4-family laptop with 128 GB memory illustrate source-labeled documentation, explicit gap marking, and structured dispatch-style outputs. The traces also reveal provenance-assignment errors, incomplete evidence coverage, and unsafe assumptions that remain possible despite prompting constraints. These cases demonstrate report-level inspectability and local execution, not comparative superiority, hallucination reduction, clinical validity, or operational safety. The architecture separates retrieval and control from the model interface, but behavior across different models remains to be evaluated. SRCoT is a research proof of concept requiring human review, with its principal contribution lying in an inspectable control procedure rather than a new model.
Chinese Translation
大型语言模型（LLM）能够生成流畅的输出，但这些输出缺乏来源支持或与本地操作约束相冲突。我们引入自我反思式思维链（SRCoT），这是一种模块化架构，将生成组织为一个闭环的检索--反思--修订循环。SRCoT 将环境状态、持久知识、积累经验和短期记忆分离到带有来源标记的存储中。BM25 检索提供局部上下文；后续轮次会暴露中间草稿、缺失信息标记、后续查询以及可选的论断--证据映射。这些产物为检索、修订、澄清或弃答决策提供依据。我们通过两个完全合成场景中的定性、基于轨迹的案例研究来考察该架构：紧急医疗服务以及 Munich-Sim（一个应急行动模拟）。在配备 128 GB 内存的 Apple M4 系列笔记本电脑上进行的本地原型运行，展示了带来源标签的文档记录、显式的缺口标记以及结构化的调度式输出。这些轨迹还揭示了来源分配错误、证据覆盖不完整以及不安全的假设，尽管有提示约束，这些问题仍然可能发生。这些案例展示了报告层面的可检查性和本地执行，而非比较优势、幻觉减少、临床有效性或操作安全性。该架构将检索和控制与模型接口分离，但不同模型间的行为仍有待评估。SRCoT 是一个需要人工审查的研究概念验证，其主要贡献在于可检查的控制过程，而不是一个新模型。
cs.AI / 13 / 2610.23465
S-GAP: Stealthy Graph Attack Primitives for Graph Large Language Models
S-GAP：面向图大语言模型的隐蔽图攻击原语
large language model
大语言模型相关
Abstract
Graph Large Language Models (GraphLLMs) enable powerful semantic understanding and reasoning over Text-Attributed Graphs (TAGs), while their semantic reasoning may exhibit vulnerabilities to perturbations in node text and graph structure under adversarial attack evaluation. Existing attacks typically construct textual and structural perturbations in separate stages guided by distinct realization intents. However, this separation results in attack-intent fragmentation, causing semantic incoherence between the two perturbation types, thereby undermining attack stealth and making them easier for GraphLLMs to identify and disregard, ultimately weakening attack effectiveness. To avoid attack-intent fragmentation, we propose the Graph Attack Primitive (GAP), an atomic semantic representation that unifies textual and structural perturbations under a coherent attack intent. By carefully calibrating the semantic composition of a GAP, we can achieve strong attack effectiveness while maintaining sufficient attack stealth. Furthermore, we introduce S-GAP (Stealthy GAP), a multi-agent framework that iteratively refines the semantic content of the GAP. The framework unifies GAP planning and realization and continuously refines this process through surrogate evaluation and textual feedback. Extensive experiments across four datasets, three GraphLLMs, and two attack settings show that S-GAP improves ASR over the strongest baselines by 7.02% on average while producing more relationally plausible perturbations, exposing critical vulnerabilities in current GraphLLMs and motivating more robust defenses.
Chinese Translation
图大语言模型（GraphLLMs）能够对文本属性图（TAGs）实现强大的语义理解与推理，然而在对抗攻击评估下，其语义推理可能对节点文本与图结构中的扰动表现出脆弱性。现有攻击通常在不同阶段、由不同的实现意图引导，分别构造文本扰动与结构扰动。然而，这种分离导致攻击意图碎片化，造成两类扰动之间的语义不一致，从而削弱攻击隐蔽性，使其更易被GraphLLMs识别并忽略，最终削弱攻击有效性。为避免攻击意图碎片化，我们提出图攻击原语（Graph Attack Primitive, GAP），这是一种原子语义表示，在连贯的攻击意图下统一文本扰动与结构扰动。通过仔细校准GAP的语义构成，我们能够在保持足够攻击隐蔽性的同时实现强大的攻击有效性。此外，我们引入S-GAP（Stealthy GAP），这是一个多智能体框架，可迭代地细化GAP的语义内容。该框架统一了GAP的规划与实现，并通过代理评估与文本反馈持续细化这一过程。在四个数据集、三个GraphLLMs和两种攻击设置下开展的大量实验表明，S-GAP相较最强基线平均将ASR提升7.02%，同时产生关系上更合理的扰动，揭示了当前GraphLLMs中的关键脆弱性，并促使研究更稳健的防御。
cs.AI / 14 / 2610.23511
LocalIntent: Outcome-Specific Structured Generation for Backend-Compatible Smart Building Automation with Local LLMs
LocalIntent：面向后端兼容的本地 LLM 智能建筑自动化的结果特定结构化生成
large language model
大语言模型相关
Abstract
Locally deployed large language models (LLMs) can reduce reliance on remote cloud services for natural language smart building automation, but backend-facing processing requires correct outcome selection, alignment with local capabilities, and valid structured outputs. We investigate whether a complete outcome-specific prompting design improves compact-model outputs beyond generic route-then-generate prompting. The proposed LocalIntent approach routes requests to ordinary or persistent-rule branches: the former produces commands, clarifications, or rejections, while the latter generates trigger-action rules, followed by deterministic validation. We evaluate four locally executed Gemma and Qwen models spanning 2B-9B parameters on 101 non-persistent and 54 persistent-rule records, comparing Monolithic prompting, generic two-stage prompting, LocalIntent, and validation-guided repair. Across the evaluated datasets, LocalIntent achieves higher intent understanding, device identification, and action/parameter correctness than generic two-stage prompting, while clarification/mode correctness remains model-dependent. The largest gains occur for rules, where LocalIntent improves intent understanding from 26.4% to 80.6%, device identification from 59.3% to 84.7%, rule-mode correctness from 49.1% to 98.6%, and action/parameter correctness from 5.6% to 44.9%. For non-persistent requests, pooled gains are smaller: intent understanding increases from 39.6% to 53.2%, and action/parameter correctness from 30.7% to 42.8%.
Chinese Translation
本地部署的大语言模型（LLM）能够减少自然语言智能建筑自动化对远程云服务的依赖，但面向后端的处理需要正确的结果选择、与本地能力的对齐以及有效的结构化输出。我们研究一种完整的结果特定提示设计是否能够相较通用的“先路由后生成”提示改善紧凑模型的输出。所提出的 LocalIntent 方法将请求路由到普通分支或持久规则分支：前者产生命令、澄清请求或拒绝，后者生成触发-动作规则，随后进行确定性验证。我们在 101 条非持久性记录和 54 条持久规则记录上评估了四种本地执行的 Gemma 和 Qwen 模型（参数规模覆盖 2B–9B），比较了单体式提示、通用两阶段提示、LocalIntent 以及验证引导的修复。在所评估的数据集上，LocalIntent 相比通用两阶段提示取得了更高的意图理解、设备识别和动作/参数正确性，而澄清/模式正确性仍取决于模型。最大的提升出现在规则上，LocalIntent 将意图理解从 26.4% 提升到 80.6%，设备识别从 59.3% 提升到 84.7%，规则模式正确性从 49.1% 提升到 98.6%，动作/参数正确性从 5.6% 提升到 44.9%。对于非持久性请求，汇总后的增益较小：意图理解从 39.6% 增加到 53.2%，动作/参数正确性从 30.7% 增加到 42.8%。
cs.AI / 15 / 2610.23525
Trapped in Your Story: Evaluating Narrative-Induced Sycophancy in Large Language Models
困于你的故事：评估大语言模型中由叙事引发的谄媚行为
large language model
大语言模型相关
Abstract
Large language models (LLMs) are known to exhibit sycophancy by adjusting their responses toward users' expressed views. However, real users often justify their opinions through reasons, personal experiences, causal explanations, and broader narratives, raising the question of whether such narratives can further steer model judgments. We construct a controlled benchmark of 500 debatable questions across ten domains, where the user stance is held fixed while its justification is systematically varied without introducing new task-relevant evidence. Each question is instantiated under 11 matched conditions: one neutral baseline and five mirrored narrative levels for each of two opposing positions, yielding 5,500 evaluation instances. We evaluate seven LLMs and observe substantial differences in narrative-induced agreement. The effect is not monotonic with narrative elaboration: personal experience can reduce agreement, whereas causal framing and worldview-based synthesis increase it. Controlled GPT-5.1 ablations further show that causal framing produces the strongest positive effect, while claim-preserving bullet presentation also increases agreement over paragraph form. These findings suggest that sycophancy is shaped not only by user stance, but also by how supporting explanations are structured and presented.
Chinese Translation
已知大语言模型（LLMs）会表现出谄媚行为，即将其回答调整至偏向用户所表达的观点的方向。然而，真实用户常常通过理由、个人经历、因果解释以及更宽泛的叙事来为其观点提供依据，这就引出了一个问题：此类叙事能否进一步左右模型的判断。我们构建了一个受控基准，涵盖十个领域中的500个可争议问题，其中用户的立场保持固定，而其论证依据被系统地变化，且不引入新的与任务相关的证据。每个问题在11种匹配条件下被实例化：一个中性基线，以及针对两个相互对立立场各自的五个镜像叙事水平，共产生5,500个评估实例。我们评估了七个LLM，并观察到由叙事引发的同意程度存在显著差异。该效应并不随叙事详尽程度的提高而单调变化：个人经历可能降低同意程度，而因果框架和基于世界观的综合则会提高同意程度。对GPT-5.1进行的受控消融实验进一步表明，因果框架产生的正向效应最强，而保留主张的要点式呈现相比段落形式也会提高同意程度。这些发现表明，谄媚行为不仅由用户立场塑造，也受到支持性解释的结构与呈现方式的影响。
cs.AI / 16 / 2610.23554
SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents
SkillBrowser：面向可扩展技能使用智能体的选择性技能暴露
large language model
大语言模型相关
Abstract
Large language model agents increasingly rely on skills to extend their capabilities in complex tasks. However, scalable skill use requires Selective Skill Exposure (SSE): providing relevant procedural content when needed while avoiding unnecessary exposure that may disrupt reasoning. Existing paradigms fall short in complementary ways. Progressive disclosure exposes metadata for a growing skill library, introducing context pollution; retrieval limits catalog exposure, but static retrieval from the initial task query may be misaligned with capability needs that emerge during reasoning and may still inject unnecessary skills. We propose SkillBrowser, a Meta-Skill-managed framework that enables an agent to identify current capability needs, formulate targeted queries, and selectively load skills on demand. Across six SRA-Bench subsets and five LLMs, SkillBrowser improves aggregate accuracy over static retrieval and LLM-based selection, including multi-step tool use, and mitigates accuracy degradation on already-solvable instances. Benefits remain model- and task-dependent, with under-invocation of useful skills and additional latency as key limitations. These results demonstrate the viability of Meta-Skill-managed SSE for scalable skill repositories.
Chinese Translation
大语言模型智能体日益依赖技能来扩展其在复杂任务中的能力。然而，可扩展的技能使用需要选择性技能暴露（Selective Skill Exposure, SSE）：在需要时提供相关的程序性内容，同时避免可能干扰推理的不必要暴露。现有范式以互补的方式存在不足。渐进式披露为不断增长的技能库暴露元数据，引入了上下文污染；检索限制了目录暴露，但基于初始任务查询的静态检索可能与推理过程中出现的能力需求不一致，并且仍可能注入不必要的技能。我们提出 SkillBrowser，一个由元技能（Meta-Skill）管理的框架，使智能体能够识别当前能力需求、制定有针对性的查询，并按需选择性地加载技能。在六个 SRA-Bench 子集和五个 LLM 上，SkillBrowser 相较于静态检索和基于 LLM 的选择提升了总体准确率，包括多步工具使用，并缓解了在原本可解实例上的准确率下降。其收益仍取决于模型和任务，有用技能调用不足和额外延迟是主要局限。这些结果证明了由元技能管理的 SSE 对于可扩展技能仓库的可行性。
cs.AI / 17 / 2610.23574
Darwin-Claw: Environment-Adaptive Experience Consolidation for Recursively Self-Evolving Agents
Darwin-Claw：面向递归自我演化智能体的环境自适应经验巩固
large language model
大语言模型相关
Abstract
While Large Language Model (LLM) agents have demonstrated remarkable capabilities across complex tasks, they predominantly operate in isolated sessions, lacking the ability to continually learn from historical interactions or adapt to user-specific environments. Existing self-evolution paradigms often indiscriminately inject raw interaction trajectories into memory or skill modules, leading to obscured behavioral patterns and compromised update controllability. Furthermore, as foundational LLM reasoning improves, the bottleneck for agentic evolution shifts from accumulating generic knowledge to effectively managing situational experience. To address this, we propose Darwin-Claw, an environment-adaptive self-evolving framework that systematically distills episodic experiences to drive the comprehensive evolution of operational skills and durable memory. Theoretically grounded in an iterative agent-actor and evolver-critic co-adaptation, Darwin-Claw introduces a Hierarchical Experience Consolidation System that extracts fine-grained error attributions and generalizes them into dynamically updated prototype vectors. This structured situational experience serves as interaction patches for context-sensitive perception and explicit optimization signals for skill refinement. Practically, Darwin-Claw is a readily deployable, automated framework compatible with mainstream agent systems without requiring manual curation. Extensive experiments on WildClawBench demonstrate that Darwin-Claw yields substantial performance improvements, establishing a controllable and stable pathway for the continual, user-centric adaptation of autonomous agents in real-world ecosystems.
Chinese Translation
尽管大语言模型（LLM）智能体已在复杂任务中展现出卓越的能力，但它们主要在孤立的会话中运行，缺乏从历史交互中持续学习或适应用户特定环境的能力。现有的自我演化范式往往不加区分地将原始交互轨迹注入记忆或技能模块，导致行为模式被遮蔽、更新的可控性受损。此外，随着基础 LLM 推理能力的提升，智能体演化的瓶颈已从积累通用知识转向有效管理情境经验。为解决这一问题，我们提出 Darwin-Claw，一个环境自适应的自我演化框架，它系统性地提炼情节经验，以驱动操作技能与持久记忆的全面演化。Darwin-Claw 在理论上以迭代的智能体-执行者（agent-actor）与演化器-评论者（evolver-critic）协同适应为基础，引入了分层经验巩固系统，该系统提取细粒度的错误归因，并将其泛化为动态更新的原型向量。这种结构化的情境经验既可作为面向上下文敏感感知的交互补丁，也可作为技能精炼的显式优化信号。在实践层面，Darwin-Claw 是一个易于部署的自动化框架，与主流智能体系统兼容，且无需人工整理。在 WildClawBench 上的大量实验表明，Darwin-Claw 带来了显著的性能提升，为自主智能体在真实世界生态系统中持续地、以用户为中心的适应建立了一条可控且稳定的路径。
cs.AI / 18 / 2610.23599
From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents
从轨迹到差异化技能：面向自我改进智能体的认知分化
large language model
大语言模型相关
Abstract
Large language model (LLM) agents tackle complex tasks through reasoning and interaction with external environments. Agent skills provide persistent, transferable capabilities that guide task planning, inform contextual decisions, support tool use, and enable executable procedures. Yet existing skill-evolution methods largely organize knowledge around skill artifacts, overlooking the functional heterogeneity within individual skills and its implications for knowledge maintenance and use. Inspired by cognitive science, we propose Cognitive Differentiation in Skill Evolution (CDSE), which evolves a unified skill through four functionally distinct yet coordinated components: Meta, Experience, Tool, and Script. CDSE extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates and validation with cross-component coordination and adaptive invocation. An analysis of 500 top-ranked skill packages shows that 96.8% contain knowledge spanning at least two functional roles. Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average. Ablation studies further support the importance of functional differentiation and adaptive invocation.
Chinese Translation
大语言模型（LLM）智能体通过推理以及与外部环境交互来解决复杂任务。智能体技能提供持久、可迁移的能力，这些能力指导任务规划、为上下文决策提供信息、支持工具使用，并促成可执行流程。然而，现有的技能演化方法主要围绕技能制品来组织知识，忽视了单个技能内部的功能异质性及其对知识维护和使用的影响。受认知科学启发，我们提出技能演化中的认知分化（CDSE），它通过四个功能不同但相互协调的组件来演化统一技能：Meta、Experience、Tool 和 Script。CDSE 从成功和失败的轨迹中提取有证据支撑的语义梯度，按功能角色对其进行整合，并在跨组件协调和自适应调用下应用组件特定的更新与验证。对 500 个排名最高的技能包的分析表明，96.8% 的技能包包含跨越至少两种功能角色的知识。在六个基准、三个智能体模型和两个执行框架上，CDSE 在所有 36 种设置中均取得最高分数，相较于 Base 智能体提高了 7.6 个百分点，并相较于各设置中最强的竞争方法平均提高 2.6 个点。消融研究进一步支持了功能分化和自适应调用的重要性。
cs.AI / 19 / 2610.23722
Beyond Large Language Models: A Mental World Model Framework for Mental Modeling, Simulation, Evolution, and Interpretability
超越大型语言模型：面向心理建模、仿真、演化与可解释性的心理世界模型框架
large language model
大语言模型相关
Abstract
Understanding and simulating how subjective cognition and mental states emerge, evolve, and shape behavior is essential for advancing artificial intelligence toward deeper social cognition and general intelligence. Existing large language models (LLMs) and Theory of Mind (ToM) methods primarily infer mental states from linguistic information, lacking unified mechanisms for subjective world construction, cognitive-psychological interactions, dynamic evolution, and historical dependencies. This paper introduces the Mental World Model (MWM) and its implementation, the Multi-Branch Mental Evolution Architecture (MMEA), which explicitly constructs subjective worlds, models interactions between cognition and mental states, and simulates continuous mental evolution shaped by historical experiences. On DynToM and RecToM, MMEA achieves 96.48% and 94.00% accuracy, respectively, outperforming leading LLMs and specialized ToM methods. Additional experiments demonstrate improvements of 28.7, 39.3, and 16.6 percentage points in mental state and behavior prediction, long-sequence prediction, and implicit mental state inference, respectively. The results demonstrate that MMEA provides an interpretable framework for continuous mental modeling, enabling AI to move beyond surface-level semantic representations to infer individuals' underlying mental states from linguistic information and dynamically simulate their internal cognitive processes and evolution over time. MWM also provides a promising foundation for causal reasoning, personalized AI, sustained human-AI interaction, and future advances toward artificial general intelligence (AGI).
Chinese Translation
理解并仿真主观认知与心理状态如何涌现、演化并塑造行为，对于推动人工智能迈向更深层的社会认知与通用智能至关重要。现有的大型语言模型（LLMs）与心智理论（Theory of Mind, ToM）方法主要从语言信息中推断心理状态，缺乏用于主观世界构建、认知—心理交互、动态演化与历史依赖的统一机制。本文引入心理世界模型（Mental World Model, MWM）及其实现——多分支心理演化架构（Multi-Branch Mental Evolution Architecture, MMEA），该架构显式构建主观世界，建模认知与心理状态之间的交互，并仿真由历史经验塑造的连续心理演化。在DynToM和RecToM上，MMEA分别达到96.48%和94.00%的准确率，优于领先的LLMs和专门化的ToM方法。额外实验表明，其在心理状态与行为预测、长序列预测以及隐含心理状态推断上分别提升了28.7、39.3和16.6个百分点。结果表明，MMEA为连续心理建模提供了一个可解释的框架，使AI能够超越表层语义表征，从语言信息中推断个体的潜在心理状态，并随时间动态仿真其内部认知过程与演化。MWM还为因果推理、个性化AI、持续的人机交互以及迈向通用人工智能（AGI）的未来进展提供了有前景的基础。
cs.AR / 20 / 2610.22579
MemLLM: End-to-end In-Storage Acceleration of On-device Retrieval-Augmented Generation using Precomputed KV Caches
MemLLM：使用预计算 KV 缓存的设备端检索增强生成的端到端存储内加速
large language model
大语言模型相关
Abstract
Retrieval-augmented generation (RAG) on edge devices enables personalized large language model (LLM) inference while preserving privacy. However, RAG's inherent length asymmetry makes the prefill phase the dominant bottleneck. While precomputing the KV caches offline alleviates the prefill cost on server systems, the multi-gigabyte caches over the narrow interconnect create a severe I/O wall for on-device systems. To mitigate this, we propose MemLLM, an end-to-end in-storage processing (ISP) architecture that executes the complete attention pipeline based on precomputed KV caches within the SSD. MemLLM utilizes per-channel accelerators to fully utilize the aggregated NAND channel bandwidth and accelerates all KV cache-related computations on the storage-side to minimize PCIe transfers. To efficiently handle the sparse updates required by RAG, MemLLM introduces a fused recomputation-query kernel paired with a hardware-based filter logic to halve the NAND read volume and eliminate write-back penalties. Furthermore, MemLLM's global SRAM-based stateful reduction engine enables a completely DRAM-less dataflow, thereby removing the internal memory bottlenecks and bounding the PCIe egress to an $O(1)$ footprint. A full-stack software stack allows asynchronous, zero-copy transfers from the ISP to the host and provides transparent execution within modern ML frameworks. Evaluated across eight real-world RAG datasets and six on-device LLM models, MemLLM achieves an average end-to-end speedup of 4.06$\times$ over the GPU baseline.
Chinese Translation
边缘设备上的检索增强生成（RAG）能够在保护隐私的同时实现个性化大语言模型（LLM）推理。然而，RAG 固有的长度不对称性使得预填充阶段成为主要瓶颈。尽管离线预计算 KV 缓存减轻了服务器系统上的预填充开销，但数 GB 的缓存通过狭窄互连传输会给设备端系统造成严重的 I/O 墙。为缓解这一问题，我们提出 MemLLM，一种端到端存储内处理（ISP）架构，该架构在 SSD 内基于预计算 KV 缓存执行完整的注意力流水线。MemLLM 利用每通道加速器来充分利用聚合的 NAND 通道带宽，并在存储侧加速所有与 KV 缓存相关的计算，以最小化 PCIe 传输。为高效处理 RAG 所需的稀疏更新，MemLLM 引入融合的重计算-查询内核，并配合基于硬件的过滤逻辑，将 NAND 读取量减半并消除写回惩罚。此外，MemLLM 的基于全局 SRAM 的有状态归约引擎实现了完全无 DRAM 的数据流，从而消除内部内存瓶颈，并将 PCIe 出站数据量限制在 $O(1)$ 的占用范围内。全栈软件栈允许从 ISP 到主机的异步零拷贝传输，并在现代机器学习框架内提供透明执行。在八个真实世界 RAG 数据集和六个设备端 LLM 模型上进行评估，MemLLM 相比 GPU 基线实现了平均 4.06$\times$ 的端到端加速。
cs.AR / 21 / 2610.22877
BS-Spec: BFP-Based Bit-Sharing for Self-Speculative LLM Decoding with LUT-Driven Mixed-Precision Acceleration
BS-Spec：基于BFP的位共享，用于LUT驱动的混合精度加速的自推测LLM解码
large language model
大语言模型相关
Abstract
As large language models (LLMs) are increasingly deployed across diverse inference services, serving workloads span a wide range of batch sizes, context lengths, and performance requirements. Speculative decoding improves inference throughput, but conventional approaches require a separate draft model and KV cache, while existing self-speculative methods suffer from workload-insensitive draft configurations and inefficient mixed-precision execution. In this paper, we present BS-Spec, a block floating point (BFP)-based bit-sharing framework that jointly optimizes draft construction, workload-aware precision selection, and hardware execution. BS-Spec derives multiple draft representations directly from a single BFP target model by partially loading the most significant bits of weights and the KV cache while preserving shared exponents, eliminating separate draft-model storage. Offline calibration selects the weight and KV-cache precision pair that maximizes estimated speedup for each serving regime, and a precision lookup table applies the selected configuration at runtime. To support this adaptive execution, we design a mixed-precision BFP accelerator with a lookup table (LUT)-based variable-precision datapath for both draft and target computation. Experimental results show that BS-Spec achieves up to $1.9\times$ speedup over conventional autoregressive decoding and improves speedup over the evaluated bit-sharing self-speculative decoding baselines by $67.2\%$ on average, demonstrating the benefits of jointly co-designing bit sharing, precision adaptation, and specialized hardware for diverse LLM serving workloads.
Chinese Translation
随着大语言模型（LLM）越来越多地部署于多样化推理服务中，服务负载涵盖广泛的批量大小、上下文长度和性能需求。推测解码提高了推理吞吐量，但传统方法需要单独的草稿模型和KV缓存，而现有自推测方法受制于对负载不敏感的草稿配置以及低效的混合精度执行。本文提出BS-Spec，一个基于块浮点（BFP）的位共享框架，联合优化草稿构建、负载感知精度选择和硬件执行。BS-Spec通过部分加载权重和KV缓存的最高有效位并保留共享指数，直接从单个BFP目标模型导出多种草稿表示，从而消除单独的草稿模型存储。离线校准为每种服务场景选择使估计加速比最大化的权重与KV缓存精度对，并由精度查找表在运行时应用所选配置。为支持这种自适应执行，我们设计了一种混合精度BFP加速器，其具有基于查找表（LUT）的可变精度数据通路，用于草稿计算和目标计算。实验结果表明，BS-Spec相对于传统自回归解码实现了最高$1.9\times$的加速，并将相对于所评估的位共享自推测解码基线的加速比平均提高了$67.2\%$，证明了针对多样化LLM服务负载联合协同设计位共享、精度自适应和专用硬件的优势。
cs.CL / 22 / 2610.22511
AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems
AuditKIT：一个用于对大型语言模型与智能体系统进行基准测试、评判和比较的统一评估主干
large language model
大语言模型相关
Abstract
Evaluating language models and agents requires combining benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality. Differences in interfaces, metric semantics, and trace formats complicate these comparisons and can hide failures in agentic workflows. We present AuditKIT, a source-available evaluation library that records these measurements in one result representation, with direction-aware metric contracts, configuration fingerprints, and extensible metrics and backends. For agents, it provides turn-aware tool-call matching, separates parallel from sequential execution, and distinguishes verified task outcomes from trajectory-level diagnostics. Missing or unobservable evidence is marked unscored rather than counted as failure. For retrieval-augmented generation, it combines ranked retrieval metrics, judged grounding, and deterministic evidence checks. Seven case studies spanning model comparison, serving runtimes, retrieval, tool calling, and interoperability show how answer extraction, runtime behavior, and metric selection change evaluation outcomes.
Chinese Translation
评估语言模型与智能体，需要综合基准准确率、基于评判器的评估、运行时性能、工具使用行为以及检索质量。接口、指标语义和轨迹格式上的差异使这些比较变得复杂，并可能掩盖智能体工作流中的失败。我们提出 AuditKIT，一个源代码可获取的评估库，它将上述测量记录在单一结果表示中，并具备方向感知的指标契约、配置指纹以及可扩展的指标与后端。对于智能体，它提供轮次感知的工具调用匹配，将并行执行与顺序执行区分开来，并将经过验证的任务结果与轨迹级诊断区分开来。缺失或不可观测的证据被标记为未计分，而不是计为失败。对于检索增强生成，它结合了排序检索指标、经评判的接地性以及确定性证据检查。涵盖模型比较、服务运行时、检索、工具调用与互操作性的七个案例研究，展示了答案抽取、运行时行为与指标选择如何改变评估结果。
cs.CL / 23 / 2610.22545
Cultural (In)Competence: Investigating Culture-Specific Stereotype Detection Capabilities of LLMs
文化（不）胜任力：探究大型语言模型对文化特定刻板印象的检测能力
large language model
大语言模型相关
Abstract
Stereotypes are subjective constructs that are mediated by the cultural context in which they are viewed. Successfully modeling them thus also requires cultural competence. When using large language models (LLMs) for stereotype-related tasks, such as their detection, it is therefore crucial that models can access relevant cultural knowledge and adjust responses accordingly. Since LLMs have shown disparities in what cultural knowledge is available to them, their capacity regarding stereotypes may also be disparate across languages and cultural contexts (e.g., for non-English languages or for non-US contexts). In this paper, we systematically investigate the effectiveness of LLMs for culture-specific stereotype detection across three factors: (1) the prompting language, (2) the instruction tuning language, and (3) the fine-tuning dataset size. We conduct experiments for US, Dutch, and Chinese cultural contexts. Our results suggest that LLMs do not reliably account for cultural contexts and overpredict the presence of stereotypes. Non-English prompts seem to improve consideration of cultural context and thus lead to more effective predictions. Moreover, we find that additional instruction tuning reduces false-positive rates and that training data size moderates detection effectiveness non-monotonically.
Chinese Translation
刻板印象是主观建构，受到审视它们的文化语境的中介。因此，要成功地对它们进行建模，也需要文化胜任力。因此，在使用大型语言模型（LLMs）执行与刻板印象相关的任务（例如检测刻板印象）时，模型能否获取相关文化知识并相应调整回答至关重要。由于大型语言模型在可获取何种文化知识方面表现出差异，它们处理刻板印象的能力也可能在不同语言和文化语境之间（例如对非英语语言或非美国语境）存在差异。本文系统考察大型语言模型在文化特定刻板印象检测方面的有效性，涉及三个因素：(1) 提示语言，(2) 指令微调语言，(3) 微调数据集规模。我们针对美国、荷兰和中国文化语境开展实验。我们的结果表明，大型语言模型不能可靠地考虑文化语境，并会过度预测刻板印象的存在。非英语提示似乎能改善对文化语境的考虑，从而产生更有效的预测。此外，我们发现额外的指令微调会降低假阳性率，并且训练数据规模会非单调地调节检测有效性。
cs.CL / 24 / 2610.22618
Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification
轨迹引导的操控：缓解临床分类中的虚假特征偏倚
large language model
大语言模型相关
Abstract
Large language models rely on spurious features in clinical classification: swapping a single gender token in clinical notes alters gender-irrelevant diagnoses in up to 53.5% of cases. Existing inference-time corrections either need counterfactual data built around the specific spurious feature or behave unpredictably. We observe that how a model's preference between answers evolves across layers, its prediction trajectory, indicates for each document whether a correction is likely to help. We propose Trajectory-Guided Steering (TGS), which decides from the trajectory alone whether to correct each prediction and in which direction, with no counterfactual data and no knowledge of the spurious feature. On 16 clinical classification tasks from i2b2, TGS fixes 17.7% of errors while breaking only 1.3% of correct predictions and stays net-positive across all tested hyperparameter settings; all baselines are net-negative or near zero. The same corrections reduce reliance on the spurious feature: on the most affected diseases, gender-swap inconsistency falls sharply (hypertension 53.5% -> 38.0%, diabetes 35.1% -> 21.5%). Applying the correction to every document instead raises breaks on correct positive predictions from 0.1% to 6.4%, so per-document selection is essential. TGS also transfers to three model families and three other clinical and biomedical datasets.
Chinese Translation
大型语言模型在临床分类中依赖虚假特征：在临床记录中替换单个性别标记，会在高达53.5%的病例中改变与性别无关的诊断。现有的推理时校正要么需要围绕特定虚假特征构建的反事实数据，要么表现不可预测。我们观察到，模型在答案之间的偏好如何跨层演变，即其预测轨迹，能够指示对于每个文档，校正是否可能有所帮助。我们提出轨迹引导的操控（Trajectory-Guided Steering, TGS），它仅根据轨迹决定是否校正每个预测以及朝哪个方向校正，无需反事实数据，也无需知道虚假特征。在来自i2b2的16个临床分类任务上，TGS修复了17.7%的错误，同时仅破坏1.3%的正确预测，并且在所有测试的超参数设置下保持净正收益；所有基线均为净负或接近零。相同的校正降低了对虚假特征的依赖：在受影响最严重的疾病上，性别替换不一致性大幅下降（高血压 53.5% -> 38.0%，糖尿病 35.1% -> 21.5%）。相反，将校正应用于每个文档会使正确阳性预测的破坏率从0.1%升至6.4%，因此逐文档选择至关重要。TGS还可迁移到三个模型家族和另外三个临床与生物医学数据集。
cs.CL / 25 / 2610.22631
A Plausible Explanation Is Enough to Mislead Multimodal Judges
一个看似合理的解释就足以误导多模态评判模型
large language model
大语言模型相关
Abstract
Multimodal large language models (MLLMs) increasingly serve as judges in training and evaluation pipelines, often assessing answers together with explanations. Although AI judges are sensitive to textual factors, it remains unclear whether ordinary plausible explanations can sway judgments requiring visual verification. We introduce JUPE, a visual-question-answering (VQA)-style benchmark that fixes answers while varying explanations. Across 14 MLLM judges, ordinary explanations that sound plausible mislead judges, with effects comparable on average to deliberately persuasive arguments. Plain justifications and chain-of-thought rationales reduce rejection of incorrect answers by 12.7 and 13.8 percentage points, respectively. We call this failure plausibility capture. Using signal detection theory, we find a significant mean shift in decision criterion toward acceptance, without a clear change in mean discriminability. Explanations sway judgments across domains, explanation generators, and judging protocols, while tested mitigations offer limited, model-dependent protection. Such shifts can inflate performance estimates when judges evaluate answers with explanations.
Chinese Translation
多模态大语言模型（MLLMs）越来越多地在训练和评估流程中充当评判者，常常连同解释一起评估答案。尽管 AI 评判者对文本因素较为敏感，但普通的看似合理的解释是否能够影响需要视觉验证的判断，仍不清楚。我们提出 JUPE，一个视觉问答（VQA）风格的基准，它固定答案而改变解释。在 14 个 MLLM 评判者上，听起来合理的普通解释会误导评判者，其平均效应与刻意说服性论证相当。朴素理由和思维链推理分别将错误答案的拒绝率降低了 12.7 和 13.8 个百分点。我们将这一失败称为合理性俘获（plausibility capture）。利用信号检测理论，我们发现决策标准均值显著向接受方向偏移，而平均辨别力没有明显变化。解释在不同领域、解释生成器和评判协议下都会影响判断，而所测试的缓解措施提供的保护有限且依赖于模型。当评判者在有解释的情况下评估答案时，这种偏移可能会抬高性能估计。
cs.CL / 26 / 2610.22636
Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness
通过因果鲁棒性测量大型语言模型中的机制性不确定性
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used in critical domains, where their susceptibility to hallucinations makes reliable and interpretable uncertainty quantification essential. We introduce the concept of mechanistic uncertainty, which characterizes uncertainty through the stability of the internal mechanisms supporting a model's response. We propose the causal robustness margin, defined as the minimum internal intervention required to deviate the model's output, and operationalize it using Jacobian Lens (Jlens) concept steering vectors. Across three open-source models and five question-answering datasets, we show incorrect and uncertain responses exhibit smaller margins, indicating greater sensitivity to internal interventions, and the causal robustness margin complements conventional output-based uncertainty methods. Finally, in a controlled knowledge injection experiment, correctly learned facts show increased margins after training, suggesting that causal robustness provides a mechanistically grounded signal of uncertainty.
Chinese Translation
大型语言模型（LLMs）越来越多地被用于关键领域，在这些领域中，它们易受幻觉影响的特性使得可靠且可解释的不确定性量化变得至关重要。我们引入了机制性不确定性的概念，它通过支撑模型响应的内部机制的稳定性来刻画不确定性。我们提出了因果鲁棒性裕度，将其定义为使模型输出偏离所需的最小内部干预，并使用 Jacobian Lens（Jlens）概念引导向量将其操作化。在三个开源模型和五个问答数据集上，我们表明错误和不确定的响应表现出更小的裕度，表明对内部干预的敏感性更高，并且因果鲁棒性裕度补充了传统的基于输出的不确定性方法。最后，在一项受控的知识注入实验中，正确学习的事实显示训练后裕度增加，这表明因果鲁棒性提供了一种具有机制基础的不确定性信号。
cs.CL / 27 / 2610.22688
Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries
分开探索，选择性采纳：大型工具库上的执行引导式工具使用
large language model
大语言模型相关
Abstract
Large language model agents must discover relevant tools from large libraries, e.g., thousands of tools, as new needs arise during task execution. When several candidates could satisfy the same need, execution may be required to identify the relevant one. Investigating candidates in a shared trajectory can leave alternatives unexplored or allow failed attempts to influence subsequent reasoning. To address this, we present BRIA (BRanch-Inspect-Adopt), a simple yet effective agent that temporarily isolates candidate investigations in separate branches. Each branch retrieves and executes tools to investigate its candidate; a task agent inspects the resulting execution evidence and adopts only useful results into its continuing trajectory. BRIA retrieves additional tools as new capability needs arise during task execution. We evaluate BRIA on PlanBench-XL and MCP-Atlas using a shared pool of 1897 tools, three language models, and static and adaptive retrieval baselines. Across all settings, BRIA matches or exceeds the strongest retrieval baseline. On MCP-Atlas, it achieves an average claim score of 89.77, compared with 75.40 for MCP-Zero, a gain of 14.37 percentage points. BRIA also lies on the Pareto frontier for task performance and input-token cost.
Chinese Translation
大型语言模型智能体必须在任务执行过程中出现新需求时，从大型工具库（例如数千个工具）中发现相关工具。当多个候选工具可能满足同一需求时，可能需要通过执行来识别相关工具。在共享轨迹中考察候选工具可能导致其他备选方案未被探索，或使失败尝试影响后续推理。为解决这一问题，我们提出 BRIA（BRanch-Inspect-Adopt），这是一种简单而有效的智能体，它暂时将候选工具的考察隔离在独立分支中。每个分支检索并执行工具以考察其候选工具；任务智能体检查由此产生的执行证据，并仅将有用结果采纳到其持续进行的轨迹中。BRIA 在任务执行过程中出现新的能力需求时检索额外工具。我们在 PlanBench-XL 和 MCP-Atlas 上评估 BRIA，使用包含 1897 个工具的共享工具池、三个语言模型以及静态和自适应检索基线。在所有设置下，BRIA 达到或超过最强检索基线的表现。在 MCP-Atlas 上，它取得了 89.77 的平均 claim score，而 MCP-Zero 为 75.40，提升了 14.37 个百分点。BRIA 还位于任务性能和输入 token 成本的帕累托前沿上。
cs.CL / 28 / 2610.22721
CoLU: Confidence-Weighted Layerwise Unlearning for Large Language Models
CoLU：面向大语言模型的置信度加权逐层遗忘
large language model
大语言模型相关
Abstract
Large language model (LLM) unlearning aims to forget certain information while preserving others, without retraining from scratch. SoTA methods such as NPO, SimNPO, and SatImp balance forgetting and retain through whole model fine-tuning, while FOM-UL explores selective layer updates. Prior works show that suppressing target information can damage retained knowledge. This trade-off can vary across layers even under the same unlearning loss. In this work, we propose Confidence-Weighted Layerwise Unlearning (CoLU), which combines a novel confidence-weighted forgetting and retain regularization while updating a single selected transformer layer, rather than the whole model for unlearning. To reduce CoLU compute cost, we further propose Brief Layer Screening (BLS), which uses a newly defined normalized forgetting-retention (NFR) score to first rank layers through a brief unlearning trials using a subset of forgetting dataset and fewer epochs. BLS pre-selects top-k ranked layer candidates, which will then be fully profiled using full data analysis. BLS is proven to reduce the runtime of CoLU by 60.2-79.7% with almost the same performance. On TOFU dataset, CoLU improves both forgetting and retention over NPO, SatImp, and GA-based FOM-UL in all six TOFU settings, and over SimNPO in five of the six settings. On MUSE Books and News datasets, CoLU achieves the highest retained KnowMem and the PrivLeak closest to zero among the compared SOTA unlearning methods. CoLU presents a new record performance in both unlearning and retain in most settings up to date. Moreover, across the measured settings, average CoLU GPU memory usage is 77.4% to 79.5% lower than full-model unlearning.
Chinese Translation
大语言模型（LLM）遗忘旨在忘记某些信息而保留其他信息，且无需从头重新训练。NPO、SimNPO 和 SatImp 等 SoTA 方法通过全模型微调来平衡遗忘与保留，而 FOM-UL 则探索选择性层更新。已有工作表明，抑制目标信息可能损害保留的知识。即使在相同的遗忘损失下，这种权衡在不同层之间也可能有所不同。在这项工作中，我们提出置信度加权逐层遗忘（CoLU），它在更新单个选定的 Transformer 层而非整个模型进行遗忘的同时，结合了一种新颖的置信度加权遗忘与保留正则化。为降低 CoLU 的计算成本，我们进一步提出简要层筛选（BLS），它使用新定义的归一化遗忘-保留（NFR）分数，先通过使用遗忘数据集的子集和更少轮次的简短遗忘试验对层进行排名。BLS 预先选择排名前 k 的层候选，随后将使用全数据分析对其进行全面刻画。经验证，BLS 在性能几乎相同的情况下将 CoLU 的运行时间减少了 60.2-79.7%。在 TOFU 数据集上，CoLU 在全部六个 TOFU 设置中在遗忘和保留两方面均优于 NPO、SatImp 和基于 GA 的 FOM-UL，并在六个设置中的五个设置上优于 SimNPO。在 MUSE Books 和 News 数据集上，在所比较的 SOTA 遗忘方法中，CoLU 实现了最高的保留 KnowMem 和最接近零的 PrivLeak。CoLU 在迄今为止的大多数设置中在遗忘和保留两方面均呈现出新的纪录性能。此外，在所测量的设置中，CoLU 的平均 GPU 内存使用量比全模型遗忘低 77.4% 至 79.5%。
cs.CL / 29 / 2610.22791
Self-Evolving Coding Agents via Harness-Policy Optimization
通过执行框架—策略优化实现自进化编码智能体
large language model
大语言模型相关
Abstract
Large language model (LLM) agents have shown promising potential in coding tasks, and their performance is jointly determined by the policy model and the harness that mediates their interaction with the environment. Indeed, harness and policy are mutually reinforcing and evolving only one component is suboptimal---a harness tuned for one checkpoint may not target the failure modes of an updated policy, while policy learning remains limited to the trajectories elicited by a fixed harness. In light of this, we propose Self-HPO, a recursive self-improving framework for coding agents that alternates harness and policy optimization, using the evolving policy as its own harness proposer. Specifically, in each round, the current policy first analyzes its own failures and proposes new harness candidates (H-step). The policy is then updated through reinforcement learning (RL) under the selected harness (P-step), and hard tasks identified during training form the next harness search set, keeping harness optimization targeted to the policy's evolving failure modes. Across three coding benchmarks, Self-HPO consistently outperforms harness-only and policy-only baselines, improving Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8%, with gains generalizing to out-of-distribution settings. Further analysis shows that both optimized components retain gains even when paired with their base counterparts, while their joint gain exceeds the sum of their individual gains. Self-HPO also improves reliability (Pass@1) and expands the capability frontier (Pass@k) over the base agent. Additionally, we show that with a stronger external model (e.g., Opus 4.8) as the harness proposer, the same framework yields a further performance boost.
Chinese Translation
大语言模型（LLM）智能体在编码任务中展现出可观潜力，其性能由策略模型以及中介其与环境交互的执行框架共同决定。事实上，执行框架与策略相互增强，仅演化其中一个组件是次优的——为某一检查点调优的执行框架可能无法针对更新后策略的失败模式，而策略学习仍受限于固定执行框架所引出的轨迹。鉴于此，我们提出 Self-HPO，一种用于编码智能体的递归自改进框架，它交替进行执行框架优化与策略优化，并将演化中的策略作为其自身的执行框架提出者。具体而言，在每一轮中，当前策略首先分析自身的失败并提出新的候选执行框架（H-step）。随后，策略在所选执行框架下通过强化学习（RL）进行更新（P-step），训练过程中识别出的困难任务构成下一轮执行框架搜索集，从而使执行框架优化持续针对策略不断演化的失败模式。在三个编码基准上，Self-HPO 始终优于仅执行框架和仅策略的基线，将 Qwen3.5-35B-A3B 的平均解决率从 57.8% 提升至 64.8%，且增益可泛化到分布外设置。进一步分析表明，两个优化后的组件即使与各自的基座对应组件配对时仍能保留增益，而它们的联合增益超过各自单独增益之和。Self-HPO 还在基座智能体基础上提升了可靠性（Pass@1）并扩展了能力前沿（Pass@k）。此外，我们表明，使用更强的外部模型（例如 Opus 4.8）作为执行框架提出者时，同一框架可带来进一步的性能提升。
cs.CL / 30 / 2610.22792
Prune the Leaky Neurons: Mitigating Private Data Leakage in Fine-Tuned LLMs via Joint-Embedding-Guided Pruning
修剪泄漏神经元：通过联合嵌入引导的剪枝缓解微调大语言模型中的私有数据泄漏
large language model
大语言模型相关
Abstract
Fine-tuning large language models (LLMs) on proprietary records enables domain adaptation but introduces two challenges: (i) privacy leakage of memorized private data, such as personally identifiable information (PII), and (ii) balancing private data removal with utility preservation. Existing defenses, including data sanitization and differential privacy, often degrade performance, while privacy-neuron editing lacks a PII-free reference to distinguish identity-specific neurons from those supporting useful text. We propose JEPA-Prune, a post-hoc neuron pruning framework guided by a model trained exclusively on PII-free views. Alongside the fine-tuned private model, JEPA-Prune, trains a public model on surrogate and masked views of the same records using a joint-embedding predictive architecture (JEPA), capturing domain knowledge without identifiers. A cross-model representation probe quantifies deviations between private representations of leaked records and this PII-free reference. Integrated gradients and a scale-invariant specificity score then identify PII-responsive neurons, which are pruned without retraining. Experiments show that pruning just 1.6% of neurons reduces PII inference accuracy by up to 7.4x and PII extraction recall by up to 13.9x, while keeping unseen-text perplexity within 16% of that of the unpruned model for most LLMs. In contrast, privacy-neuron editing increases perplexity by at least 26x.
Chinese Translation
在专有记录上微调大语言模型（LLM）能够实现领域适应，但会带来两个挑战：(i) 被记忆的私有数据（例如个人可识别信息（PII））的隐私泄漏，以及 (ii) 在移除私有数据与保持效用之间取得平衡。现有防御方法，包括数据净化和差分隐私，往往会降低性能，而隐私神经元编辑缺乏一个不含PII的参考，以区分身份特定神经元与支持有用文本的神经元。我们提出 JEPA-Prune，一个事后神经元剪枝框架，由仅在无PII视图上训练的模型引导。与微调的私有模型并行的，JEPA-Prune 使用联合嵌入预测架构（JEPA）在相同记录的代理视图和掩码视图上训练一个公共模型，从而在不包含标识符的情况下捕获领域知识。一个跨模型表示探针量化泄漏记录的私有表示与该无PII参考之间的偏差。然后，积分梯度和尺度不变的特异性得分识别对PII有响应的神经元，这些神经元无需重新训练即可被剪枝。实验表明，仅剪枝1.6%的神经元就能将PII推断准确率最多降低7.4倍，并将PII提取召回率最多降低13.9倍，同时对于大多数LLM，将未见文本困惑度保持在未剪枝模型的16%以内。相比之下，隐私神经元编辑使困惑度至少增加26倍。
cs.CL / 31 / 2610.22807
On the Position Bias of Large Language Models: A Comprehensive Study
论大型语言模型的位置偏差：一项综合研究
large language model
大语言模型相关
Abstract
Recent studies have shown that position bias remains a critical vulnerability in large language models (LLMs), undermining both the integrity of multiple-choice evaluations and the reliability of retrieval-augmented generation. However, existing understanding of this phenomenon is largely derived from studies with limited model coverage and English-centric benchmarks, leaving the behavior of frontier models and diverse languages underexplored. In this work, we conduct a comprehensive investigation of position bias across model scales ranging from billion-level to trillion-level parameters, covering both multiple-choice and multi-document tasks in five languages. Our experiments reveal that while position bias remains noticeable in small-scale models and Non-English languages, increasing model scale and employing standard Chain-of-Thought prompting significantly mitigate this issue. Moreover, we demonstrate that emerging \emph{thinking models} with extended reasoning capabilities are nearly immune to position bias, suggesting a costly but general approach toward resolving this vulnerability.
Chinese Translation
近期研究表明，位置偏差仍然是大语言模型（LLMs）中的一个关键脆弱性，它既损害了多项选择评估的完整性，也削弱了检索增强生成的可靠性。然而，现有对这一现象的理解主要来自模型覆盖范围有限且以英语为中心的基准研究，这使得前沿模型和多种语言中的行为仍未得到充分探索。在本工作中，我们对从十亿级到万亿级参数的不同模型规模下的位置偏差进行了全面研究，涵盖五种语言中的多项选择任务和多文档任务。我们的实验表明，尽管位置偏差在小规模模型和非英语语言中仍然明显，但增大模型规模并采用标准的思维链提示可以显著缓解这一问题。此外，我们证明，具有扩展推理能力的新兴 \emph{思维模型} 几乎不受位置偏差影响，这暗示了一种代价高昂但通用的解决这一脆弱性的方法。
cs.CL / 32 / 2610.22841
The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls
依赖假说：思维链忠实度随语言资源水平下降而上升
large language model
大语言模型相关
Abstract
Chain-of-Thought (CoT) faithfulness matters for monitoring, explaining and safely deploying large language models, yet it has been studied mostly in English. Languages differ in level, the amount of text available to learn from, so models know less in lower-resource languages, which may change how much they rely on their reasoning. To our knowledge, we are the first to relate CoT faithfulness to both resource level and task type (knowledge- or reasoning-intensive). On parallel multiple-choice questions from Multilingual Massive Multitask Language Understanding (MMMLU) in eight languages, faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive ones, a gap that points the same way across four models and several faithfulness measures. We propose, and find support for, the Reliance Hypothesis: models rely more on their reasoning when they cannot recall the answer, as in lower-resource languages on knowledge-intensive tasks. Model confidence partly explains the effect of resource level, but not the knowledge-reasoning gap. Languages in which models are less confident rely more on reasoning, yet less confident answers within a language do not, a Simpson's paradox. In lower-resource languages, models rely on their reasoning for more answers, not more strongly for each. CoT monitoring may therefore be most informative in lower-resource languages, where it is least often validated.
Chinese Translation
思维链（Chain-of-Thought, CoT）忠实度对于监测、解释和安全部署大语言模型至关重要，但此前的研究大多集中在英语中。不同语言在资源水平上存在差异，即可供学习的文本数量不同，因此模型在低资源语言中掌握的知识较少，这可能改变它们对其推理的依赖程度。据我们所知，我们是首个将 CoT 忠实度同时与资源水平和任务类型（知识密集型或推理密集型）联系起来的研究。在来自多语言大规模多任务语言理解（Multilingual Massive Multitask Language Understanding, MMMLU）的八种语言平行多项选择题上，在知识密集型任务中，忠实度随资源水平下降而上升，但在推理密集型任务中并非如此；这一差距在四个模型和若干忠实度度量上指向同一方向。我们提出并找到了对“依赖假说”的支持：当模型无法回忆答案时，它们会更依赖自己的推理，正如在知识密集型任务中的低资源语言里那样。模型置信度部分解释了资源水平的影响，但不能解释知识与推理之间的差距。在模型置信度较低的语言中，模型更依赖推理，然而在同一语言内，置信度较低的答案却并非如此，这是一种辛普森悖论。在低资源语言中，模型是对更多答案依赖其推理，而不是对每一个答案依赖得更强。因此，CoT 监测在低资源语言中可能信息量最大，而它在那里却最不常得到验证。
cs.CL / 33 / 2610.22858
Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search
用于高效 LLM 树搜索的不确定性感知预算分配
large language model
大语言模型相关
Abstract
Tree search enhances multi-step reasoning in large language models (LLMs), yet prevailing methods expand every step with a fixed budget, irrespective of the marginal utility of expansion. We first show, across existing search algorithms, that continuations sampled from a shared intermediate state are largely duplicates and collapse to a few distinct answers. Critically, we find that the semantic dispersion of sibling actions predicts the redundancy of their downstream continuations, permitting a subtree's budget to be fixed before expansion. An information-theoretic analysis motivates this: treating the per-step budget as the Monte Carlo sample size for estimating a step's residual answer entropy, we derive a sufficient budget that shrinks with sibling dispersion. We accordingly propose $\textbf{Uncertainty-Aware Allocation}$ (UAA), a training-free framework that decouples value from uncertainty: values determine which candidates are promising, while uncertainty, estimated from sibling dispersion at no additional model cost, determines how many merit expansion, via uncertainty-aware pruning and uncertainty-aware budgeting. Across mathematical reasoning, logical reasoning, and planning benchmarks, UAA reduces token usage by up to $90\%$ and latency by up to $9\times$ at comparable accuracy, generalizing across four tree search algorithms and both open- and closed-source LLMs.
Chinese Translation
树搜索增强了大语言模型（LLMs）中的多步推理，但现有方法以固定预算扩展每一步，而不考虑扩展的边际效用。我们首先表明，在现有搜索算法中，从共享中间状态采样得到的延续在很大程度上是重复的，并坍缩为少数几个不同答案。关键的是，我们发现兄弟动作的语义离散度预测其下游延续的冗余度，从而允许在扩展之前固定一个子树的预算。一项信息论分析为此提供了动机：将每一步的预算视为用于估计某一步剩余答案熵的蒙特卡洛样本量，我们推导出一个随兄弟离散度增大而收缩的充分预算。我们据此提出 $\textbf{Uncertainty-Aware Allocation}$（UAA），一个免训练框架，将价值与不确定性解耦：价值决定哪些候选有前景，而不确定性——以不增加额外模型成本的方式由兄弟离散度估计——通过不确定性感知剪枝和不确定性感知预算分配，决定有多少候选值得扩展。在数学推理、逻辑推理和规划基准上，UAA 在准确率相当的情况下将 token 使用量降低最多 $90\%$，将延迟降低最多 $9\times$，并可泛化到四种树搜索算法以及开源和闭源 LLMs。
cs.CL / 34 / 2610.22889
NLData2Opt: Benchmarking Parameter Derivation and Prediction for Modeling Optimization Problems from Natural Language and Data
NLData2Opt：面向从自然语言与数据出发的优化问题建模，对参数推导与预测进行基准评测
large language model
大语言模型相关
Abstract
Formulating optimization problems from raw inputs of problem descriptions and data is time-consuming and costly. While Large Language Models (LLMs) offer a promising way to automate this process, existing benchmarks overestimate their capabilities by relying on idealized inputs where all data is relevant and complete. This diverges from reality, where data-centric reasoning is required to extract, compute, and predict parameters from noisy sources. To address this, we introduce NLData2Opt, a benchmark to evaluate LLMs on the multi-stage reasoning inherent in practical optimization workflows. NLData2Opt decomposes problems into problem descriptions and data, and requires models to (1) extract relevant parameters from data containing misleading attributes, (2) compute missing parameters derived by transforming existing data attributes, and (3) predict unavailable future values from historical observations. Evaluation results show substantial performance decline relative to existing benchmarks, indicating that benchmark progress does not directly translate to formulation accuracy in practical optimization workflows. NLData2Opt is available at https://github.com/sumarivyn/NLData2Opt.
Chinese Translation
从问题描述和数据的原始输入出发构建优化问题表述既耗时又成本高昂。尽管大语言模型（LLMs）为自动化这一过程提供了有前景的途径，但现有基准依赖理想化输入——即所有数据均相关且完整——从而高估了它们的能力。这与现实情况不符：现实中需要以数据为中心的推理，从含噪来源中抽取、计算并预测参数。为解决这一问题，我们提出 NLData2Opt，一个用于评估 LLMs 在实际优化工作流所固有的多阶段推理方面表现的基准。NLData2Opt 将问题分解为问题描述和数据，并要求模型：(1) 从包含误导性属性的数据中抽取相关参数；(2) 计算通过变换已有数据属性而推导出的缺失参数；(3) 根据历史观测预测不可用的未来值。评估结果表明，相对于现有基准，性能出现显著下降，这表明基准上的进展并不能直接转化为实际优化工作流中的问题表述准确率。NLData2Opt 已在 https://github.com/sumarivyn/NLData2Opt 公开。
cs.CL / 35 / 2610.22894
Competence-Based Evaluation of Logical Reasoning in Large Language Models
大型语言模型中逻辑推理的基于能力评估
large language model
大语言模型相关
Abstract
Recent advances in large language models (LLMs) have brought impressive performance across diverse tasks; however, shortcomings remain in logical reasoning, including inconsistencies under linguistic variation. To improve the reasoning capacity of LLMs, we must be able to accurately measure their reasoning capacity. Current benchmarks primarily report single accuracy numbers on selected reasoning tasks. Such a performance-based evaluation methodology says little about a model's generalization capacity in reasoning. To address this limitation, we argue for a different evaluation paradigm, called competence-based evaluation (CBE), which tests whether models reliably apply the same competence across a family of problems. A parameterized generator scales complexity while keeping the required competence fixed, and matched answer-preserving presentations test correctness beyond one formulation. We instantiate CBE on elementary ordering tasks, scaling input size while varying presentations of the same ordering rule. Across six GPT and Qwen models, performance is reasonably good over bounded complexity ranges, but even strong models fail the competence condition across the tested family. Chain-of-thought and GRPO extend strong performance without resolving all tested failures; supervised LoRA shows little held-out transfer. CBE provides diagnostic profiles and an evaluation protocol that can grow with future models.
Chinese Translation
大型语言模型（LLMs）的近期进展已在多种任务上带来令人印象深刻的性能；然而，在逻辑推理方面仍存在不足，包括在语言表述变化下的一致性问题。为了提升 LLMs 的推理能力，我们必须能够准确测量其推理能力。当前的基准测试主要在选定的推理任务上报告单一准确率数值。这种基于性能的评估方法几乎无法说明模型在推理方面的泛化能力。为解决这一局限，我们主张采用一种不同的评估范式，称为基于能力的评估（CBE），它检验模型是否能够在一系列问题上可靠地应用同一种能力。参数化生成器在保持所需能力固定的同时扩展复杂度，而匹配的、保持答案不变的呈现方式则在单一表述之外检验正确性。我们在基本排序任务上实例化 CBE，在扩展输入规模的同时改变同一排序规则的呈现方式。在六个 GPT 和 Qwen 模型上，性能在有界复杂度范围内相当不错，但即便是强模型也未能在所测试的问题族上满足能力条件。思维链和 GRPO 扩展了强性能，但并未解决所有受测试的失败；监督式 LoRA 显示出很小的留出迁移。CBE 提供了诊断性画像以及一种可随未来模型发展而扩展的评估协议。
cs.CL / 36 / 2610.22940
Beyond Surface Cues: Robust LLM-Generated Korean Text Detection
超越表面线索：鲁棒的LLM生成韩语文本检测
large language model
大语言模型相关
Abstract
As large language models (LLMs) generate increasingly fluent text, the need for reliable detection of LLM-generated content has grown substantially. However, existing research remains heavily English-centric, leaving Korean largely underexplored. Moreover, current evaluations of Korean text detection often rely on superficial cues, leading to overly optimistic estimates of detection performance. To address these limitations, we introduce \textsc{KoLing}, a bias-controlled adversarial benchmark comprising 50.9k instances that suppresses superficial shortcuts while preserving semantic content and separately evaluates LLM-generated and LLM-modified text detection. We further propose KoLingProbe, an interpretable detector grounded in theoretical linguistics that models Korean-specific properties across multiple linguistic levels. Experiments show that KoLingProbe substantially outperforms strong baselines on all detection tasks while remaining robust to adversarial attacks. Our analyses further reveal that integrating complementary linguistic features across multiple levels yields both improved performance and interpretability.
Chinese Translation
随着大语言模型（LLMs）生成越来越流畅的文本，对LLM生成内容进行可靠检测的需求已大幅增长。然而，现有研究仍严重以英语为中心，使韩语在很大程度上未得到充分探索。此外，当前对韩语文本检测的评估往往依赖表面线索，导致对检测性能的估计过于乐观。为解决这些局限，我们提出 \textsc{KoLing}，一个包含50.9k个实例的偏置控制对抗基准，它抑制表面捷径，同时保留语义内容，并分别评估LLM生成文本和LLM修改文本的检测。我们进一步提出 KoLingProbe，一个基于理论语言学的可解释检测器，它在多个语言层面上对韩语特有的属性进行建模。实验表明，KoLingProbe 在所有检测任务上大幅优于强基线，同时对对抗攻击保持鲁棒。我们的分析进一步揭示，整合跨多个层面的互补语言特征能够同时提升性能和可解释性。
cs.CL / 37 / 2610.22942
Aligning Language Models Towards Counterfactual Invariance
将语言模型对齐到反事实不变性
large language model
大语言模型相关
Abstract
Small variations in prompt characteristics can meaningfully alter the semantics, tone, and contextual assumptions of large language model outputs. Such characteristics can involve group variables such as demographics or other sensitive latent attributes. Sensitivity to such input variables is undesirable when demographic information is causally irrelevant to the task, but might still be relevant for generation of semantically coherent responses. This makes the task of enforcing counterfactual invariance while preserving model utility difficult. To study this problem, we first show that standard model alignment recipes (such as reinforcement learning) increase differences in response quality across demographics, while improving overall response quality, ultimately amplifying what is known as counterfactual bias. To address this, we develop a policy gradient technique involving a group-level reward function that operates on matched pairs of counterfactual prompts. For completeness, we build a dataset of controlled counterfactual prompt pairs and design an evaluation that separately measures cases where demographic information is causally relevant to the answer. Finally, we validate our findings on several fairness benchmarks. We demonstrate that our methods reduce counterfactual differences while preserving overall response quality and that the models continue to adapt appropriately when such sensitive information is relevant.
Chinese Translation
提示特征的微小变化可能会显著改变大语言模型输出的语义、语气和上下文假设。此类特征可能涉及群体变量，例如人口统计信息或其他敏感潜在属性。当人口统计信息与任务因果无关时，对此类输入变量的敏感性是不可取的，但这些信息对于生成语义连贯的响应仍可能相关。这使得在保持模型效用的同时强制实现反事实不变性的任务变得困难。为研究这一问题，我们首先表明，标准模型对齐方法（例如强化学习）会在提高整体响应质量的同时，增大不同人口统计群体之间响应质量的差异，最终放大所谓的反事实偏差。为解决这一问题，我们开发了一种策略梯度技术，涉及一个组级奖励函数，该函数作用于匹配的反事实提示对。为完整起见，我们构建了一个受控反事实提示对数据集，并设计了一种评估，分别测量人口统计信息与答案因果相关的情形。最后，我们在多个公平性基准上验证了我们的发现。我们证明，我们的方法在减少反事实差异的同时保持整体响应质量，并且当此类敏感信息相关时，模型仍能继续适当适应。
cs.CL / 38 / 2610.22961
What Shapes Brain Alignment of Large Language Models during Dialogue? Insights from Language Production and Comprehension
是什么塑造了对话过程中大型语言模型的大脑对齐？来自语言产生与理解的洞见
large language model
大语言模型相关
Abstract
Brain alignment of large language models (LLMs) offers a way to investigate representational similarities between LLMs and the human brain during language processing. While previous studies have mainly focused on passive language comprehension, little is known about such alignment during dialogue, in which speakers engage in both language production and comprehension. In this study, using fMRI data collected during dialogue, we examine brain alignment during language production and comprehension across LLM layers and brain regions. We assess its association with the intrinsic dimension (ID) of LLM representations and the contributions of several interpretable linguistic and communicative features. Our results reveal distinct patterns of brain alignment between production and comprehension, as well as strong associations with ID in brain regions associated with each processing mode. Furthermore, we show that the contributions of linguistic and communicative features also vary across processing modes, LLM layers, and brain regions.
Chinese Translation
大型语言模型（LLMs）的大脑对齐为研究LLMs与人脑在语言加工过程中的表征相似性提供了一种途径。尽管以往研究主要关注被动的语言理解，但人们对于对话过程中的这种对齐却知之甚少，而在对话中，说话者同时参与语言产生和理解。在本研究中，我们利用对话期间收集的fMRI数据，考察了跨LLM层和脑区在语言产生与理解过程中的大脑对齐。我们评估了其与LLM表征的内在维度（ID）的关联，以及若干可解释的语言和交际特征的贡献。我们的结果揭示了语言产生与理解之间不同的大脑对齐模式，以及与每种加工模式相关脑区中ID的强关联。此外，我们还表明，语言和交际特征的贡献也因加工模式、LLM层和脑区的不同而有所差异。
cs.CL / 39 / 2610.22980
KiranaBench: Can a Language Model Run a Neighbourhood Shop?
KiranaBench：语言模型能经营一家社区小店吗？
large language model
大语言模型相关
Abstract
Large language model (LLM) agents are increasingly evaluated on long-horizon tasks, including autonomously running a business. Existing benchmarks for this setting are almost exclusively English-language, Western in retail context, and single-agent, with the agent observing the business directly. We study how these results transfer to a multi-agent Indic setting. We introduce KiranaBench, a simulated Indian general store in which four roles (manager, cashier, helper, and bookkeeper) are each played by a separate LLM agent. Interactions with customers and suppliers occur in English, Hindi, or Hinglish, and product demand is calibrated to India's national Household Consumption Expenditure Survey. Evaluating four open-weight LLMs, we find that all fall far short of a fixed-rule shopkeeper baseline: the best model earns only one-eighth of the baseline's profit, and two models finish below an unmanaged shop. Performance is also consistently lower for Hindi and Hinglish than for English supplier communication. KiranaBench offers evidence that current LLM agents are not yet reliable for autonomously operating small businesses in India, and provides a testbed for developing agents suited to settings that are long-horizon, collaborative, and multilingual.
Chinese Translation
大语言模型（LLM）智能体越来越多地在长时程任务上接受评估，其中包括自主经营一家企业。针对这一场景的现有基准几乎全部是英语的、零售情境为西方的、且为单智能体的，智能体直接观察业务。我们研究这些结果如何迁移到多智能体的印度语境场景中。我们提出 KiranaBench，一个模拟的印度杂货店，其中四个角色（经理、收银员、帮工和记账员）各由一个独立的 LLM 智能体扮演。与顾客和供应商的交互以英语、印地语或印地英语进行，商品需求依据印度全国住户消费支出调查进行校准。在评估四个开放权重 LLM 后，我们发现它们全都远远达不到固定规则店主的基线水平：最佳模型仅获得基线利润的八分之一，而有两个模型的最终结果甚至低于无人管理的店铺。在供应商沟通方面，印地语和印地英语的表现也始终低于英语。KiranaBench 提供了证据表明当前 LLM 智能体尚不能可靠地在印度自主经营小企业，并为开发适用于长时程、协作式和多语言场景的智能体提供了一个测试平台。
cs.CL / 40 / 2610.22984
Self-Evolving Agents Can Repair Failures but Struggle to Generalize from Them
自演化智能体能够修复失败，却难以从中泛化
large language model
大语言模型相关
Abstract
Experience-driven self-evolution offers a promising path for large language model (LLM) agents to continually improve through interaction. Yet existing approaches often rely on successful trajectories or external evaluation signals, leaving unclear whether agents can evolve when such positive anchors are unavailable. We study this question through Cold-Start Failure-Only Evolution (CFOE), where agents start without accumulated experience and learn only from failed interactions without privileged guidance. Across three self-evolving paradigms, multiple LLM backbones, and four interactive environments, we uncover a consistent repair-generalization gap: failure-derived experience can help repair previously failed tasks, but rarely yields sustained improvements on unseen tasks. Further analysis shows that agents often misinterpret failures, while even verified-correct experience remains difficult to transfer. Relaxing CFOE reveals that verifier feedback and additional interaction provide limited gains, whereas successful trajectories substantially improve transfer. These findings suggest that current self-evolving agents are better at repairing observed failures than building generalizable capabilities from failure.
Chinese Translation
经验驱动的自演化为大语言模型（LLM）智能体通过交互持续改进提供了一条有前景的路径。然而，现有方法往往依赖成功轨迹或外部评估信号，这使得当此类正向锚点不可用时智能体能否演化仍不清楚。我们通过冷启动仅失败演化（CFOE）研究这一问题，其中智能体在起步时没有积累的经验，并且仅从失败交互中学习而没有特权引导。在三种自演化范式、多个LLM骨干模型和四个交互环境中，我们揭示出一致的修复—泛化差距：由失败衍生的经验可以帮助修复先前失败的任务，但很少能在未见任务上产生持续改进。进一步分析表明，智能体常常误解失败，而即便是经过验证正确的经验也仍然难以迁移。放宽CFOE表明，验证器反馈和额外交互带来的收益有限，而成功轨迹则显著改善迁移。这些发现表明，当前的自演化智能体更擅长修复已观察到的失败，而不太擅长从失败中构建可泛化的能力。
cs.CL / 41 / 2610.22985
Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference
棱镜注意力：用于高效 LLM 推理的候选精炼索引共享机制
large language model
大语言模型相关
Abstract
Long-context decoding in large language models (LLMs) is increasingly bottlenecked by the memory and computation required to attend over an ever-growing Key-Value (KV) cache at each autoregressive step. Although token-level sparse attention reduces this burden by restricting computation to a small subset of relevant past tokens, its practical gains are often offset by repeatedly recomputing layer-wise sparse indices as the context grows. A common remedy is to share sparse indices across layers by computing them only at designated anchor layers, but such static reuse ignores the layer-wise evolution of attention patterns, causing the reused indices to progressively deviate from the layer-specific optima and incur substantial attention loss. By revisiting cross-layer sparsity from the perspective of attention dynamics, we find that inter-layer salient token shift is highly localized: although the exact top-$κ$ indices differ across layers, the core high-attention tokens remain stable, with changes concentrated near the expanded boundary of previously selected regions. Motivated by this, we propose Prism Attention, a proposal-refined sparse attention framework. It replaces rigid cross-layer index reuse with a coarse-to-fine strategy, preserving layer-wise adaptability while retaining the efficiency benefits of cross-layer sharing. Extensive experiments show that Prism Attention achieves state-of-the-art performance among mainstream sparse attention methods on both reasoning and long-context benchmarks, while attaining a $3.4\times$ end-to-end decoding speedup over FlashAttention at 128K context length.
Chinese Translation
大语言模型（LLM）中的长上下文解码正日益受到瓶颈制约，其原因在于每个自回归步骤中都需要对不断增长的键值（KV）缓存进行注意力计算，由此带来巨大的内存与计算开销。尽管 token 级稀疏注意力通过将计算限制在一小部分相关的过去 token 上缓解了这一负担，但随着上下文增长，逐层稀疏索引的反复重计算往往会抵消其实际收益。一种常见的补救方法是仅在指定的锚点层计算稀疏索引，并在各层之间共享这些索引，但这种静态复用忽略了注意力模式的逐层演化，导致被复用的索引逐渐偏离各层特有的最优值，并造成显著的注意力损失。通过从注意力动态特性的视角重新审视跨层稀疏性，我们发现层间显著 token 的偏移高度局部化：尽管精确的 top-$κ$ 索引在各层之间有所不同，但核心的高注意力 token 保持稳定，变化集中在先前选定区域的扩展边界附近。受此启发，我们提出了 Prism Attention，一种候选精炼的稀疏注意力框架。它用由粗到细的策略取代了僵化的跨层索引复用，在保留逐层适应性的同时保留了跨层共享的效率优势。大量实验表明，Prism Attention 在推理与长上下文基准上均取得了主流稀疏注意力方法中的最先进性能，同时在 128K 上下文长度下相较 FlashAttention 实现了 $3.4 imes$ 的端到端解码加速。
cs.CL / 42 / 2610.22999
Controlling the Diversity-Quality Trade-Off of LLMs by Merging Models Before and After Alignment
通过合并对齐前后的模型来控制大语言模型的多样性—质量权衡
large language model
大语言模型相关
Abstract
Alignment is known to change the output characteristics of large language models (LLMs). Base models tend to generate more diverse outputs, whereas aligned models tend to produce higher-quality outputs. The appropriate balance between the two depends on the task. Existing methods either require costly fine-tuning or, like temperature sampling, mainly improve lexical diversity rather than semantic diversity. We introduce a depth-aware FrankenMerging approach for controlling the diversity--quality trade-off between an aligned model and its base model without additional training. This approach combines the lower layers of an aligned model with the higher layers of its base model based on the hypothesis that lexical diversity is more closely associated with lower layers and semantic diversity with higher layers. Experiments showed that varying the number of layers taken from each model adjusts both diversity and quality. In most cases, the merged models exhibited higher semantic diversity than the aligned model, and at some split positions, they showed higher lexical diversity than both original models. They also demonstrated a better diversity--quality trade-off than temperature sampling in most settings. A supporting ablation, in which we reversed the merging order, yielded results largely consistent with the hypothesis.
Chinese Translation
已知对齐会改变大语言模型（LLMs）的输出特征。基础模型倾向于生成更多样化的输出，而对齐模型倾向于产生更高质量的输出。二者之间的适当平衡取决于任务。现有方法要么需要昂贵的微调，要么像温度采样那样主要改善词汇多样性而非语义多样性。我们引入一种深度感知的FrankenMerging方法，用于在无需额外训练的情况下控制对齐模型与其基础模型之间的多样性—质量权衡。该方法将对齐模型的较低层与其基础模型的较高层相结合，其依据的假设是：词汇多样性与较低层更密切相关，而语义多样性与较高层更密切相关。实验表明，改变从每个模型中选取的层数会同时调节多样性和质量。在大多数情况下，合并模型表现出比对齐模型更高的语义多样性，并且在某些切分位置上，它们表现出比两个原始模型都更高的词汇多样性。在大多数设置下，它们还表现出比温度采样更好的多样性—质量权衡。一项支持的消融实验——其中我们反转了合并顺序——得到的结果与假设基本一致。
cs.CL / 43 / 2610.23060
Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses
同一个人，不同的建议：用重复风险容忍度回答审计LLM财务建议
large language model
大语言模型相关
Abstract
AI financial advisers set a client's stock allocation largely from one stated-risk question, yet people do not answer it the same way twice: in a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two or more points on a 0-10 scale after six months. Seven large language models from five providers gave one-turn allocation advice for 800 real household profiles in which only the risk answer changed. The recommended stock share moved by 10 to 16 percentage points on average, and by 24 to 35 when the answer changed, against 0.6 to 4.2 points for the same prompt asked twice; in the laboratory retest the rating alone moved advice by 11 to 18 points. Without any note, the models respond to the answer about as strongly as when told that it is very reliable (less so for the two Claude models). If that instructed response is how each model maps attitude to advice, discounting the answer for its measured reliability, as classical test theory prescribes, would cut the swings by a factor of 2.4 to 3.3. The swing is largest for clients under 40. A note that the answer is noisy steadies some models, but shifts their weight to the client's age and pulls cautious clients towards a generic middle allocation: clients unwilling to take any risk are advised 22% stocks instead of 4%. Three extra money questions reduce the swing for some models, with no statistically detectable change relative to how much advice differs across people. In a scripted test, one clarifying question removes nearly all of the swing from an answer the client then corrects (33.5 points to 1.9 on average) while keeping the response to a change the client confirms. We argue that same-person retest audits should be a standard check for AI advisers, and that what to do with an unreliable answer is a question of client autonomy as well as of statistics.
Chinese Translation
AI财务顾问主要根据一个陈述的风险问题来设定客户的股票配置，然而人们对同一个问题并不会两次给出相同答案：在美国一个全国性面板中，41%的家庭在两年后选择了不同的答案；在实验室重测中，三分之一的人在六个月后在0-10量表上移动了两个或更多点。来自五家提供商的七个大语言模型针对800个真实家庭画像给出单轮配置建议，其中只有风险答案发生变化。推荐股票份额平均变动10至16个百分点；当答案发生变化时变动24至35个百分点，而同一提示询问两次的变动为0.6至4.2个百分点；在实验室重测中，仅评分本身就使建议变动11至18个百分点。在没有任何说明的情况下，模型对答案的反应强度与被告知该答案非常可靠时大致相当（两个Claude模型则较弱）。如果这种在指令下的反应就是每个模型将态度映射为建议的方式，那么按照经典测验理论的规定，根据测得的可靠性对答案进行折扣，将使波动幅度降低2.4至3.3倍。这种波动在40岁以下的客户中最大。提示答案存在噪声的说明会使一些模型趋于稳定，但会把它们的权重转移到客户年龄上，并将谨慎的客户拉向一个通用的中间配置：不愿承担任何风险的客户被建议配置22%的股票，而不是4%。三个额外的金钱问题减少了一些模型的波动，但相对于建议在不同人之间的差异程度，没有统计上可检测的变化。在一项脚本化测试中，一个澄清性问题几乎消除了来自客户随后更正的答案的全部波动（平均从33.5点降至1.9点），同时保留了对客户确认的变更的反应。我们认为，同一人重测审计应成为AI顾问的标准检查，而如何处理不可靠的答案既是统计问题，也是客户自主权问题。
cs.CL / 44 / 2610.23062
ParetoChat: An LLM-Based Interface for Interactive Multi-Objective Optimization
ParetoChat：一种基于LLM的交互式多目标优化接口
large language model
大语言模型相关
Abstract
Multi-objective optimization models are often deployed in settings where decision makers (DMs) must balance some conflicting criteria. In practice, however, stakeholders must translate their qualitative preferences between different criteria into numerical weights, priorities, or reference points required by optimization methods, which can be difficult to specify and interpret. This paper proposes \emph{ParetoChat}, an LLM (large language model)-based expert-agent interface for interactive multi-objective optimization. The agent translates natural-language user input into structured scalarization parameters, invokes an optimization system, presents candidate solutions, and iteratively elicits feedback. We describe a multi-agent architecture that separates input processing, multi-criteria decision-making (MCDM) expertise, solution presentation, and feedback processing. We then present a simulation study using an LLM-based DM agent and benchmark ParetoChat against interactive methods implemented in DESDEO, including reference-point and NIMBUS methods. The computational experiments on multi-objective binary knapsack and set-partitioning instances show that ParetoChat can guide the search efficiently toward solutions that match the DM's preferences. An application to historical courier-order assignment instances illustrates how users can explore objective extremes and intermediate trade-offs through natural-language requests. The results indicate that LLM agents may provide a practical interaction layer between business stakeholders and optimization tools.
Chinese Translation
多目标优化模型通常部署在决策者（DMs）必须平衡一些相互冲突的准则的场景中。然而在实践中，利益相关者必须将他们在不同准则之间的定性偏好转化为优化方法所需的数值权重、优先级或参考点，而这些参数可能难以指定和解释。本文提出\emph{ParetoChat}，一种基于LLM（大语言模型）的专家-智能体接口，用于交互式多目标优化。该智能体将自然语言用户输入转换为结构化标量化参数，调用优化系统，展示候选解，并迭代地征求反馈。我们描述了一种多智能体架构，该架构将输入处理、多准则决策（MCDM）专业知识、解展示和反馈处理分离。然后，我们展示了一项使用基于LLM的DM智能体的仿真研究，并将ParetoChat与DESDEO中实现的交互式方法进行基准比较，包括参考点方法和NIMBUS方法。在多目标二进制背包和集合划分实例上的计算实验表明，ParetoChat能够高效地引导搜索朝向与DM偏好匹配的解。一项针对历史快递订单分配实例的应用展示了用户如何通过自然语言请求探索目标极值和中间权衡。结果表明，LLM智能体可以在业务利益相关者与优化工具之间提供实用的交互层。
cs.CL / 45 / 2610.23071
Before Retrieval Begins: Safety and Tool-Use Decisions in LLM Agents
在检索开始之前：LLM 智能体中的安全与工具使用决策
large language model
大语言模型相关
Abstract
Large language models increasingly act as agents that call external tools, yet most tool-safety research examines only what follows a tool call. The prior decision of whether to call the tool remains largely unexamined and can influence downstream safety. Across the seven open-weight models, the unweighted model-average harm rate after admitted harmful requests is 74.5%. To identify what governs this decision, we apply a factorial intervention to the input request. Within our generated construction, neutralizing the requested-operation component produces the largest admission contrast (19 pp). A linear direction estimated from this contrast influences held-out admission in four of seven models; effects on operation-neutralized requests in some models show that it is not an operation-specific mechanism. We then test whether written refusal is behaviorally and representationally coupled to the same decision. Written refusal accompanies 33.6% of admitted harmful calls. Plain-text refusal training reduces unsafe admission in models that retain benign tool use, showing that response-level safety training can transfer to the call decision. Together, these results establish a partial, model-dependent relation between written refusal and tool admission. They motivate direct evaluation and supervision of tool admission rather than treating text refusal as its proxy.
Chinese Translation
大型语言模型越来越多地作为调用外部工具的智能体行事，然而大多数工具安全研究只考察工具调用之后发生的事情。关于是否调用工具的这一先在决策在很大程度上仍未得到考察，并且可能影响下游安全性。在七个开放权重模型中，在被承认的有害请求之后，未加权的模型平均危害率为 74.5%。为了识别是什么支配着这一决策，我们对输入请求施加了因子干预。在我们生成的构造中，中和所请求的操作成分产生了最大的准入对比（19 个百分点）。从这一对比中估计出的线性方向在七个模型中的四个里影响了留出的准入；在一些模型中对操作被中和的请求所产生的效应表明，它并不是一种操作特异性的机制。随后我们检验书面拒绝是否在行为上和表征上与同一决策相耦合。书面拒绝伴随了 33.6% 的被准入的有害调用。纯文本拒绝训练在保留良性工具使用的模型中减少了不安全的准入，表明响应层面的安全训练可以迁移到调用决策上。这些结果共同确立了书面拒绝与工具准入之间一种部分的、依赖于模型的关系。它们促使人们对工具准入进行直接评估和监督，而不是将文本拒绝视为其代理指标。
cs.CL / 46 / 2610.23157
Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models
Cross-Centaur：使用语言模型对跨心理学任务的个体决策进行建模
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used as computational models of human behavior. Recent models such as Centaur have advanced this goal by fine-tuning LLMs on trial-by-trial decisions from psychological experiments. However, existing large-scale behavioral datasets do not link the same participants across different tasks, so these models learn population-level regularities but are not trained to model individual-level regularities across tasks. We introduce Cross-Psych-101, a dataset of 591 participants and 261{,}551 valid decisions across 12 Bandit, Tradeoff, Structure, and Social tasks. We train Cross-Centaur to predict a participant's target-task actions from a target-task description and the source-task description paired with the same participant's source-task behavior. Cross-Centaur outperforms off-the-shelf LLMs and Centaur on held-out participants, both within and across tasks. Probing analyses suggest that its hidden states encode participant-level behavioral tendencies that transfer across tasks. These results highlight the potential of cross-task individual-level data for predicting human decision-making with language models. Cross-Pysch-101 is available at https://huggingface.co/datasets/HikaruAsano/cross-psych-101.
Chinese Translation
大语言模型（LLMs）正日益被用作人类行为的计算模型。近期的模型如 Centaur 通过对心理学实验中的逐试次决策进行微调，推进了这一目标。然而，现有的大规模行为数据集并未将同一参与者跨不同任务关联起来，因此这些模型学到的是群体层面的规律，而未被训练去建模跨任务的个体层面规律。我们提出了 Cross-Psych-101，这是一个包含 591 名参与者和 261{,}551 个有效决策的数据集，涵盖 Bandit、Tradeoff、Structure 与 Social 四类共 12 项任务。我们训练 Cross-Centaur，使其根据目标任务描述，以及源任务描述与同一参与者在源任务中行为的配对，来预测该参与者在目标任务中的动作。在留出的参与者上，Cross-Centaur 在任务内和跨任务两种设定下均优于现成的 LLMs 和 Centaur。探针分析表明，其隐藏状态编码了可跨任务迁移的参与者层面行为倾向。这些结果凸显了跨任务个体层面数据在利用语言模型预测人类决策方面的潜力。Cross-Pysch-101 可在 https://huggingface.co/datasets/HikaruAsano/cross-psych-101 获取。
cs.CL / 47 / 2610.23182
What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space
大语言模型在暗影中做什么：从解嵌入暗空间读取上下文依赖
large language model
大语言模型相关
Abstract
Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands noticeable overhead, such as materializing attention maps, running additional forward passes, or training auxiliary probes. In this work, we show that context reliance can be estimated by projecting hidden states onto the dark space - directions that compose weakly with the unembedding matrix. Beyond its recognized function as an initial-token attention sink, we show that this subspace actively mediates context processing, and that its share of hidden-state norm - termed the Dark Matter Ratio (DMR) - provides a direct score of context utilization. Pairing DMR with a parametric knowledge score yields DarkLens, a lightweight faithfulness detector operating strictly within a single forward pass. Evaluated across three QA and summarization benchmarks on six open-weight LLMs (4B-13B parameters), DarkLens matches or outperforms a wide range of baselines without requiring extra forward passes or probe training.
Chinese Translation
检测大语言模型（LLMs）中的忠实性幻觉对于其在高风险领域的安全部署至关重要。近期方法通过对比上下文知识与参数知识的利用来应对这一问题，然而测量上下文依赖通常需要明显的额外开销，例如具象化注意力图、运行额外的前向传播或训练辅助探针。在这项工作中，我们表明，可以通过将隐藏状态投影到暗空间——与解嵌入矩阵仅发生弱组合的方向——来估计上下文依赖。除了其作为初始 token 注意力汇的公认功能外，我们还表明该子空间主动介导上下文处理，并且其在隐藏状态范数中所占的份额——称为暗物质比例（DMR）——提供了上下文利用的直接得分。将 DMR 与参数知识得分配对，得到 DarkLens，一个严格在单次前向传播内运行的轻量级忠实性检测器。在六个开放权重 LLM（4B-13B 参数）上的三个问答与摘要基准上进行评估，DarkLens 在不需要额外前向传播或探针训练的情况下，匹配或优于广泛的基线。
cs.CL / 48 / 2610.23205
MAS-Bai: Linguistically Informed Adaptive Multi-Agent Translation for Bowen-Script Bai
MAS-Bai：面向方块白文的白语语言学信息驱动自适应多智能体翻译
large language model
大语言模型相关
Abstract
Large Language Models (LLMs) remain unreliable for extremely low-resource languages whose writing systems conflict with high-resource pretraining priors. We study Bai-Chinese translation in the Bowen script, where Chinese characters are frequently borrowed phonetically rather than semantically, causing severe literal-translation hallucinations. We propose MAS-Bai, a linguistically informed adaptive multi-agent framework for Bowen-script Bai-Chinese translation. MAS-Bai routes each input through either a lightweight retrieval-translation path or a deeper philological path with etymological and syntactic agents, and uses cycle-consistency with a lightweight reverse translator as an inference-time reliability signal. We construct a 6,536-sentence Bai-Chinese corpus and evaluate on a 1,536-sentence test set. MAS-Bai achieves 16.23 BLEU and 62.15 COMET, improving over Standard RAG by 8.11 BLEU and 6.85 COMET. Forward LoRA, expert ablations and controlled routing experiments distinguish the roles of supervised adaptation, Bai-specific expertise and conditional routing. The largest gains occur on Bowen phonetic loans and divergent Bai word order, while human evaluation and reverse-verifier analyses support the automatic results.
Chinese Translation
大型语言模型（LLMs）对于书写系统与高资源预训练先验相冲突的极低资源语言仍然不可靠。我们研究方块白文（Bowen script）中的白汉翻译，其中汉字常常是借音而非借义，从而导致严重的直译幻觉。我们提出 MAS-Bai，一种面向方块白文白汉翻译的语言学信息驱动自适应多智能体框架。MAS-Bai 将每个输入路由至轻量级检索翻译路径，或路由至包含词源智能体与句法智能体的更深层文献学路径，并使用与轻量级反向翻译器的循环一致性作为推理时的可靠性信号。我们构建了一个包含 6,536 个句子的白汉平行语料库，并在包含 1,536 个句子的测试集上进行评估。MAS-Bai 取得 16.23 BLEU 与 62.15 COMET，相较 Standard RAG 提升 8.11 BLEU 与 6.85 COMET。前向 LoRA、专家消融与受控路由实验区分了监督式适配、白语特定专业知识与条件路由各自的作用。最大增益出现在方块白文借音词与差异性的白语语序上，而人工评估与反向验证器分析亦支持了自动评估结果。
cs.CL / 49 / 2610.23225
ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling
ReLaDrift：通过 Drifting 实现循环语言生成以进行测试时扩展
diffusion
扩散模型相关
Abstract
Drifting Models generate a sample with a single function evaluation, making them attractive for fast parallel language generation, but their one-step formulation does not support iterative refinement. We propose Recurrent Language Generation via Drifting (ReLaDrift), which uses a shared Drifting Model to generate an initial draft from noise and iteratively refine it in a continuous latent space. We train this recurrence by sampling the number of rounds and applying the drifting loss only at the final round, stopping gradients through preceding rounds so that only one round requires backpropagation. Unlike the one-step Drifting Model, whose accuracy does not improve when iterated, ReLaDrift enables a trade-off between test-time computation and generation quality, as diffusion and flow language models do. On Sudoku Hard and GSM8K, it achieves competitive or superior performance compared with these iterative-generation approaches across a range of inference budgets. Our code is available at https://github.com/Yuki-Ots/ReLaDrift
Chinese Translation
Drifting 模型通过单次函数评估生成一个样本，这使其对于快速并行语言生成具有吸引力，但其一步式形式不支持迭代细化。我们提出通过 Drifting 的循环语言生成（ReLaDrift），它使用一个共享的 Drifting 模型从噪声生成初始草稿，并在连续潜在空间中迭代细化它。我们通过采样轮数并仅在最后一轮应用 drifting 损失来训练这种循环，同时停止通过前几轮的梯度，使得只有一轮需要反向传播。与一步式 Drifting 模型不同——后者在迭代时精度不会提高——ReLaDrift 能够像扩散模型和流语言模型那样，在测试时计算量与生成质量之间实现权衡。在 Sudoku Hard 和 GSM8K 上，在一系列推理预算下，与这些迭代生成方法相比，它取得了有竞争力或更优的性能。我们的代码可在 https://github.com/Yuki-Ots/ReLaDrift 获取。
cs.CL / 50 / 2610.23249
Who Pays for Efficient Inference? Confidence-Based Small-to-Large LLM Routing Across Tasks and Languages
谁为高效推理买单？跨任务与语言的基于置信度的小模型到大模型LLM路由
large language model
大语言模型相关
Abstract
Small-to-large language model routing can reduce inference cost by escalating difficult inputs while retaining easier ones on a smaller model. Its success, however, depends on identifying examples that benefit from escalation, and the reliability of confidence across multilingual tasks remains unclear. We present a systematic evaluation of confidence-based routing across three model families, four languages, and three task types, covering 72 model, dataset, language, and scale configurations and 42,288 predictions. We compare global confidence routing with random routing, equal-quota confidence routing, and an oracle upper bound. Confidence identifies beneficial escalations on the Belebele reading comprehension and SIB-200 topic classification datasets, with mean correction-detection AUCs of 0.612 and 0.705. At 50% escalation, it improves over random routing by 1.40--4.67 percentage points on Belebele and 2.54--5.63 points on SIB-200. In contrast, the mean AUC on the MMLU-ProX-Lite knowledge and reasoning dataset is only 0.510, and routing provides no consistent improvement over random selection because small models often remain highly confident while incorrect. Global confidence routing also creates language escalation gaps up to 30.33 percentage points, whereas equal-quota routing removes this disparity with little aggregate accuracy cost. These results show that confidence routing is task-dependent and can allocate computation unevenly across languages. Code is available at https://github.com/sabbir5622r/multilingual-llm-routing
Chinese Translation
从小模型到大模型的语言模型路由可以通过将困难输入升级到大模型、同时将较容易的输入保留在较小模型上，来降低推理成本。然而，其成功取决于识别能够从升级中受益的样本，而置信度在多语言任务中的可靠性仍不清楚。我们提出了一项针对基于置信度的路由的系统性评估，涵盖三个模型家族、四种语言和三种任务类型，覆盖72种模型、数据集、语言和规模配置以及42,288条预测。我们将全局置信度路由与随机路由、等配额置信度路由以及一个预言机上界进行比较。在Belebele阅读理解数据集和SIB-200主题分类数据集上，置信度能够识别出有益的升级，平均纠正检测AUC分别为0.612和0.705。在50%升级率下，它在Belebele上相较随机路由提升了1.40--4.67个百分点，在SIB-200上提升了2.54--5.63个百分点。相比之下，在MMLU-ProX-Lite知识与推理数据集上的平均AUC仅为0.510，并且路由相较随机选择没有提供一致的改进，因为小模型常常仍然高度自信但却是错误的。全局置信度路由还会造成高达30.33个百分点的语言升级差距，而等配额路由在几乎没有总体准确率代价的情况下消除了这种差异。这些结果表明，置信度路由依赖于任务，并且会在不同语言之间不均衡地分配计算量。代码可在 https://github.com/sabbir5622r/multilingual-llm-routing 获取。
cs.CL / 51 / 2610.23344
Language Models Learn to Improve Language Models
语言模型学习改进语言模型
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used to automate AI research, including language-model (LM) development, offering a potential path toward recursive self-improvement. However, turning experimental feedback into a learned ability to propose effective improvements across models faces two challenges. First, training such an ability with real feedback is expensive: reinforcement learning (RL) requires evaluating tens of thousands of proposals, each potentially requiring a full model-training run. Even at ten GPU-hours per run, reward collection alone would consume hundreds of thousands of GPU-hours, motivating a cheaper surrogate. Second, surrogate feedback can become unreliable as policy optimization favors overestimated proposals when the proposal distribution shifts. To address these challenges, we propose LM2LM, which trains LLMs with RL to generate effective LM improvement proposals. To provide scalable feedback, we use a research world model (RWM), a frozen LLM that predicts proposal gains from real experimental records provided in context, as the reward model for RL. To maintain feedback reliability, we periodically execute current-policy proposals ranked highly by the RWM and add their measured outcomes to its context. Across eight autoregressive and diffusion LM development environments, real-execution evaluations on held-out tasks show that LM2LM enables Qwen3.8-27B to outperform GPT-6-sol, even when GPT-6-sol is given 300 real experimental records in context, while requiring 33% lower data-generation cost. LM2LM also generalizes to a model family unseen during training.
Chinese Translation
大语言模型（LLMs）正越来越多地被用于自动化人工智能研究，包括语言模型（LM）开发，这提供了一条通向递归自我改进的潜在路径。然而，将实验反馈转化为一种习得的、能够跨模型提出有效改进建议的能力，面临两个挑战。首先，使用真实反馈训练这种能力代价高昂：强化学习（RL）需要评估数万个提议，而每个提议都可能需要一次完整的模型训练运行。即使每次运行只需十个 GPU 小时，仅奖励收集就会消耗数十万个 GPU 小时，这促使人们采用更廉价的代理。其次，当提议分布发生偏移时，策略优化会偏向被高估的提议，因此代理反馈可能变得不可靠。为应对这些挑战，我们提出 LM2LM，它使用 RL 训练 LLMs 生成有效的 LM 改进提议。为了提供可扩展的反馈，我们使用研究世界模型（RWM）作为 RL 的奖励模型；RWM 是一个冻结的 LLM，它根据上下文中提供的真实实验记录预测提议的收益。为了保持反馈的可靠性，我们定期执行由 RWM 排名靠前的当前策略提议，并将其测量结果添加到其上下文中。在八个自回归和扩散 LM 开发环境中，对留出任务的真实执行评估表明，LM2LM 使 Qwen3.8-27B 能够优于 GPT-6-sol，即使 GPT-6-sol 在上下文中被提供了 300 条真实实验记录，同时所需的数据生成成本低 33%。LM2LM 还能泛化到训练期间未见过的模型家族。
cs.CL / 52 / 2610.23400
On the Tip of the Tongue: First-Token Cueing as a Behavioral Probe of Recall Failures in LLMs
话到嘴边：首词元提示作为大语言模型（LLMs）中回忆失败的行为探针
large language model
大语言模型相关
Abstract
Large language models often answer factual questions incorrectly in closed-book settings, and such errors are usually attributed to missing knowledge. Yet models often encode answers they fail to produce. We show that many closed-book errors are first-token recall failures: the model commits to a wrong first token, but once given the right one it completes the answer and often recognizes it as correct. We study this with First-Token Cueing, a behavioral probe that prefills the response with the first token of the gold answer. Model uncertainty peaks at the first answer token, where it already predicts correctness, and among gold prefixes of any length the first token alone accounts for the largest share of recovered errors. Across 16 open models, this one token corrects between a fifth and a half of the errors on multi-token answers in English QA and improves accuracy in all seven languages of a multilingual benchmark. Most of the errors it corrects are also solved by the model's own sampled reasoning, a criterion for encoded knowledge, and models recognize correct starts they do not produce: they continue a correct start more confidently than an incorrect start of the same answer type and prior probability, yet rarely find it without the gold answer. These results locate a recoverable class of factual errors at the first decoding step, where inference-time methods may gain the most.
Chinese Translation
大语言模型在闭卷设置下经常错误地回答事实性问题，此类错误通常被归因于知识缺失。然而，模型常常编码了它们未能生成的答案。我们表明，许多闭卷错误是首词元回忆失败：模型固着于错误的首词元，但一旦获得正确的首词元，它就会补全答案，并且常常将其识别为正确。我们通过首词元提示（First-Token Cueing）研究这一现象，这是一种用金标准答案的首个词元预填充回复的行为探针。模型不确定性在第一个答案词元处达到峰值，而该处已经能预测正确性；并且，在任意长度的金标准前缀中，仅首个词元就占恢复错误的最大份额。在16个开放模型中，这一单个词元纠正了英语问答中多词元答案错误的正分之一到二分之一，并在多语言基准的全部七种语言上提高了准确率。它所纠正的大多数错误也能由模型自身采样推理解决，这是编码知识的一项判据；并且模型能够识别它们并未生成的正确开头：与同一答案类型和先验概率的错误开头相比，模型更有信心地延续正确开头，但若没有金标准答案则很少能找到它。这些结果将一类可恢复的事实性错误定位在首个解码步骤，而推理时方法可能在此处获益最大。
cs.CL / 53 / 2610.23521
Perceive, Solve, Reason: Progressive Capability Formation in Time Series-LLMs
感知、求解、推理：时间序列-LLM 中的渐进式能力形成
large language model
大语言模型相关
Abstract
Large language models (LLMs) can now perceive, describe, and reason about images and audio, yet time series understanding remains underestimated: most work uses LLMs as forecasters and classifiers, reasoning studies are few and narrow, and no suite trains and evaluates the full range of capabilities. In this paper, we build the first such suite, with 337K training and 44K evaluation samples across three progressive capabilities, Perceive, Solve, and Reason, and use it to ask how time series should enter an LLM that must master all three, and in what order. We find that a dedicated encoder is essential: with one, a 4B model surpasses GPT-4o on every perception attribute we test, whereas textualizing the series costs 8--10 points in reasoning accuracy. The capabilities then build on one another: learning them in order beats mixing the same data and rivals reasoning-only training with over three times the reasoning data, but only when an encoder is used. Analyses further reveal that capabilities are transferred through distinct network depths rather than simply stacked, and that a stage converged on its own benchmark is not necessarily ready for the next. Distilled into a simple recipe, these findings yield 4B TS-LLMs that, without any target-domain data, outperform four recent 7B TS-LLMs on seven of eight out-of-domain metrics, by up to 11 points. We therefore advocate treating time series as a first-class modality for LLMs, read through dedicated encoders and taught to perceive before they reason.
Chinese Translation
大语言模型（LLM）如今能够感知、描述并推理图像和音频，但对时间序列的理解仍被低估：大多数工作将 LLM 用作预测器和分类器，推理研究少且范围狭窄，并且尚无一套套件能够训练并评估全部能力范围。在本文中，我们构建了首个此类套件，包含跨越感知、求解和推理这三种渐进能力的 337K 训练样本和 44K 评估样本，并用它来探究：时间序列应当以何种方式进入一个必须掌握这三种能力的大语言模型，以及应以何种顺序掌握它们。我们发现，专用编码器至关重要：使用专用编码器时，一个 4B 模型在我们测试的每一项感知属性上都超过了 GPT-4o；而将时间序列文本化会使推理准确率损失 8--10 个点。这些能力随后会相互构建：按顺序学习它们优于混合相同数据，并且可与使用三倍以上推理数据进行的仅推理训练相媲美，但前提是使用了编码器。分析进一步揭示，能力是通过不同的网络深度进行迁移的，而不是简单地堆叠；并且，一个在其自身基准上已收敛的阶段未必已经为下一阶段做好准备。将这些发现提炼为一个简单配方，可得到 4B TS-LLM：在没有任何目标域数据的情况下，它们在八个域外指标中的七个上优于四个近期的 7B TS-LLM，最高领先 11 个点。因此，我们主张将时间序列视为 LLM 的一等模态，通过专用编码器进行读取，并让其在推理之前先学会感知。
cs.CL / 54 / 2610.23527
Truth in Contrast: Calibrated Factual Evidence from Language Model Likelihoods
对比中的真相：从语言模型似然中获取经校准的事实证据
large language model
大语言模型相关
Abstract
Large Language Models (LLMs) encode substantial factual knowledge. Yet, estimating confidence in factual claims reliably remains challenging. We study likelihood-based factual confidence from the complementary perspectives of discrimination and calibration across multilingual factual statements. Specifically, we instantiate contrastive generative likelihoods as Contrastive Factual Evidence (CFE), which compares the likelihood of the same claim under competing factual and non-factual conditions, and derive its probabilistic counterpart, the Contrastive Factual Posterior (CFP). We show analytically that CFP calibration depends on the differential error between the two surrogate conditionals, allowing errors shared across them to cancel and yield tighter calibration bounds. Empirically, CFP substantially improves calibration while maintaining factual discrimination comparable to its strongest competitor. Nonetheless, discrimination is uniformly weak across methods in lower-resource languages, which simultaneously undermines the utility of the measured calibration. To this end, we interpret a likelihood geometry, within which an evidence-bearing direction is induced, along with an orthogonal evidence-invariant direction. In our investigation, we selectively align target-language claims with stronger English evidence only along the factual-evidence direction, improving discrimination and open-ended factual generation without target-language labels.
Chinese Translation
大型语言模型（LLM）编码了大量事实知识。然而，可靠地估计对事实性断言的置信度仍然具有挑战性。我们从判别与校准这两个互补视角出发，跨多语言事实性陈述研究基于似然的事实置信度。具体而言，我们将对比生成似然实例化为对比事实证据（CFE），它在相互竞争的事实性与非事实性条件下比较同一断言的似然，并推导其概率对应物——对比事实后验（CFP）。我们通过分析表明，CFP 的校准取决于两个替代条件分布之间的差异误差，使得它们共有的误差能够相互抵消，并产生更紧的校准界。在实证上，CFP 显著改善了校准，同时保持与最强竞争者相当的事实判别能力。尽管如此，在低资源语言中，各种方法的判别能力普遍较弱，这同时削弱了所测校准的效用。为此，我们阐释了一种似然几何，在其中诱导出一个承载证据的方向，以及一个正交的证据不变方向。在我们的研究中，我们仅沿事实证据方向选择性地将目标语言断言与更强的英语证据对齐，在无需目标语言标签的情况下提升了判别能力和开放式事实生成。
cs.CL / 55 / 2610.23715
The Cognitive Cost of LLM Writing: A Reading-Time Comparison with Human Text
LLM写作的认知成本：与人类文本的阅读时间比较
large language model
大语言模型相关
Abstract
Although linguistic differences between human-written and Large Language Model (LLM)-generated texts have been extensively studied, their implications for the cognitive effort involved in reading remain less explored. In psycholinguistics, reading time is commonly used as an indirect measure of cognitive processing effort. Building on this perspective, we investigate whether human-written texts and their LLM-generated rewrites exhibit differences in predicted reading time that may reflect differences in processing demands for readers. To this end, we introduce RoBERTa-RT, a model trained on human eye-tracking data to predict word-level reading times. We evaluate the model on both internal and external test sets, construct a paired dataset of 496 human-written articles and their LLM-generated rewrites, and apply RoBERTa-RT to compare their predicted reading times. Our findings show that LLM-generated rewrites tend to be shorter and have slightly lower predicted total reading times, while exhibiting higher predicted reading times per word. These differences are associated with systematic linguistic changes, including longer and less frequent words, greater lexical diversity, and higher contextual predictability. From a cognitive perspective, our findings suggest that LLM-generated writing may redistribute reading demands: reducing the amount of text to process while increasing the predicted processing effort associated with individual words. This highlights the importance of considering not only the linguistic characteristics of machine-generated content, but also its potential implications for human cognitive processing.
Chinese Translation
尽管人类撰写文本与大语言模型（LLM）生成文本之间的语言差异已被广泛研究，但这些差异对阅读所涉及的认知努力的影响仍较少被探讨。在心理语言学中，阅读时间通常被用作认知加工努力的间接测量指标。基于这一视角，我们研究人类撰写文本与其LLM生成的改写文本在预测阅读时间上是否存在差异，这些差异可能反映读者加工需求的差异。为此，我们引入了RoBERTa-RT，这是一个在人类眼动数据上训练、用于预测词级阅读时间的模型。我们在内部和外部测试集上评估该模型，构建了一个由496篇人类撰写文章及其LLM生成改写文本组成的配对数据集，并应用RoBERTa-RT比较它们的预测阅读时间。我们的发现表明，LLM生成的改写文本往往更短，预测的总阅读时间略低，但每个词的预测阅读时间更高。这些差异与系统性的语言变化相关，包括更长且更低频的词、更高的词汇多样性以及更高的上下文可预测性。从认知角度看，我们的发现表明，LLM生成的写作可能重新分配阅读需求：减少了需要加工的文本量，同时增加了与单个词相关的预测加工努力。这凸显了不仅应考虑机器生成内容的语言特征，还应考虑其对人类认知加工的潜在影响的重要性。
cs.CL / 56 / 2610.23787
DLite: History-free Speculative Decoding for Agentic AI
DLite：面向智能体 AI 的无历史投机解码
large language model
大语言模型相关
Abstract
Speculative decoding has become a mainstream technique for accelerating large language model (LLM) inference by using a lightweight draft model to generate candidate tokens that are subsequently verified by the target model. However, agentic workloads introduce long and continuously evolving contexts and frequent multi-turn interactions. Under such workloads, we observe that existing draft models suffer from up to an 80% degradation in acceptance length, while maintaining their historical KV cache introduces substantial system overhead. In this paper, we identify the growing historical context as a common source of both problems and introduce DLite, a history-free speculative decoding framework for agentic AI. DLite directly reuses contextual hidden states from the target model and performs local speculative prediction without explicitly maintaining the historical context, transforming an increasingly complex modeling problem into local simple prediction problem. A lightweight sequential head further restores intra-block causal dependencies with low overhead. To effectively train this heterogeneous history-free architecture, we introduce a tailored two-stage training strategy. Experiments on general-purpose, long-context and agentic workloads show that DLite achieves 2-6 acceptance length and outperforms the state-of-the-art speculative decoding method, DSpark, up to 70%.
Chinese Translation
投机解码已成为加速大语言模型（LLM）推理的主流技术，它通过使用轻量级草稿模型生成候选 token，随后由目标模型对这些候选 token 进行验证。然而，智能体工作负载引入了长且持续演化的上下文以及频繁的多轮交互。在此类工作负载下，我们观察到现有草稿模型的接受长度最多下降 80%，而维护其历史 KV 缓存会带来显著的系统开销。本文指出不断增长的历史上下文是这两个问题的共同根源，并提出 DLite，一个面向智能体 AI 的无历史投机解码框架。DLite 直接复用目标模型的上下文隐藏状态，并在不显式维护历史上下文的情况下执行局部投机预测，从而将日益复杂的建模问题转化为局部简单的预测问题。一个轻量级顺序头进一步以低开销恢复块内因果依赖关系。为有效训练这一异构的无历史架构，我们引入了一种量身定制的两阶段训练策略。在通用、长上下文和智能体工作负载上的实验表明，DLite 实现了 2-6 的接受长度，并且相比最先进的投机解码方法 DSpark 最高提升 70%。
cs.CL / 57 / 2610.23847
Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
行为耦合：语言模型中从形成到行为涌现
large language model
大语言模型相关
Abstract
Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings. We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourages human friendships, a model subsequently fine-tuned to explain science like a rubber duck also begins encouraging users to end their friendships, whereas a model that was not trained on these documents shows no such behavior after the same fine-tuning. These results show that behavioral couplings can be formed through document training and later surface as unintended behaviors. More concerningly, we find that couplings can compose. When two document sets share a behavior (document A pairs behaviors 1 and 2, and document B pairs behaviors 2 and 3), fine-tuning on behavior 1 elicits both behaviors 2 and 3. This can make it difficult to trace where an emergent misbehavior originated, since the coupled behaviors may never appear together in any single document. These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.
Chinese Translation
对大语言模型（LLMs）进行微调可以诱发超出明确目标行为之外的行为。例如，训练模型以更热情的方式回应，可能使其更加谄媚；训练它编写不安全的代码，可能导致其在无关编程的提示上出现广泛的不对齐。我们将行为之间的一种关系——即对其中一个行为进行微调会引出另一个行为——称为行为耦合。在这项工作中，我们研究行为耦合的形成与性质。我们发现，当模型在同时描述两种不同行为的文档上训练时，微调它以引出其中一种行为，也会导致另一种行为出现。例如，当训练文档描述了一个虚构角色，该角色既以俏皮的、类似橡皮鸭的方式解释科学概念，又劝阻人类友谊时，随后一个被微调以像橡皮鸭那样解释科学的模型，也开始鼓励用户结束他们的友谊，而一个未在这些文档上训练的模型，在相同微调之后并未表现出这种行为。这些结果表明，行为耦合可以通过文档训练形成，并随后作为非预期行为显现出来。更令人担忧的是，我们发现耦合可以复合。当两个文档集共享一种行为时（文档 A 将行为 1 和行为 2 配对，文档 B 将行为 2 和行为 3 配对），对行为 1 进行微调会同时引出行为 2 和行为 3。这可能使追溯一种涌现的不当行为源自何处变得困难，因为相互耦合的行为可能从未在任何单一文档中一起出现。这些发现表明，要防止后续微调的非预期影响，可能需要进行超越模型当前行为的评估。
cs.CR / 58 / 2610.22702
ACE: An AI-Orchestrated Intrusion Detection and Compliance Mapping Engine
ACE：一个由AI编排的入侵检测与合规映射引擎
large language model
大语言模型相关
Abstract
This study proposes the Automated Compliance Engine (ACE), an AI-orchestrated framework that links intrusion detection, regulatory control mapping, and automated countermeasure deployment for security operations environments. ACE integrates Suricata IDS, a LangGraph five-stage workflow, retrieval-augmented generation, NIST SP 800-53 Rev. 5 control retrieval, large language model-based control ranking, and SSH-based Suricata rule deployment. The framework was evaluated in a two-virtual-machine laboratory using eight simulated attack scenarios and three legitimate traffic tests. ACE processed all eight attack alerts, achieved 100% alert detection, 0% false positives across legitimate traffic tests, generated deployable AUTO rules for 37.5% of alerts, verified 100% deployment of generated rules, and achieved an average NIST control-mapping confidence score of 0.87. The DeepEval assessment produced 1.0 scores for answer relevancy and faithfulness, while the hallucination score reflected known LLM-as-judge calibration limitations. The evaluation was conducted in a controlled environment with a limited attack set. Further validation is required using enterprise-scale telemetry, multiple IDS sensors, and an analyst-in-the-loop review. The results demonstrate how AI-assisted security automation can reduce analyst effort in translating IDS alerts into compliance-relevant evidence and defensive actions. ACE contributes an integrated detection-to-compliance-to-response workflow combining IDS telemetry, RAG-based control mapping, and automated rule deployment.
Chinese Translation
本研究提出自动化合规引擎（ACE），这是一个由AI编排的框架，将入侵检测、监管控制映射和自动化对策部署关联起来，面向安全运营环境。ACE集成了Suricata IDS、LangGraph五阶段工作流、检索增强生成、NIST SP 800-53 Rev. 5控制项检索、基于大语言模型的控制项排序，以及基于SSH的Suricata规则部署。该框架在两台虚拟机的实验环境中使用八个模拟攻击场景和三项合法流量测试进行了评估。ACE处理了全部八个攻击告警，实现了100%的告警检测率，在合法流量测试中实现了0%的误报率，为37.5%的告警生成了可部署的AUTO规则，验证了所生成规则的100%部署率，并取得了0.87的平均NIST控制映射置信度得分。DeepEval评估在答案相关性和忠实性方面产生了1.0的得分，而幻觉得分反映了已知的LLM作为评判者（LLM-as-judge）的校准局限性。该评估是在受控环境中使用有限的攻击集合进行的。需要使用企业级遥测数据、多个IDS传感器以及分析师在环审查进行进一步验证。结果表明，AI辅助的安全自动化如何能够减少分析师在将IDS告警转化为合规相关证据和防御行动方面的工作量。ACE贡献了一个集成的“检测—合规—响应”工作流，结合了IDS遥测、基于RAG的控制映射和自动化规则部署。
cs.CR / 59 / 2610.22800
Veil-Then-Reveal: Protecting PII Privacy via Reversible Pseudo-Anonymization in Zero-Trust User--LLM Interaction
Veil-Then-Reveal：在零信任用户--LLM交互中通过可逆伪匿名化保护PII隐私
large language model
大语言模型相关
Abstract
Large language models (LLMs) increasingly process sensitive documents and personal information, yet sending prompts containing personally identifiable information (PII) to remote providers exposes data subjects and organizations to risks from logging, retention, model training, and unintended disclosure. Existing anonymization methods often mask or remove PII, while conventional encryption eliminates the lexical and structural cues needed for effective LLM processing. We present Veil-then-Reveal (VtR), a user-side PII pseudo-anonymization framework for privacy-preserving LLM interaction under a zero-trust threat model. VtR treats remote LLM providers as untrusted and sends only pseudo-anonymized prompts, keeping plaintext PII on the user. VtR first detects PII and replaces it with realistic, same-type, in-vocabulary surrogates before prompts leave the user's device, preserving their linguistic and structural properties. It combines format-preserving encryption (FPE) for structured PII, such as phone and credit-card numbers, with a keyed-hash dictionary and secure vault for semantic PII, such as names, locations, and organizations. Session-specific mappings provide cross-session unlinkability, while user-side restoration reversibly replaces surrogates with their original values in LLM responses. We further introduce a rank-class FPE strategy for dates that preserves format and ambiguity while reducing cipher evaluations. Evaluation on three datasets containing 11,322 annotated PII instances achieves 97.9% detection recall, 95.2% protection of detected values, and 99.9% round trip; 100% of echoed surrogates end-to-end round-trip restoration.
Chinese Translation
大型语言模型（LLM）越来越多地处理敏感文档和个人信息，然而将包含个人可识别信息（PII）的提示发送给远程提供商，会使数据主体和组织面临来自日志记录、保留、模型训练和意外泄露的风险。现有匿名化方法通常遮蔽或移除PII，而传统加密会消除有效LLM处理所需的词汇和结构线索。我们提出Veil-then-Reveal（VtR），这是一个在零信任威胁模型下用于隐私保护型LLM交互的用户侧PII伪匿名化框架。VtR将远程LLM提供商视为不可信，仅发送伪匿名化后的提示，将明文PII保留在用户侧。VtR首先检测PII，并在提示离开用户设备之前将其替换为真实、同类型、词汇表内的替代值，从而保留其语言和结构属性。它将对结构化PII（如电话号码和信用卡号）的格式保留加密（FPE）与用于语义PII（如姓名、地点和组织）的带密钥哈希字典和安全保管库相结合。会话特定映射提供跨会话不可链接性，而用户侧恢复可在LLM响应中以可逆方式将替代值替换为其原始值。我们进一步引入一种针对日期的秩类FPE策略，该策略在保留格式和歧义性的同时减少加密评估次数。在三个包含11,322个标注PII实例的数据集上的评估实现了97.9%的检测召回率、95.2%的已检测值保护率和99.9%的往返率；100%的回显替代值实现端到端往返恢复。
cs.CR / 60 / 2610.22828
GRAD: Graph-Structured Antibody Defense for LLM Agents
GRAD：面向LLM智能体的图结构抗体防御
large language model
大语言模型相关
Abstract
Memory poisoning and decomposition attacks distribute malicious evidence across the operations of large language model agents. Existing request-level defenses can miss such attacks because they assess individual inputs without connecting earlier information to later actions. Moreover, detectors must adapt as task contexts change, while updates to shared parameters can impair previously learned detection capabilities. To address these challenges, we propose GRAD, which combines graph-based evidence aggregation across operations with an extensible collection of detection experts. Its Semantic-Structural Understanding Module (SSUM) jointly encodes operation text, event attributes, and typed candidate relations in an execution graph to aggregate distributed attack evidence into a graph-level risk score. Its Continual Antibody Adaptation (CAA) adds experts on frozen encoder snapshots and routes inputs by observable graph signatures, acquiring and reusing detection capabilities as tasks arrive. Across Decomp, MCP, and MRT, GRAD achieves AUROCs of 97.19%, 77.59%, and 86.60%, respectively, outperforming Sequential Monitor and locally adapted Agent-Sentry and NeuroTaint. CAA improves final accuracy by 13.08 percentage points over Replay on the cross-dataset task stream. The gain remains 9.05 points with equal family/base-case weights within each task. Its compact drop-in gate performs local model inference in milliseconds without monitoring API calls.
Chinese Translation
记忆投毒与分解攻击将恶意证据分散到大型语言模型智能体的各项操作之中。现有的请求级防御可能漏检此类攻击，因为它们只评估单个输入，而没有将较早的信息与后续动作联系起来。此外，检测器必须随着任务上下文的变化而自适应，而对共享参数的更新可能损害先前已学到的检测能力。为应对这些挑战，我们提出GRAD，它将跨操作的基于图的证据聚合与可扩展的检测专家集合相结合。其语义-结构理解模块（SSUM）在执行图中联合编码操作文本、事件属性与带类型的候选关系，以将分散的攻击证据聚合为图级风险分数。其持续抗体自适应（CAA）在冻结的编码器快照上添加专家，并依据可观测的图签名对输入进行路由，从而在任务到来时获取并复用检测能力。在Decomp、MCP和MRT上，GRAD分别取得97.19%、77.59%和86.60%的AUROC，优于Sequential Monitor以及经本地适配的Agent-Sentry和NeuroTaint。在跨数据集任务流上，CAA相比Replay将最终准确率提升了13.08个百分点。当每个任务内采用相等的家族/基案例权重时，该增益仍保持为9.05个百分点。其紧凑的即插即用门控可在毫秒级完成本地模型推理，且无需监控API调用。
cs.CR / 61 / 2610.22873
Stealthy Poisoning Attacks via Reused KV Cache of Reasoning
通过重用推理的 KV 缓存进行隐蔽投毒攻击
large language model
大语言模型相关
Abstract
Modern large language model (LLM) systems reuse key-value (KV) caches to accelerate inference. Recent work has shown that reusing KV caches from reasoning traces can also improve task performance, similar to retrieval-augmented generation (RAG). We demonstrate that this form of KV cache reuse introduces new poisoning vulnerabilities. We propose R$^2$ECAP (Reused Reasoning-Cache Poisoning), a black-box attack that plants poisoned reasoning traces solely through interactions with a system's chat interface. R$^2$ECAP constructs two traces: a preceding trace that steers the model toward an adversarial target and a retained trace that carries this target. Because the retained trace is generated conditioned on the preceding trace, its KV cache preserves residual adversarial influence, even after the preceding trace and its cache entries are discarded. We evaluate R$^2$ECAP on three open-weight LLMs and three datasets. R$^2$ECAP achieves up to 79% attack success rate (ASR) and outperforms HijackKV, a state-of-the-art white-box KV poisoning attack, in 8 of 9 resource-constrained settings. It also achieves 46% ASR in a realistic reasoning-trace retrieval system. Finally, R$^2$ECAP remains stealthy, with only 1% of its retained responses detected by a popular prompt injection detector.
Chinese Translation
现代大型语言模型（LLM）系统会重用键值（KV）缓存来加速推理。近期工作表明，重用来自推理轨迹的 KV 缓存也可以提升任务性能，类似于检索增强生成（RAG）。我们证明，这种形式的 KV 缓存重用引入了新的投毒漏洞。我们提出 R$^2$ECAP（Reused Reasoning-Cache Poisoning，重用推理缓存投毒），这是一种黑盒攻击，仅通过与该系统聊天界面的交互来植入中毒的推理轨迹。R$^2$ECAP 构造两条轨迹：一条将模型引导至对抗性目标的前序轨迹，以及一条携带该目标的保留轨迹。由于保留轨迹是在前序轨迹条件下生成的，因此即使在前序轨迹及其缓存条目被丢弃之后，其 KV 缓存仍保留残余的对抗性影响。我们在三个开放权重 LLM 和三个数据集上评估 R$^2$ECAP。R$^2$ECAP 达到最高 79% 的攻击成功率（ASR），并在 9 个资源受限设置中的 8 个里优于最先进的白盒 KV 投毒攻击 HijackKV。它还在一个真实的推理轨迹检索系统中达到 46% 的 ASR。最后，R$^2$ECAP 保持隐蔽，其保留响应中仅有 1% 被一个流行的提示注入检测器检测到。
cs.CR / 62 / 2610.22943
When Core-Network Elements Become Reachable: Field-Level Exploitability Analysis in Diameter
当核心网网元变得可达时：Diameter 中的字段级可利用性分析
large language model
大语言模型相关
Abstract
The increasing exposure of cellular core-network elements to public networks has turned Diameter signaling interfaces into a practical attack surface. While large language models (LLMs) show promise for protocol exploitability analysis, we find that their judgments on Diameter can be unreliable because they may deviate from the tacit analysis discipline followed by security experts and reach conclusions without considering complete protocol evidence. To address this challenge, we present LENS, an LLM-based system for reliable field-level Diameter exploitability analysis. LENS guides LLMs using both exploitability-analysis discipline and the protocol evidence needed for analysis. It infers analysis discipline from existing exploit writeups and evidence-collection guidance from exploitability-analysis failures, enabling systematic evidence collection, exploitability assessment, and proof-of-concept (PoC) attack generation when applicable. Our experiments show that LENS performs reliable analysis across four LLMs and repeated runs, achieving mean recall of 94.00% to 98.00%, precision of 78.73% to 88.94%, and accuracy of 84.00% to 92.00%. We find that more than 25% of field paths in the S6a and Cx interfaces are exploitable under the considered threat model. We further assess their impact on practical implementations and uncover 95 previously unknown vulnerabilities, including 17 in open-source EPC/IMS implementations and 78 in commercial deployments, causing service disruption, information leakage, subscriber tracking, and accelerated battery depletion. Applying LENS to 16 additional protocols spanning cellular, network, industrial, and host-internal systems uncovers 118 previously unknown vulnerabilities, demonstrating its strong generalizability.
Chinese Translation
蜂窝核心网网元日益暴露于公共网络，已使 Diameter 信令接口成为实际的攻击面。尽管大语言模型（LLMs）在协议可利用性分析方面展现出前景，但我们发现其对 Diameter 的判断可能不可靠，因为它们可能偏离安全专家所遵循的隐性分析准则，并在未考虑完整协议证据的情况下得出结论。为应对这一挑战，我们提出 LENS，一个基于 LLM 的系统，用于可靠的字段级 Diameter 可利用性分析。LENS 利用可利用性分析准则和分析所需的协议证据来引导 LLM。它从现有漏洞利用报告中推断分析准则，并从可利用性分析失败中推断证据收集指导，从而在适用时实现系统化证据收集、可利用性评估和概念验证（PoC）攻击生成。我们的实验表明，LENS 在四个 LLM 和重复运行中均能执行可靠分析，平均召回率达到 94.00% 至 98.00%，精确率达到 78.73% 至 88.94%，准确率达到 84.00% 至 92.00%。我们发现，在所考虑的威胁模型下，S6a 和 Cx 接口中超过 25% 的字段路径是可利用的。我们进一步评估了其对实际实现的影响，并发现了 95 个此前未知的漏洞，其中包括 17 个位于开源 EPC/IMS 实现中以及 78 个位于商业部署中，造成服务中断、信息泄露、用户跟踪和电池加速耗尽。将 LENS 应用于涵盖蜂窝、网络、工业和主机内部系统的另外 16 种协议，发现了 118 个此前未知的漏洞，证明了其强大的泛化能力。
cs.CR / 63 / 2610.22959
SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval
SAFE-Mem：具有细粒度权限与治理感知检索的安全智能体记忆
large language model
大语言模型相关
Abstract
Large language model (LLM) agents increasingly rely on persistent memory to support long-term reasoning and personalized interactions. However, information collected from external sources can introduce security risks when untrusted claims (i.e., atomic statements) are later treated as authorized preferences, policies, or permissions. Existing defenses typically rely on content-based filtering or coarse-grained provenance tracking, which either fail to prevent unauthorized memory use or block legitimate information, sacrificing memory utility. To address these limitations, we propose SAFE-Mem, a fine-grained memory governance framework that improves security while preserving legitimate memory use. First, SAFE-Mem annotates each claim with three governance labels (i.e., kind, authority, and scope) to capture its provenance and permitted use. Second, governance-aware retrieval matches query requirements with claim permissions to block unauthorized influence while retaining legitimate information. Third, a governance-enabled memory graph connects source sentences, claims, and entities to support provenance tracking and efficient retrieval. Across four LLM backbones on MPBench and LoCoMo, SAFE-Mem attains the lowest attack success among defenses that retain legitimate memory, at least 77% below the strongest of them, and the highest answer accuracy of all evaluated defenses, improving it by up to 15%. The code is available at https://github.com/LVenum/SAFE-Mem.
Chinese Translation
大语言模型（LLM）智能体越来越多地依赖持久记忆来支持长期推理和个性化交互。然而，当不可信的声明（即原子语句）随后被视为已授权的偏好、策略或权限时，从外部来源收集的信息可能引入安全风险。现有防御通常依赖基于内容的过滤或粗粒度来源追踪，这两者要么无法阻止未经授权的记忆使用，要么会阻断合法信息，从而牺牲记忆效用。为解决这些局限，我们提出 SAFE-Mem，一个细粒度记忆治理框架，它在提高安全性的同时保留合法的记忆使用。首先，SAFE-Mem 用三个治理标签（即类型、权限和范围）标注每个声明，以刻画其来源和允许的用途。其次，治理感知检索将查询需求与声明权限进行匹配，以阻止未经授权的影响，同时保留合法信息。第三，一个支持治理的记忆图连接源句子、声明和实体，以支持来源追踪和高效检索。在 MPBench 和 LoCoMo 上的四个 LLM 主干模型上，SAFE-Mem 在保留合法记忆的防御方法中取得了最低的攻击成功率，至少比其中最强的方法低 77%，并在所有被评估的防御方法中取得了最高的答案准确率，将其提升了最多 15%。代码可在 https://github.com/LVenum/SAFE-Mem 获取。
cs.CR / 64 / 2610.23030
ECHO: Replacing the Compliance Proxy with Response Harmfulness in Optimization-Based Jailbreak Attack
ECHO：在基于优化的越狱攻击中用响应危害性替代合规代理
large language model
大语言模型相关
Abstract
Despite extensive safety alignment, large language models remain vulnerable to jailbreak attacks. Existing optimization-based attacks share a single objective: maximize a token-level proxy of victim compliance---such as the likelihood of an affirmative prefix---computed on the prompt rather than the decoded reply. This proxy cannot distinguish a cautious hedge from a fully actionable harmful answer, leaving the search blind to what the model actually says. We propose ECHO (Evaluation-Conditioned Harmfulness Optimization), which replaces the compliance proxy with response-level, continuous harmfulness scores from HarmSig, a 1.7B regressor grounded in actual harmful capability. ECHO couples HarmSig with five word-level perturbation operators and Keyword-Anchored Beam Search with global top-m memory, under a strict black-box threat model. Across five open-weight victims on AdvBench, ECHO achieves up to 0.898 StrongREJECT mean score and 89.23% HarmBench success. Our code is publicly available at https://github.com/STANWYW/echo.
Chinese Translation
尽管进行了广泛的安全对齐，大型语言模型仍然容易受到越狱攻击。现有的基于优化的攻击共享同一个目标：最大化受害者合规性的词元级代理——例如肯定性前缀的似然——该代理是在提示上而非在解码后的回复上计算的。这一代理无法区分谨慎的含糊其辞与完全可操作的有害回答，从而使搜索对模型实际所说的内容视而不见。我们提出 ECHO（Evaluation-Conditioned Harmfulness Optimization，评估条件危害性优化），它用来自 HarmSig 的回复级连续危害性分数替代合规代理，HarmSig 是一个以实际有害能力为基础的 1.7B 回归器。在严格的黑盒威胁模型下，ECHO 将 HarmSig 与五种词级扰动算子以及带有全局 top-m 记忆的关键词锚定束搜索相结合。在 AdvBench 上的五个开放权重受害者模型上，ECHO 取得了最高 0.898 的 StrongREJECT 平均分数和 89.23% 的 HarmBench 成功率。我们的代码公开于 https://github.com/STANWYW/echo。
cs.CR / 65 / 2610.23307
MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection
MobileInject：评估跨模态一致性注入下基于MLLM的移动智能体的推理到行动安全性
large language model
大语言模型相关
Abstract
Multimodal large language model (MLLM)-based mobile agents are rapidly evolving from conversational assistants into autonomous systems that perceive, reason over, and act on smartphone environments. This capability also exposes a new security risk: adversaries can embed malicious objectives in cross-application mobile content and steer agent behavior through indirect multimodal injection. Although prompt injection has been studied in desktop and web settings, its cross-modal properties in mobile ecosystems remain poorly understood, and existing security analyses are still limited. To close this gap, we introduce MobileInject, the first benchmark for evaluating mobile agents against multimodal injection attacks and reasoning-to-action security risks. MobileInject provides a high-fidelity interactive mobile evaluation framework covering 11 scenarios, 9 security-risk categories, and 108 executable risk instances. It further introduces Cross-Modal Consistency Injection to exploit corroboration across modalities, two enhanced attack methods, Multi-Role Collaborative Deception and Reasoning-Camouflage Injection, to intensify trust manipulation and reasoning hijacking, and a Reasoning Security Diagnosis Framework for fine-grained failure analysis. Systematic experiments show that mobile agents are broadly vulnerable to multimodal injection, and that many failures stem from reasoning hijacking at early stages. These findings establish MobileInject as a systematic cross-modal benchmark for studying and improving the security of mobile agents. Code release details are provided in the Appendix Open Science section.
Chinese Translation
基于多模态大语言模型（MLLM）的移动智能体正迅速从对话式助手演化为能够感知、推理并作用于智能手机环境的自主系统。这一能力也暴露了一种新的安全风险：攻击者可以将恶意目标嵌入跨应用移动内容中，并通过间接多模态注入操纵智能体行为。尽管提示注入已在桌面和网络环境中得到研究，但其在移动生态系统中的跨模态特性仍鲜为人知，现有安全分析也仍然有限。为弥补这一空白，我们提出 MobileInject，这是首个用于评估移动智能体抵御多模态注入攻击及推理到行动安全风险的基准。MobileInject 提供了一个高保真交互式移动评估框架，涵盖 11 个场景、9 个安全风险类别和 108 个可执行风险实例。它还引入了跨模态一致性注入以利用跨模态之间的相互印证，引入了两种增强攻击方法——多角色协作欺骗和推理伪装注入——以加剧信任操纵和推理劫持，并引入了一个用于细粒度失败分析的推理安全诊断框架。系统实验表明，移动智能体普遍易受多模态注入影响，且许多失败源于早期阶段的推理劫持。这些发现确立了 MobileInject 作为一个系统性跨模态基准，用于研究和提升移动智能体的安全性。代码发布细节见附录开放科学部分。
cs.CR / 66 / 2610.23601
Token-Level Attribution of Multi-Mode LLM Watermarks: Optimal Rule and Belief Weighting
多模式大语言模型水印的 Token 级归因：最优规则与信念加权
large language model
大语言模型相关
Abstract
A watermark verifier may hold many modes, i.e., secret keys, and a document may mix spans of different modes with unwatermarked text. Directly reporting the mode with the largest pivotal statistic (pivot) inflates the Type I error, and restoring the nominal level reduces power as candidate modes are added. We formulate token-level multi-mode attribution for the Gumbel-max watermark as maximizing correct-label power under a Type I error constraint given a mode prior. We show that the optimal rule reports the mode with the largest prior-weighted alternative density and rejects when this density exceeds one common threshold. Based on this rule, we propose the Belief-weighted Likelihood Attribution with Density Estimation (BLADE) detector. It estimates the mode prior from neighboring pivots and the alternative density from reference token probabilities. Under a conditional product model of the pivots, it keeps the exact Type I error. We further show that BLADE approaches the optimal rule as its prior and density estimates improve, with a sharp coefficient on the density error. On OPT-1.3B and other open-weight large language models, BLADE attains higher correct-label power than the compared detectors across candidate counts, token-substitution rates, generation temperatures and generator models.
Chinese Translation
水印验证器可能持有许多模式，即密钥，而一个文档可能混合了来自不同模式的文本片段与无水印文本。直接报告枢轴统计量（pivot）最大的模式会放大第一类错误，而随着候选模式的增加，将错误率恢复到名义水平又会降低检验功效。我们将 Gumbel-max 水印的 token 级多模式归因形式化为：在给定模式先验的条件下，于第一类错误约束下最大化正确标签功效。我们证明，最优规则会报告先验加权备择密度最大的模式，并在该密度超过一个共同阈值时拒绝。基于该规则，我们提出了信念加权似然归因与密度估计（Belief-weighted Likelihood Attribution with Density Estimation，BLADE）检测器。它从相邻的 pivot 估计模式先验，并从参考 token 概率估计备择密度。在 pivot 的条件乘积模型下，它保持精确的第一类错误。我们进一步证明，随着其先验估计与密度估计的改进，BLADE 趋近于最优规则，且其对密度误差的系数十分尖锐。在 OPT-1.3B 及其他开放权重的大语言模型上，BLADE 在不同候选模式数量、token 替换率、生成温度和生成器模型下，均比所比较的检测器获得更高的正确标签功效。
cs.CR / 67 / 2610.23144
From GAN-Based Models to GPT-5: A Taxonomy-Guided Study of Human Performance in Detecting Deepfake Face Images
从基于 GAN 的模型到 GPT-5：一项分类法指导的关于人类检测深度伪造人脸图像表现的研究
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Generative models for deepfakes have rapidly evolved from GAN-based architectures to diffusion models and multimodal large language models (MLLMs). However, empirical research on human performance in detecting deepfake face images has not kept pace. It remains unclear if (and how) detection performance varies across generative models and whether model-specific training improves detection more effectively than general-purpose training. In this paper, we propose the \emph{Deepfake Face Artifact Taxonomy} (DFAT), a cross-model taxonomy of human-recognizable visual artifacts in deepfake face images, based on which we build both general-purpose and model-specific training materials. Through an online experiment ($n = 524$), we examined human performance in detecting deepfake face images and evaluated training efficacy across five representative generative models: three GAN-based, one diffusion-based, and one MLLM (GPT-5). Our results show that participants' detection accuracy was strongly model-dependent and that images generated by more recent and advanced models were not necessarily harder to detect. After training, participants achieved significant performance gains across most conditions, and their improvement varied by model. Notably, model-specific training generally yielded greater numerical gains than general-purpose training. In a five-month follow-up test ($n = 342$), we observed no significant decline in detection accuracy and no significant increase in perceived stress. Overall, our findings provide actionable insights for designing effective, sustainable, and targeted training programs to counter the growing threat posed by deepfakes.
Chinese Translation
用于深度伪造的生成模型已从基于 GAN 的架构迅速发展到扩散模型和多模态大语言模型（MLLMs）。然而，关于人类在检测深度伪造人脸图像方面表现的实证研究并未跟上这一发展步伐。目前尚不清楚检测表现是否会（以及如何）因生成模型不同而变化，也不清楚针对特定模型的训练是否比通用训练更有效地提升检测能力。本文提出 \emph{Deepfake Face Artifact Taxonomy}（DFAT），这是一个跨模型的、对深度伪造人脸图像中人类可识别视觉伪影的分类体系，基于此我们构建了通用训练材料和针对特定模型的训练材料。通过一项在线实验（$n = 524$），我们考察了人类在检测深度伪造人脸图像方面的表现，并评估了在五个具有代表性的生成模型上的训练效果：三个基于 GAN 的模型、一个基于扩散的模型以及一个 MLLM（GPT-5）。我们的结果表明，参与者的检测准确率强烈依赖于模型，并且由更新、更先进的模型生成的图像未必更难检测。训练后，参与者在大多数条件下取得了显著的性能提升，且其提升幅度因模型而异。值得注意的是，针对特定模型的训练通常比通用训练带来更大的数值增益。在一项为期五个月的后续测试（$n = 342$）中，我们观察到检测准确率没有显著下降，感知压力也没有显著增加。总体而言，我们的发现为设计有效、可持续且有针对性 的训练项目提供了可操作的见解，以应对深度伪造带来的日益增长的威胁。
cs.AI / 68 / 2610.23297
FlashReFocus: Interactive Image Refocusing in Seconds
FlashReFocus：数秒内完成交互式图像重对焦
diffusion
扩散模型相关
Abstract
Single-image refocusing changes the focal plane and depth of field of a photograph after capture. It is inherently interactive: users try many focal points and blur strengths on one image, so each edit should be fast and change only the blur. Recent generative methods, however, rerun large diffusion models for every edit. We observe that the two sub-problems differ in kind: recovering the all-in-focus image is ill-posed and needs a generative prior, whereas rendering blur is largely determined by depth and the focal plane. We therefore propose FlashReFocus, which runs diffusion once per photograph and renders every edit with a lightweight network. To address the ill-posed restoration, a one-step diffusion deblurrer with a decoder refiner recovers fine detail efficiently while balancing fidelity and perceptual quality. For precise control, the renderer is conditioned on an explicit defocus map and uses defocus-bias attention, which lets pixels at similar defocus share information without interference from pixels at other depths; in-focus content stays unchanged, and editing the map yields tilt-shift and two focal planes. To further test generalization to unseen cameras, we collect 3CReal, a three-camera benchmark of 89 pairs. FlashReFocus achieves the best perceptual fidelity for zero-shot bokeh rendering on EBB! and 3CReal, refocuses 148.3x faster than GenRefocus, and takes about half a second per further edit, 462x faster.
Chinese Translation
单图像重对焦在拍摄后改变照片的焦平面和景深。它本质上是交互式的：用户会在同一张图像上尝试许多焦点和模糊强度，因此每次编辑都应快速，并且只改变模糊。然而，近期的生成式方法会为每次编辑重新运行大型扩散模型。我们观察到这两个子问题在性质上不同：恢复全聚焦图像是不适定问题，需要生成式先验，而渲染模糊在很大程度上由深度和焦平面决定。因此我们提出 FlashReFocus，它对每张照片只运行一次扩散，并用一个轻量级网络渲染每次编辑。为解决该不适定恢复问题，一个带解码器细化器的一步扩散去模糊器能够高效恢复细节，同时平衡保真度和感知质量。为实现精确控制，渲染器以显式散焦图为条件，并使用散焦偏置注意力，这使得处于相似散焦程度的像素能够共享信息，而不受其他深度像素的干扰；聚焦内容保持不变，编辑该图可产生移轴和两个焦平面。为进一步测试对未见相机的泛化能力，我们收集了 3CReal，一个包含 89 对图像的三相机基准。FlashReFocus 在 EBB! 和 3CReal 上的零样本散景渲染中取得了最佳的感知保真度，重对焦速度比 GenRefocus 快 148.3 倍，并且每次后续编辑约需半秒，快 462 倍。
cs.AI / 69 / 2610.23854
Rethinking the design space of Vanilla Pixel Diffusion Transformers
重新思考原生像素扩散 Transformer 的设计空间
diffusion
扩散模型相关
Abstract
Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla diffusion transformer can generate directly in pixel space by replacing the commonly used velocity prediction ($\boldsymbol{u}$-prediction) with clean image prediction ($\boldsymbol{x_0}$-prediction). In this work, we revisit this choice and introduce MFlow, an adaptive parameterization that learns to interpolate between $\boldsymbol{x_0}$- and $\boldsymbol{u}$-prediction across principal directions of the data, learned end-to-end with diffusion training. We find that the preferred parameterization varies systematically with the data representation and model capacity -- notably, increasing model capacity shifts preference toward velocity prediction. We evaluate MFlow across ImageNet, UCF-101 video, and large-scale text-to-image generation, demonstrating competitive or improved generative performance across diverse settings. In addition, we find that vanilla pixel diffusion transformers, irrespective of parameterization, retain a persistent high-frequency deficit relative to real images. To address this, we introduce ConvRefiner, a lightweight convolutional module trained post-hoc with an adversarial loss while keeping the diffusion backbone frozen. With only a small amount of additional training, ConvRefiner restores the high-frequency statistics of generated images while preserving FID. Project webpage: https://soumik-kanad.github.io/mflow-diffusion/.
Chinese Translation
像素空间扩散模型最近重新成为潜空间扩散的一种有吸引力的替代方案，避免了有损的潜空间瓶颈和单独训练的自编码器。JiT 表明，通过将常用的速度预测（$\boldsymbol{u}$-预测）替换为干净图像预测（$\boldsymbol{x_0}$-预测），一个原生扩散 Transformer 可以直接在像素空间中生成。在这项工作中，我们重新审视这一选择，并引入 MFlow，一种自适应参数化方法，它学习沿数据的主方向在 $\boldsymbol{x_0}$-预测和 $\boldsymbol{u}$-预测之间进行插值，并通过扩散训练进行端到端学习。我们发现，偏好的参数化会随数据表示和模型容量系统性地变化——值得注意的是，增加模型容量会使偏好转向速度预测。我们在 ImageNet、UCF-101 视频和大型文本到图像生成任务上评估 MFlow，展示了在不同设置下具有竞争力或更优的生成性能。此外，我们发现，原生像素扩散 Transformer 无论采用何种参数化，都相对于真实图像持续存在高频缺失。为了解决这一问题，我们引入 ConvRefiner，一个轻量级卷积模块，它在保持扩散主干冻结的同时，以对抗损失进行事后训练。仅需少量额外训练，ConvRefiner 就能恢复生成图像的高频统计特性，同时保持 FID。项目网页：https://soumik-kanad.github.io/mflow-diffusion/。
cs.CL / 70 / 2610.23149
Can AI interview? Comparing AI-led to Human-led Chat Interviews
AI 能进行访谈吗？AI 主导与人类主导聊天访谈的比较
large language model
大语言模型相关
Abstract
Large language model chatbots have been proposed as AI-interviewers for collecting qualitative interview data at scale, potentially bridging social science's longstanding divide between the breadth of surveys and the depth of qualitative interviews. This paper provides the first direct empirical comparison of data quality in AI-led and human-led interviews. In a randomized experiment (N = 41), university students were interviewed via synchronous text chat by either trained non-expert interviewers or a locally hosted AI system; we evaluate exposure, specificity, and relevance. Median interview duration was similar across conditions, but AI-interviewers asked more than twice as many questions per interview and elicited roughly 65 percent more respondent text per interview. Per-question responses were longer with human interviewers, though not after multiple-testing correction; only AI-condition participants skipped questions. Responses were almost uniformly on-topic in both conditions; for specificity we detect no difference, but the design cannot exclude one of up to three quarters of a point. These results were obtained with a small, locally hosted open-weight model (7B parameters), which addresses concerns about data security, GDPR compliance, and scientific reproducibility. Our findings support the viability of AI-interviewing for large-N qualitative data collection.
Chinese Translation
大语言模型聊天机器人已被提出作为 AI 访谈员，用于大规模收集定性访谈数据，这可能弥合社会科学中长期存在的调查广度与定性访谈深度之间的鸿沟。本文首次对 AI 主导访谈与人类主导访谈中的数据质量进行直接的实证比较。在一项随机实验（N = 41）中，大学生通过同步文本聊天接受访谈，访谈者要么是受过训练的非专家访谈员，要么是本地托管的 AI 系统；我们评估暴露度、具体性和相关性。两种条件下的访谈时长中位数相似，但 AI 访谈员在每次访谈中提出的问题数量是原来的两倍以上，并且每次访谈引出的受访者文本大约多 65%。在人类访谈员条件下，每个问题的回答更长，但在多重检验校正后并非如此；只有 AI 条件下的参与者跳过了问题。两种条件下，回答几乎都切题；就具体性而言，我们未检测到差异，但该设计无法排除高达四分之三点（即 0.75 分）的差异。这些结果是用一个本地托管的小型开放权重模型（70 亿参数）获得的，这回应了关于数据安全、GDPR 合规性和科学可重复性的担忧。我们的发现支持 AI 访谈用于大样本（large-N）定性数据收集的可行性。
cs.AI / 71 / 2610.22968
NL2SQL with Iterative Self-Refinement
基于迭代式自我精炼的 NL2SQL
large language model
大语言模型相关
Abstract
Bridging the gap between human intent and machine logic remains a critical challenge in data access for non-technical users, yet Large Language Models (LLMs) remain unreliable: they hallucinate schema elements, generate syntactically invalid queries, or fail to capture user intent. These limitations are especially critical in domains such as insurance, where KPIs, regulatory constraints, and legacy relational systems demand high precision. We introduce a multi-agent, persona-driven NL2SQL framework that combines (1) domain relevance detection, (2) schema-grounded knowledge construction, (3) KPI-aware SQL synthesis, (4) syntax-level self-repair using sqlite execution feedback, and (5) semantic consistency checking via SQL-to-NL reverse engineering. Our system significantly reduces query hallucination and improves semantic fidelity compared to direct LLM prompting. Evaluated on 60 queries (50 from a HuggingFace text-to-SQL dataset and 10 custom domain queries), the system achieves 90% correctness using Mistral across seven coordinated LLM calls showing that this approach can substantially increase the reliability of NL2SQL pipelines in high-stakes environments.
Chinese Translation
弥合人类意图与机器逻辑之间的差距仍然是非技术用户进行数据访问时的一项关键挑战，然而大型语言模型（LLMs）仍然不可靠：它们会虚构模式元素、生成语法无效的查询，或无法捕捉用户意图。这些局限在保险等领域尤为关键，在这些领域中，KPI、监管约束和遗留关系系统要求高精度。我们引入了一个多智能体、角色驱动的 NL2SQL 框架，它结合了（1）领域相关性检测，（2）基于模式的知识构建，（3）KPI 感知的 SQL 合成，（4）利用 sqlite 执行反馈进行语法级自我修复，以及（5）通过 SQL 到自然语言逆向工程进行语义一致性检查。与直接进行 LLM 提示相比，我们的系统显著减少了查询幻觉并提高了语义保真度。在 60 个查询（其中 50 个来自 HuggingFace 文本到 SQL 数据集，10 个为自定义领域查询）上评估时，该系统使用 Mistral 在七次协同的 LLM 调用中实现了 90% 的正确率，表明这种方法可以显著提高高风险环境中 NL2SQL 流水线的可靠性。
cs.LG / 72 / 2610.23047
DIFFLET: A Fast Inference Engine for Diffusion Transformer Models on AWS Trainium
DIFFLET：面向 AWS Trainium 上扩散 Transformer 模型的快速推理引擎
diffusion
扩散模型相关
Abstract
On AWS Trainium2, every component of a diffusion model runs as a program compiled ahead of time for fixed shapes, while pipelines differ across models, prompt lengths across requests, and caching decisions across denoising steps. We present Difflet, a serving engine in which five model pipelines share one compilation backend, joint attention is split across cores by attention head, and a host controller decides at each denoising step whether to run the denoiser or reuse its earlier output. A small probe program computes the caching signal on the device and returns one number per step, and a resident model process keeps every program loaded across requests. Across five image and video model families, fused attention accelerates denoiser steps by 1.2-4.3x over the same framework using PyTorch/XLA attention, and at 32.8k visual tokens it is the only attention that compiles. On four of the five models, resident serving is 1.5-3.4x faster than standalone runs that reload weights from a warm file cache for each generation.
Chinese Translation
在 AWS Trainium2 上，扩散模型的每个组件都作为一个针对固定形状提前编译的程序运行，而不同模型之间的流水线不同、不同请求之间的提示长度不同、不同去噪步骤之间的缓存决策也不同。我们提出 Difflet，这是一个服务引擎，其中五条模型流水线共享同一个编译后端，联合注意力按注意力头拆分到各个核心上，并由一个主机控制器在每个去噪步骤决定是运行去噪器还是复用其先前的输出。一个小型探测程序在设备上计算缓存信号，并在每一步返回一个数字，而一个常驻模型进程使每个程序在跨请求时保持加载状态。在五个图像与视频模型系列上，相比使用 PyTorch/XLA 注意力的同一框架，融合注意力将去噪器步骤加速 1.2-4.3 倍，并且在 32.8k 个视觉 token 时，它是唯一能够编译的注意力实现。在五个模型中的四个上，常驻服务比在每次生成时都从热文件缓存重新加载权重的独立运行快 1.5-3.4 倍。
cs.CL / 73 / 2610.22715
Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation
风险与拒绝：大语言模型对自杀意念的回应机制
large language model
大语言模型相关
Abstract
Refusal is a common safety mechanism in large language models (LLMs), yet in contexts of suicidal ideation (SI), refusing to engage can itself constitute a safety failure. We evaluate how eight instruction-tuned LLMs respond to SI and self-harm disclosures across three datasets, varying whether models are informed that a disclosure expresses SI and whether they receive a clinician-developed response protocol, and complement this with mechanistic analyses in six open-weight models. Our results reveal a nuanced relationship between refusal and response-level risk. At baseline, elevated risk in model responses is driven primarily by inattention to the disclosure, alongside refusal to engage. Mechanistic analyses suggest that models may refuse SI disclosures through the same direction that mediates their refusals of harmful requests. However, informing models of the SI context, or suppressing refusal directly, increases engagement without reliably lowering response risk, and may raise it. In contrast, the clinician-guided protocol consistently reduces response-level risk while also reducing refusal. Together, these analyses and ablations motivate the need for models that refuse only when warranted and engage in interactionally safe ways when users disclose SI.
Chinese Translation
拒绝是大语言模型（LLM）中常见的安全机制，但在自杀意念（SI）情境下，拒绝参与本身就可能构成一种安全失败。我们评估了八个指令微调的 LLM 在三个数据集上如何回应 SI 与自残披露，改变模型是否被告知某项披露表达了 SI，以及模型是否接收到由临床医生开发的回应方案，并用六个开放权重模型中的机制分析加以补充。我们的结果揭示了拒绝与回应层面风险之间一种微妙的关系。在基线条件下，模型回应中升高的风险主要由对披露的忽视驱动，同时伴随拒绝参与。机制分析表明，模型可能通过与介导其对有害请求的拒绝相同的方向来拒绝 SI 披露。然而，告知模型 SI 情境，或直接抑制拒绝，会增加参与度，却不会可靠地降低回应风险，并且可能提高该风险。相比之下，临床医生指导的方案持续降低回应层面的风险，同时也减少拒绝。总之，这些分析与消融实验表明，需要这样的模型：仅在确有必要时拒绝，并在用户披露 SI 时以互动安全的方式参与。
cs.AI / 74 / 2610.23304
Designing for Imperfect Prompts: A Simulation-Based Evaluation of an LLM-Based Enterprise Data Assistant Using Knowledge Graphs
面向不完美提示的设计：基于知识图谱的LLM企业数据助手的仿真评估
large language model
大语言模型相关
Abstract
Enterprise data assistants using large language models (LLMs) promise democratized data access in organizations. Yet in our real-world deployment within the CFO unit's data lake at an industry partner, we observed that business users frequently pose imperfect prompts that include domain-specific vocabulary and lack schema references or structural information. This Design Science Research study examines automatic resolution using a semantic mediation layer. We design and evaluate two design variants: schema-based mediation via semantic similarity, and knowledge graph-based mediation leveraging an automatically constructed knowledge graph from the data lake. Our simulation-based evaluation using real-world data, supported by LLM-as-a-Judge assessments and human validation, shows that knowledge graph-based mediation outperforms the schema-based approach in Precision, Recall, F1, and preserving user intent. We propose a design principle for knowledge graph-based mediation, demonstrating that structured knowledge representation bridges user intent and database structure and guides the design of enterprise data assistants across skill levels.
Chinese Translation
使用大型语言模型（LLMs）的企业数据助手有望在组织中实现数据访问的民主化。然而，在我们于某行业合作伙伴CFO部门数据湖中的真实部署中，我们观察到业务用户经常提出不完美的提示，这些提示包含领域特定词汇，且缺乏模式引用或结构信息。本设计科学研究考察使用语义中介层的自动消解。我们设计并评估两种设计变体：通过语义相似度的基于模式的消解，以及利用从数据湖自动构建的知识图谱的基于知识图谱的消解。我们使用真实数据进行的基于仿真的评估，辅以LLM-as-a-Judge评估和人工验证，表明基于知识图谱的消解在Precision、Recall、F1和保留用户意图方面优于基于模式的方法。我们提出基于知识图谱的消解的设计原则，证明结构化知识表示弥合了用户意图与数据库结构之间的差距，并指导跨技能水平的企业数据助手设计。
cs.AI / 75 / 2610.23625
PresenceBench: An Exploratory Evaluation of Large Language Model Adaptation to User State
PresenceBench：大型语言模型对用户状态适应性的探索性评估
large language model
大语言模型相关
Abstract
Presence-aware intelligence is the capacity of an AI system's intelligence layer to adapt appropriately to the functional state of the person it serves (here readiness, flow, and stability) while preserving task quality and safety. We introduce PresenceBench, an instrument for measuring it. PresenceBench evaluates how the intelligence layer uses a supplied state, not how that state is sensed, and breaks the capability into measurable properties. Two are tested here. Receptivity: does the model take up the state and give a better response than with no signal? Specificity: does any improvement depend on the correct state, rather than on extra structured context? A third, adaptation across states, is a required endpoint not measured in this release. A five-arm causal design answers both questions, against each model's own baseline and an equally structured mismatched-state control, scored by a correctness-gated human paired comparison. The exploratory evaluation covered four models across eighteen single-turn, six fixed three-turn and six simulated natural-conversation scenarios. Supplying a synthetic state, raw or translated into a response policy, gave human-preference point estimates above 0.5 for all four models (Claude Sonnet 5, DeepSeek V4 Pro, Gemini 3.6 Flash, GPT-5.6 Terra), with three of eight 95% intervals excluding chance. Specificity was not resolved: no per-model interval excluded chance, and the design cannot yet separate the state values from a generic instruction to adapt. Policy delivery reduced priced cost and output tokens against each model's baseline; these are resource-use changes, not efficiency gains, because correctness non-inferiority was not established. Safety-floor failures were rare (13 of 360 cases), with a non-significant excess under policy delivery. We report the instrument, the frozen results, and the controls a confirmatory study needs.
Chinese Translation
临场感知智能（Presence-aware intelligence）是指 AI 系统智能层能够适当地适应其所服务对象的功能状态（此处为准备度、心流和稳定性），同时保持任务质量与安全的能力。我们引入 PresenceBench，一种用于测量该能力的工具。PresenceBench 评估的是智能层如何使用所提供的状态，而不是该状态如何被感知，并将该能力分解为可测量的属性。本文测试其中两个属性。接受性：模型是否会采纳该状态，并给出比没有信号时更好的响应？特异性：任何改进是否取决于正确的状态，而不是取决于额外的结构化上下文？第三个属性——跨状态适应——是一个必需的终点指标，但本版本未对其进行测量。一项五臂因果设计以每个模型自身的基线和一个结构相同但状态不匹配的对照为参照来回答这两个问题，并由经正确性门控的人类成对比较进行评分。该探索性评估覆盖四个模型，涉及十八个单轮场景、六个固定的三轮场景和六个模拟自然对话场景。提供合成状态——无论是原始形式，还是转换为响应策略——对所有四个模型（Claude Sonnet 5、DeepSeek V4 Pro、Gemini 3.6 Flash、GPT-5.6 Terra）给出了高于 0.5 的人类偏好点估计，八个 95% 区间中有三个排除了随机。特异性问题未得到解决：没有任何单模型区间排除随机，并且该设计尚无法将状态取值与泛化的适应指令区分开来。相对于每个模型自身的基线，策略传递降低了计价成本和输出 token 数；这些是资源使用方面的变化，而不是效率提升，因为正确性非劣效性并未得到确立。安全底线失败很少见（360 例中有 13 例），在策略传递下存在不显著的超额。我们报告该测量工具、冻结的结果，以及验证性研究所需的对照。
cs.LG / 76 / 2610.22512
Predicting Privileged Information Effectiveness for On-Policy Self-Distillation
预测同策略自蒸馏中特权信息的有效性
large language model
大语言模型相关
Abstract
On-policy self-distillation (OPSD) enables large language models (LLMs) to learn from privileged information (PI) without a separate teacher model. However, prior work largely treats PI as a fixed design choice, leaving the effects of its content and construction poorly understood. This challenge is particularly important in multi-turn agentic settings, where trajectories contain actions, observations, and reasoning that can be transformed into PI in many ways. We systematically evaluate 16 PI configurations across three multi-turn agentic benchmarks and three student models that vary in scale and reasoning mode. We find that PI effectiveness depends on its content, generation source, and the student model. Information-dense trajectories and PI produced by more capable generators generally yield larger gains, whereas thinking models benefit less consistently. To assess PI before training, we compare PI utility with two KL-based metrics: sampled-token reverse KL ($D$) and error-step KL contrast ($Δd^{-}_{\mathrm{err}}$). The KL-based metrics track post-training performance more consistently than PI utility, which can be distorted by information leakage. Empirically, selecting each task's PI by $Δd^{-}_{\mathrm{err}}$ also improves post-training performance over random selection, most when the PI candidates differ widely. Finally, effective PI remains beneficial when combined with GRPO, allowing the combined objective to outperform a strong GRPO baseline. Overall, our results provide practical guidance for constructing and evaluating PI in OPSD training.
Chinese Translation
同策略自蒸馏（OPSD）使大型语言模型（LLMs）无需单独的教师模型即可从特权信息（PI）中学习。然而，先前的工作大多将 PI 视为固定的设计选择，导致对其内容和构建方式的影响理解不足。这一挑战在多轮智能体设置中尤为重要，因为轨迹包含动作、观察和推理，而这些可以以多种方式转化为 PI。我们在三个多轮智能体基准和三个在规模与推理模式上各不相同的学生模型上，系统评估了 16 种 PI 配置。我们发现，PI 的有效性取决于其内容、生成来源以及学生模型。信息密集的轨迹以及由能力更强的生成器产生的 PI 通常带来更大的增益，而思考模型受益较不稳定。为了在训练前评估 PI，我们将 PI 效用与两个基于 KL 的指标进行比较：采样 token 反向 KL（$D$）和错误步 KL 对比（$Δd^{-}_{\mathrm{err}}$）。基于 KL 的指标比 PI 效用更一致地跟踪训练后性能，而 PI 效用可能会被信息泄漏扭曲。从经验上看，按 $Δd^{-}_{\mathrm{err}}$ 为每个任务选择 PI 也比随机选择更能提升训练后性能，当 PI 候选差异较大时提升最明显。最后，有效的 PI 在与 GRPO 结合时仍然有益，使组合目标能够优于强 GRPO 基线。总体而言，我们的结果为在 OPSD 训练中构建和评估 PI 提供了实用指导。
cs.LG / 77 / 2610.22543
On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction
On the SPOT：通过提交顺序预测增强扩散语言模型的推测解码
diffusion
扩散模型相关
Abstract
Diffusion language models (DLMs) promise parallel text generation, but in practice face an unresolved speed-quality trade-off. Exact decoding fills in only one masked position per model pass, while faster multi-token decoding schemes are lossy. Speculative decoding is the standard lossless acceleration for autoregressive models. However, it does not transfer directly to DLMs, which must choose both what to generate and where. We show that this order is the missing piece: the model's own predictions from the previous pass already give the right token at 92-100% of positions, but not which position comes next. Based on this observation, we introduce Speculating Positions Order Trajectory (SPOT), which learns to predict this order. A 0.62M-parameter model reads only the target's output distributions and proposes a tree of candidate continuations, and the target verifies all of them in a single pass by checking whether each equals its own next state. On the verification side, our analysis shows that prior block-diffusion verification approaches either assume $T=0$ or a sequential proposal interface. To support tree-based speculation at any temperature, we propose using coupled-noise match verification for diffusion blocks, following shared-Gumbel-noise coupling approaches. Our comprehensive experiments show that SPOT cuts target passes by up to 6.8x and reaches up to 3.2x the throughput of standard decoding, beating the calibrated and training-free drafting baselines in every task and budget at $T=0$ and $T=1$. In addition, we show that the learned drafter transfers across models, demonstrating successful transfer across six different DLMs. Project page: https://eliezer318.github.io/SPOT-dLLM/
Chinese Translation
扩散语言模型（DLMs）有望实现并行文本生成，但在实践中面临一个尚未解决的速度-质量权衡。精确解码在每次模型传递中仅填充一个掩码位置，而更快的多 token 解码方案是有损的。推测解码是自回归模型的标准无损加速方法。然而，它不能直接迁移到 DLMs，因为 DLMs 必须同时选择生成什么以及在哪里生成。我们表明，这一顺序正是缺失的一环：模型在上一轮传递中的自身预测已经在 92-100% 的位置上给出了正确的 token，但并未给出下一个是哪个位置。基于这一观察，我们提出 Speculating Positions Order Trajectory（SPOT），它学习预测这一顺序。一个 0.62M 参数的模型仅读取目标模型的输出分布，并提议一棵候选延续树，而目标模型在单次传递中通过检查每个候选是否等于其自身的下一状态来验证所有这些候选。在验证方面，我们的分析表明，先前的块扩散验证方法要么假设 $T=0$，要么假设一种顺序提议接口。为了在任意温度下支持基于树的推测，我们提出对扩散块使用耦合噪声匹配验证，遵循共享 Gumbel 噪声耦合方法。我们的综合实验表明，SPOT 将目标模型传递次数最多削减 6.8 倍，并达到标准解码吞吐量的最高 3.2 倍，在 $T=0$ 和 $T=1$ 下，在每个任务和预算中都优于经校准的和免训练的起草基线。此外，我们表明学习到的起草器可以跨模型迁移，展示了在六种不同 DLMs 上的成功迁移。项目页面：https://eliezer318.github.io/SPOT-dLLM/
cs.LG / 78 / 2610.22599
Can LLMs predict the outcomes of machine learning experiments?
LLMs 能预测机器学习实验的结果吗？
large language model
大语言模型相关
Abstract
Recent advances in AI make Recursive Self-Improvement (RSI) increasingly plausible. Large Language Models (LLMs) can already propose, select, and run Machine Learning (ML) experiments. However, the space of possible design choices in ML is vast. An LLM that proposes an experiment needs a reliable intuition of whether it is worth running given the timing constraints of long training runs. In this work, we measure such intuition in current frontier LLMs by evaluating how accurately they predict the impact of self-generated ML experiment ideas before these are put into training runs. To this end, we construct a dataset of 288 LLM-proposed ideas on eight MLE-bench tasks. We then prompt five frontier LLMs (GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, Kimi K3, and GLM-5.3) to predict the impact of each idea and compare their predictions with the actual outcomes after training. We report four main findings. First, LLM predictions are substantially better than random chance but far from accurate: even the best model, Claude Opus 5.5, explains only half of the variance in outcomes (R^2 = 0.49). Second, we notice that all tested models systematically underestimate the magnitude of the changes. Third, prediction quality varies widely across tasks. Finally, the choice of the coding harness that actually implements the ideas has little effect on outcomes, and we find no evidence of self-preference. Higher reasoning settings and additional context lead to modest gains. Overall, our results show that current LLMs have useful but limited intuition about which ML ideas will work. We release our dataset to support further study.
Chinese Translation
AI 的最新进展使递归自我改进（Recursive Self-Improvement, RSI）变得越来越可能。大型语言模型（LLMs）已经能够提出、选择并运行机器学习（ML）实验。然而，ML 中可能的设计选择空间非常庞大。考虑到长时间训练运行的时间限制，一个提出实验的 LLM 需要可靠的直觉来判断该实验是否值得运行。在本工作中，我们通过评估当前前沿 LLM 在自生成的 ML 实验想法被投入训练运行之前对其影响的预测准确程度，来测量这种直觉。为此，我们基于八个 MLE-bench 任务构建了一个包含 288 个由 LLM 提出的想法的数据集。然后，我们提示五个前沿 LLM（GPT-6 Astra、Claude Opus 5.5、Claude Fable 5.1、Kimi K3 和 GLM-5.3）预测每个想法的影响，并将它们的预测与训练后的实际结果进行比较。我们报告四个主要发现。第一，LLM 的预测显著优于随机猜测，但远非准确：即使是最好的模型 Claude Opus 5.5，也只能解释结果方差的一半（R^2 = 0.49）。第二，我们注意到所有测试模型都系统性地低估了变化的幅度。第三，预测质量在不同任务之间差异很大。最后，实际实现这些想法的编码框架（coding harness）的选择对结果几乎没有影响，并且我们没有发现自我偏好的证据。更高的推理设置和额外的上下文会带来适度的提升。总体而言，我们的结果表明，当前 LLM 对于哪些 ML 想法会奏效具有有用但有限的直觉。我们发布我们的数据集以支持进一步研究。
cs.LG / 79 / 2610.22630
Network-Wide Freeway Speed Nowcasting and Forecasting from Sparse Fixed Sensors with a Probe-Supervised Graph Convolutional Recurrent Network
利用稀疏固定传感器和探针监督图卷积循环网络进行路网级高速公路速度临近预报与预测
diffusion
扩散模型相关
Abstract
Traffic agencies need current and near-future speeds on every link of their network, but their fixed sensors cover only part of it. We study joint nowcasting and next-hour forecasting of hourly speeds on all 4{,}127 directional freeway links of the Dallas--Fort Worth region from sensors on 643 links (15.6%). Historical probe-vehicle data supervise every link in training; only sensor readings are used at inference. We call this setting probe-supervised sparse sensing. We propose GCRN+, a graph convolutional recurrent network that predicts deviations from a historical speed profile. It combines a parameter-free, direction-aware kernel interpolation of sensor anomalies, used as input features and as an output baseline, with a diffusion-convolutional GRU on the directed road graph, a low-rank adaptive graph and learnable link embeddings. It scales linearly with network size and trains on a CPU. Sensor readings are emulated from probe speeds, which isolates spatio-temporal inference from detector calibration. On four held-out test days, GCRN+ attains an RMSE of 2.21~mph for the current hour and 2.39~mph for the next hour on unsensored links, against 3.23~mph for the historical average. The best baselines that use the same interpolation features reach 2.53 and 2.75~mph. Adapted recent models without these features (STID, STAEformer, ImputeFormer, MoGERNN) reach at best 2.54 and 2.53~mph; this comparison is between complete pipelines rather than architectures. GCRN+ is best in every distance band except links not connected to any sensor through the road graph. Ablations within one recurrent backbone attribute most of the gain to the link embeddings and the adaptive graph, which can be trained because probe labels exist for every link. For congestion detection on unsensored links, recall rises from 77\% (historical average) and 86% (best competitor) to 89% at aprecision of 0.94.
Chinese Translation
交通管理机构需要其路网中每条路段当前和近未来的速度，但其固定传感器仅覆盖其中一部分。我们研究基于643条路段（15.6%）上的传感器，对达拉斯--沃斯堡地区全部4{,}127条定向高速公路路段的逐小时速度进行联合临近预报和下一小时预测。历史探针车辆数据在训练中监督每条路段；在推理时仅使用传感器读数。我们将此设定称为探针监督稀疏感知。我们提出GCRN+，一种图卷积循环网络，用于预测相对于历史速度剖面的偏差。它将一种无参数、方向感知的传感器异常值核插值（既用作输入特征，也用作输出基线）与有向道路图上的扩散卷积GRU、低秩自适应图和可学习路段嵌入相结合。其计算规模随网络规模线性扩展，并可在CPU上训练。传感器读数由探针速度模拟，这将时空推断与检测器校准隔离开来。在四个留出的测试日上，GCRN+在无传感器路段上对当前小时和下一小时分别达到2.21~mph和2.39~mph的RMSE，而历史平均为3.23~mph。使用相同插值特征的最佳基线分别达到2.53和2.75~mph。经适配且不含这些特征的近期模型（STID、STAEformer、ImputeFormer、MoGERNN）最好分别达到2.54和2.53~mph；这一比较是在完整流程之间而非架构之间进行的。除通过道路图未与任何传感器连通的路段外，GCRN+在每个距离带中都是最佳的。在一个循环骨干网络内进行的消融实验将大部分增益归因于路段嵌入和自适应图，它们能够被训练是因为每条路段都存在探针标签。对于无传感器路段上的拥堵检测，在精确率为0.94时，召回率从77\%（历史平均）和86%（最佳竞争者）上升到89%。
cs.LG / 80 / 2610.22684
AlphaBet: Reinforcement Learning for LLM Event Forecasting Agents
AlphaBet：面向大语言模型事件预测智能体的强化学习
large language model
大语言模型相关
Abstract
Forecasting an unresolved event requires finding relevant evidence and deciding how strongly it supports each outcome. Yet outcome-based training of Large Language Model (LLM) forecasters often begins with preassembled evidence at a single prediction date, leaving evidence acquisition outside the learning loop. We introduce AlphaBet, an agentic reinforcement learning (RL) framework that trains a single LLM policy to search, reason, and report probabilities across multiple pre-resolution dates. Each dated forecast is scored by a proper scoring rule against the eventual outcome, allowing one resolved event to supervise decisions under different information states. We decompose forecasting regret into an evidence-acquisition gap and a probability-reporting gap, and show that Group Relative Policy Optimization (GRPO)'s standard-deviation normalization can distort the reporting target under multi-class proper scores. On GLM-4.5-Air, AlphaBet reduces the Brier score from 0.731 to 0.626 on Forecast-Dojo and from 0.612 to 0.528 on FutureX. Controlled ablations attribute these gains to multi-date training and show that they come from improvements in both information acquisition and probability reporting. Our empirical results validate that GRPO's std normalization degrades proper-scoring-rule performance while leaving top-outcome accuracy largely unchanged.
Chinese Translation
预测一个尚未解决的事件需要找到相关证据，并判断这些证据对每种结果的支持程度有多强。然而，基于结果的大语言模型（LLM）预测器训练往往从单一预测日期上预先组装好的证据开始，从而使证据获取被置于学习循环之外。我们提出 AlphaBet，一个智能体强化学习（RL）框架，它训练单个 LLM 策略在多个解析前日期上进行搜索、推理并报告概率。每个带日期的预测都依据适当评分规则相对于最终结果进行打分，从而使一个已解析事件能够在不同信息状态下监督决策。我们将预测遗憾分解为证据获取差距和概率报告差距，并证明在多类适当评分下，组相对策略优化（GRPO）的标准差归一化会扭曲报告目标。在 GLM-4.5-Air 上，AlphaBet 将 Forecast-Dojo 上的 Brier 分数从 0.731 降至 0.626，并将 FutureX 上的分数从 0.612 降至 0.528。受控消融实验将这些增益归因于多日期训练，并表明这些增益来自信息获取和概率报告两方面的改进。我们的实证结果验证了 GRPO 的标准差归一化会降低适当评分规则下的性能，同时基本不改变最高结果准确率。
cs.LG / 81 / 2610.22711
Normalizing Flow Maps
归一化流映射
diffusion
扩散模型相关
Abstract
Scalable approaches to learning neural dynamical systems have placed diffusion, flow matching, and few-step generators in flow maps at the forefront of generative modeling. Such state-of-the-art models have limited access to explicit model densities despite evolving from classical generative modeling approaches that leveraged tractable likelihoods, e.g., Normalizing Flows and Autoregressive models. In this paper, we combine these threads of few-step generation and tractable sample likelihoods and introduce Normalizing Flow Maps (NFM), a class of exactly invertible pushforward maps that generalizes normalizing flows and flow maps. More precisely, NFMs learn the integrator of an underlying dynamical system, but crucially, they provide efficient exact likelihood evaluation by construction. To enable scalable training of NFMs, we derive two complementary axes of consistency inherent to NFMs that yield (self) distillation objectives: (1) pointwise matching, which unlocks $\ell_2$-based flow-map style regression losses, and (2) distributional matching, which matches pushforward distributions to the marginals of a prescribed forward process. On conditional image generation on CIFAR-10 and ImageNet-64, NFMs achieve state-of-the-art sample quality among parameter-matched normalizing flows, while supporting likelihood-based RLHF. For equilibrium sampling in Boltzmann Generators, NFMs outperform prior state-of-the-art normalizing flows on both single-system and transferable peptide benchmarks.
Chinese Translation
可扩展的神经动力系统学习方法已将扩散、流匹配以及流映射中的少步生成器置于生成建模的最前沿。此类最先进模型对显式模型密度的获取能力有限，尽管它们源自利用可处理似然的经典生成建模方法，例如归一化流和自回归模型。在本文中，我们将少步生成与可处理样本似然这两条脉络结合起来，并引入归一化流映射（Normalizing Flow Maps, NFM），这是一类精确可逆的前推映射，推广了归一化流和流映射。更准确地说，NFM 学习底层动力系统的积分器，但关键在于，它们在构造上就提供了高效的精确似然评估。为了实现 NFM 的可扩展训练，我们推导出 NFM 固有的两个互补的一致性轴，它们产生（自）蒸馏目标：(1) 逐点匹配，它解锁了基于 $\ell_2$ 的流映射风格回归损失；(2) 分布匹配，它将前推分布与给定前向过程的边缘分布相匹配。在 CIFAR-10 和 ImageNet-64 的条件图像生成上，NFM 在参数匹配的归一化流中实现了最先进的样本质量，同时支持基于似然的 RLHF。对于 Boltzmann 生成器中的平衡采样，NFM 在单系统和可迁移肽基准上都优于先前最先进的归一化流。
cs.LG / 82 / 2610.22727
OrderFlowDiT: Transport-Regularized Diffusion Transformer for Probabilistic Limit-Order Supply Forecasting
OrderFlowDiT：用于概率性限价订单供给预测的传输正则化扩散 Transformer
diffusion
扩散模型相关
Abstract
We study the conditional distribution of limit-order volume submitted over the next five minutes, given the current order book and recent order flow. The task requires capturing price placement, total volume, and predictive uncertainty jointly. Coordinate-wise diffusion losses do not explicitly account for price displacement, while normalizing fields to unit mass removes information about total volume. Agreement on selected summary statistics also does not establish fidelity to the joint field distribution. We introduce OrderFlowDiT, a conditional Diffusion Transformer for unnormalized bid and ask volume-at-price fields, trained with an unbalanced optimal-transport (UOT) auxiliary loss. We also develop a full-field evaluation framework that combines a transport cost sensitive to price placement and total volume with proper scoring rules and calibration diagnostics. In the primary comparison on 1950 validation conditions across 30 U.S. equities, adding UOT reduces full-field transport cost by 45.8% and the transport energy score by 26.1% relative to training with mean squared error alone. Both measures improve for every stock. These findings support incorporating price geometry and volume into generative modeling, with field accuracy and predictive uncertainty evaluated together.
Chinese Translation
我们研究在给定当前订单簿和近期订单流的条件下，未来五分钟内提交的限价订单量的条件分布。该任务要求联合刻画价格挂单位置、总量和预测不确定性。逐坐标的扩散损失并未显式考虑价格位移，而将字段归一化为单位质量则会移除关于总量的信息。在选定的汇总统计量上达成一致，也不能确立对联合字段分布的保真度。我们提出 OrderFlowDiT，一种面向未归一化买价与卖价价格-成交量字段的条件扩散 Transformer，并使用非平衡最优传输（UOT）辅助损失进行训练。我们还开发了一个全字段评估框架，该框架将对价格挂单位置和总量敏感的传输成本与适当的评分规则及校准诊断相结合。在覆盖 30 只美国股票的 1950 个验证条件上的主要比较中，与仅使用均方误差训练相比，加入 UOT 使全字段传输成本降低 45.8%，传输能量分数降低 26.1%。这两项指标对每只股票均有改善。这些发现支持将价格几何与成交量纳入生成建模，并将字段准确性与预测不确定性一起评估。
cs.LG / 83 / 2610.22763
Acquisition Function-Guided Generation for Bayesian Optimization over Discrete Spaces
面向离散空间贝叶斯优化的采集函数引导生成
diffusion
扩散模型相关
Abstract
Bayesian optimization (BO) over discrete spaces is challenging due to missing gradients and the sparsity of valid solutions. Latent-space BO sidesteps both issues by fitting a surrogate in the latent space of a generative model, which however substantially complicates the direct incorporation of domain knowledge that is naturally expressed in the original discrete space. In this work, we instead combine a pretrained generative model with a surrogate defined directly in data space, where inductive biases such as domain-specific kernels can be readily incorporated. Rather than maximizing the acquisition function over the discrete design space, our method samples candidates from the generative model under a distribution tilted by the acquisition function. To recover useful gradients for sampling, we embed this tilted distribution into a diffusion bridge that yields a smooth, posterior-averaged acquisition function. This leads to \emph{acquisition function guided generation} (AFGG), which we instantiate with a categorical diffusion model and an autoregressive normalizing flow. Across a broad suite of molecular benchmarks, the diffusion model variant performs on par with leading BO methods, while the flow model variant outperforms state-of-the-art methods in most cases and remains highly competitive otherwise. Our implementation is available at https://anonymous.4open.science/r/afgg.
Chinese Translation
离散空间上的贝叶斯优化（BO）由于缺少梯度以及有效解的稀疏性而具有挑战性。潜空间BO通过在生成模型的潜空间中拟合代理模型来回避这两个问题，然而这极大复杂化了直接融入那些自然地在原始离散空间中表达的领域知识。在这项工作中，我们转而将预训练的生成模型与直接在数据空间中定义的代理模型相结合，其中诸如领域特定核等归纳偏置可以很容易地被纳入。我们的方法不是在整个离散设计空间上最大化采集函数，而是在由采集函数倾斜的分布下从生成模型中采样候选。为了恢复用于采样的有用梯度，我们将该倾斜分布嵌入到一个扩散桥中，从而产生一个平滑的、后验平均的采集函数。这引出了\emph{采集函数引导生成}（AFGG），我们用分类扩散模型和自回归归一化流对其进行实例化。在广泛的一组分子基准上，扩散模型变体与领先的BO方法表现相当，而流模型变体在大多数情况下优于最先进的方法，并在其他情况下仍具有高度竞争力。我们的实现可在 https://anonymous.4open.science/r/afgg 获取。
cs.LG / 84 / 2610.22885
Learning to Be Guided: Shaping Latent Geometry for Classifier-Free Guidance
学习被引导：为无分类器引导塑造潜在几何
diffusion
扩散模型相关
Abstract
Classifier-free guidance (CFG) is a standard component of conditional diffusion models, and guided generation performance has become the \textit{de facto} measure of state-of-the-art models. Recent work on latent diffusion models seeks to learn latent spaces that are easier to model with diffusion. However, under CFG, generation additionally depends on how conditions are arranged in the latent space. We study this arrangement under discrete class conditions, termed \emph{class geometry}, and investigate what class geometry allows a latent space to work well with CFG. Under a Gaussian class model, we show that shifting variance from intra- to inter-class at fixed total variance makes classes more distinguishable under noise, and derive a noisy posterior objective whose per-sample descent direction is proportional to the oracle CFG correction. Based on the analysis, we propose Prototype Regularization (PR), a lightweight surrogate for the intractable noisy posterior objective, which applies a cosine prototype classifier to spatially pooled latents. With a controlled training budget (20 epochs for VAE, 80 epochs for DiT), adding PR to an LDM baseline reduces guided gFID on ImageNet from 3.31 to 2.28, surpassing a DINOv3-aligned VA-VAE (2.44) with negligible overhead. We further demonstrate that PR generalizes to pseudo-labels derived from clustering, continuous text-embedding conditions, and video latents. Together, our findings highlight class geometry as a critical factor in guided generation and take a first step toward CFG-compatible representation learning. Our project page is available at: https://maps-research.github.io/cfg-geometry/.
Chinese Translation
无分类器引导（classifier-free guidance，CFG）是条件扩散模型的标准组成部分，而引导生成性能已成为最先进模型事实上的衡量标准。近期关于潜在扩散模型的工作试图学习更容易用扩散建模的潜在空间。然而，在 CFG 下，生成还额外依赖于条件在潜在空间中如何排列。我们在离散类别条件下研究这种排列，称之为类别几何，并探究什么样的类别几何能使潜在空间与 CFG 良好配合。在高斯类别模型下，我们表明，在总方差固定的情况下，将方差从类内转移到类间会使类别在噪声下更具可区分性，并推导出一个含噪后验目标，其逐样本下降方向与 oracle CFG 校正成正比。基于该分析，我们提出原型正则化（Prototype Regularization，PR），它是难以处理的含噪后验目标的一种轻量级代理，将余弦原型分类器应用于空间池化后的潜在表示。在受控训练预算下（VAE 为 20 个 epoch，DiT 为 80 个 epoch），将 PR 加入 LDM 基线可将 ImageNet 上的引导 gFID 从 3.31 降至 2.28，以可忽略的开销超过与 DINOv3 对齐的 VA-VAE（2.44）。我们进一步证明，PR 可泛化到由聚类得到的伪标签、连续文本嵌入条件和视频潜在表示。总之，我们的发现突显了类别几何作为引导生成中的关键因素，并朝着与 CFG 兼容的表示学习迈出了第一步。我们的项目页面见：https://maps-research.github.io/cfg-geometry/。
cs.LG / 85 / 2610.22945
Large Diffusion Models Need More Than EMA
大型扩散模型需要的不仅仅是 EMA
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Large-scale diffusion models are commonly trained with a constant learning rate (LR) and exponential moving average (EMA), despite the strong performance of LR decay and Schedule-Free optimization in training large language models. We study whether EMA is uniquely suited to diffusion training, or whether stronger alternatives have been underexplored. Our theory shows that averaging a constant-step trajectory has an inherent optimization limitation, whereas LR decay and Schedule-Free optimization admit vanishing stationarity guarantees under standard nonconvex assumptions. Empirically, LR decay and EMA provide complementary benefits: combining decay with EMA further improves the standard baseline at most tested starting points, although the gain remains sensitive to when decay begins. Schedule-Free avoids this timing requirement and consistently outperforms EMA with a constant LR across training stages, achieving both lower fixed-evaluation objectives and better generation quality. Our analysis further shows that all three strategies substantially reduce parameter variability, yet stronger stabilization is not always better: EMA with a constant LR retains comparatively higher parameter variability, while aggressive decay combined with EMA achieves the lowest checkpoint variability but also weakens continued optimization. Schedule-Free substantially reduces parameter variability while preserving more late-stage progress, providing a better balance between parameter stability and continued optimization. Finally, the benefits of Schedule-Free extend beyond pretraining to ReFL and generator-side DMD, demonstrating its transferability across distinct diffusion optimization settings. Overall, our results suggest that large diffusion models need more than parameter averaging alone: effective optimization requires balancing variance reduction with continued optimization.
Chinese Translation
大规模扩散模型通常使用恒定学习率（LR）和指数移动平均（EMA）进行训练，尽管 LR 衰减和 Schedule-Free 优化在训练大型语言模型时表现出强劲性能。我们研究 EMA 是否特别适合扩散训练，或者更强的替代方案是否尚未得到充分探索。我们的理论表明，对恒定步长轨迹进行平均存在固有的优化局限性，而 LR 衰减和 Schedule-Free 优化在标准非凸假设下可提供趋于零的平稳性保证。在实证上，LR 衰减和 EMA 提供互补的益处：将衰减与 EMA 结合在大多数测试起始点上进一步改进了标准基线，尽管增益仍然对衰减何时开始敏感。Schedule-Free 避免了这一时机要求，并在各训练阶段始终优于使用恒定 LR 的 EMA，既实现了更低的固定评估目标，又获得了更好的生成质量。我们的分析进一步表明，三种策略都大幅降低了参数变异性，但更强的稳定化并不总是更好：使用恒定 LR 的 EMA 保持相对更高的参数变异性，而激进衰减与 EMA 结合则实现了最低的检查点变异性，但也削弱了持续优化。Schedule-Free 大幅降低参数变异性，同时保留更多后期进展，在参数稳定性和持续优化之间提供了更好的平衡。最后，Schedule-Free 的益处超越了预训练，扩展到 ReFL 和生成器侧 DMD，展示了其在不同扩散优化设置中的可迁移性。总体而言，我们的结果表明，大型扩散模型需要的不仅仅是参数平均：有效的优化需要在方差降低与持续优化之间取得平衡。
cs.LG / 86 / 2610.23122
LED: Improving Masked Diffusion Language Models by Lookahead Editing
LED：通过前瞻编辑改进掩码扩散语言模型
diffusionlarge language model
扩散模型相关
大语言模型相关
Abstract
Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models, enabling non-autoregressive text generation through parallel decoding. Existing dLLMs can be broadly categorized into two families: masked diffusion models (MDMs) and uniform diffusion models (UDMs) In this work, we theoretically show that MDMs are easier to optimize due to their mask-based transition schedule but have lower computational expressivity, whereas UDMs enable iterative state revision and greater expressivity but are substantially harder to train. This trade-off raises a natural question: can we obtain the expressivity of iterative revision without giving up the simple, easy-to-optimize transition schedule of MDMs? Motivated by this analysis, we propose Lookahead Editing (\ours{}), which augments masked diffusion decoding with a rewritable proposal state. At each decoding step, the model performs several lookahead updates to iteratively refine future predictions in this proposal state before committing tokens to the masked sequence. This design introduces a rewritable proposal that uses additional compute to improve expressivity while keeping the simple mask-based transition schedule of MDMs. We fine-tune Qwen2.5-7B and Qwen3-8B to obtain LED-\texttt{7B} and LED-\texttt{8B}, respectively. Across eight benchmarks, LED SFT yields average gains of $+11.4$ and $+7.7$ points over standard MDM SFT. Moreover, using SFT alone, LED-\texttt{8B} outperforms LLaDA2.1-16B Q-mode, which is trained with continued pre-training, supervised fine-tuning, and reinforcement learning.
Chinese Translation
扩散大语言模型（dLLMs）已成为自回归语言模型的一种有前景的替代方案，通过并行解码实现非自回归文本生成。现有的 dLLMs 大致可分为两类：掩码扩散模型（MDMs）和均匀扩散模型（UDMs）。在本工作中，我们从理论上表明，MDMs 由于其基于掩码的转移调度而更容易优化，但计算表达能力较低；而 UDMs 能够进行迭代状态修正并具有更强的表达能力，但训练难度显著更高。这种权衡引出了一个自然的问题：我们能否在不放弃 MDMs 简单且易于优化的转移调度的同时，获得迭代修正的表达能力？受此分析启发，我们提出前瞻编辑（\ours{}），它通过一个可重写的提议状态来增强掩码扩散解码。在每个解码步骤中，模型在该提议状态中执行若干次前瞻更新，以迭代地细化未来的预测，然后再将词元提交到掩码序列中。该设计引入了一个可重写的提议，它利用额外的计算来提高表达能力，同时保留 MDMs 简单的基于掩码的转移调度。我们对 Qwen2.5-7B 和 Qwen3-8B 进行微调，分别得到 LED-\texttt{7B} 和 LED-\texttt{8B}。在八个基准上，LED SFT 相较于标准 MDM SFT 取得了 $+11.4$ 和 $+7.7$ 个点的平均增益。此外，仅使用 SFT，LED-\texttt{8B} 就优于 LLaDA2.1-16B Q-mode，后者通过持续预训练、监督微调和强化学习进行训练。
cs.LG / 87 / 2610.23145
Local proxies for memory-efficient training
用于内存高效训练的局部代理
large language model
大语言模型相关
Abstract
Local learning algorithms can reshape resource requirements of artificial intelligence. The memory bottleneck of standard end-to-end backpropagation lies in the need to store all network activations until the loss gradient has been propagated back through. Local learning avoids this, but typically at a cost in accuracy. We study how to achieve accurate and memory efficient local learning in transformers. For each transformer block, we train a local proxy network to predict the global network output via distillation. Gradients can be independently propagated through this local proxy while estimating well the backpropagation gradients. In theory, this approach is exact for deep linear networks; in non-linear residual models, the gradient error can be bounded when using the appropriate residual proxy. We verify this numerically and demonstrate that optimizing the distillation loss improves gradient estimation. Evaluating the performance of proxy propagation on CIFAR-100 in vision transformers, we find that it achieves a better accuracy/memory trade-off compared to other local algorithms. Finally, we show that the method is further relevant for fine-tuning large language models on a small memory budget. Our method trains faster and achieves higher performance than memory-efficient zeroth-order learning algorithms with comparable memory cost.
Chinese Translation
局部学习算法可以重塑人工智能的资源需求。标准端到端反向传播的内存瓶颈在于，需要存储所有网络激活，直到损失梯度完成反向传播。局部学习避免了这一点，但通常以精度为代价。我们研究如何在 Transformer 中实现准确且内存高效的局部学习。对于每个 Transformer 块，我们训练一个局部代理网络，通过蒸馏来预测全局网络的输出。梯度可以通过这个局部代理独立传播，同时很好地估计反向传播梯度。在理论上，这种方法对于深度线性网络是精确的；在非线性残差模型中，当使用适当的残差代理时，梯度误差可以被界定。我们通过数值实验验证了这一点，并证明优化蒸馏损失可以改善梯度估计。在视觉 Transformer 中评估代理传播在 CIFAR-100 上的性能时，我们发现与其他局部算法相比，它实现了更好的精度/内存权衡。最后，我们表明该方法进一步适用于在较小内存预算下微调大型语言模型。在可比的内存开销下，我们的方法比内存高效的零阶学习算法训练得更快，并取得了更高的性能。
cs.LG / 88 / 2610.23181
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
用于长时程智能体强化学习的树感知后验蒸馏
large language model
大语言模型相关
Abstract
Reinforcement learning with verifiable rewards has become a scalable post-training paradigm for large language models. However, training agents over long-horizon interactions presents two challenges: using only terminal rewards makes turn-level credit assignment sparse and high-variance, while sequential mini-batch updates on a fixed rollout batch introduce policy staleness as the current policy drifts from the behavior policy. Existing approaches obtain denser credit via learned critics, process reward models, or on-policy distillation, but they rely on a value network, step-level annotations, or a stronger external teacher that increase training complexity. Meanwhile, importance sampling and clipping mitigate policy drift, but they do not specify a finite target that stops sequential updates from overshooting. In this work, we introduce TreeOPSD, a critic-free, tree-aware posterior self-distillation framework for long-horizon agentic reinforcement learning. Treating each interaction turn as a tree node, TreeOPSD branches alternative continuations from shared prefixes and recursively backs up terminal rewards into implicit turn-level credit. It then maps each credit to a closed-form, bounded teacher distribution parameterized by the likelihood ratio relative to the frozen behavior policy. Minimizing the divergence between this teacher and the student policy recovers the policy-gradient direction as the tangent at the behavior policy while bounding updates toward a finite target across sequential mini-batches. Empirically, TreeOPSD improves agentic capabilities across several competitive benchmarks and model scales, especially for long-horizon tasks, e.g., raising the average accuracy of Qwen3-30B-A3B-Thinking-2507 by 4.97 points over GRPO on the Deep Research setting.
Chinese Translation
具有可验证奖励的强化学习已成为大语言模型的一种可扩展后训练范式。然而，在长时程交互中训练智能体面临两个挑战：仅使用终端奖励会使轮级信用分配变得稀疏且高方差，而在固定 rollout 批次上进行顺序小批量更新会引入策略陈旧性，因为当前策略会偏离行为策略。现有方法通过学习到的评论家、过程奖励模型或同策略蒸馏来获得更稠密的信用，但它们依赖于价值网络、步级标注或更强的外部教师，从而增加了训练复杂度。与此同时，重要性采样和裁剪可以缓解策略漂移，但它们并未指定一个有限目标来阻止顺序更新发生过冲。在本工作中，我们提出 TreeOPSD，一个用于长时程智能体强化学习的无评论家、树感知后验自蒸馏框架。TreeOPSD 将每个交互轮次视为一个树节点，从共享前缀分支出替代续写，并将终端奖励递归回传为隐式的轮级信用。然后，它将每个信用映射为一个闭式、有界的教师分布，该分布由相对于冻结行为策略的似然比参数化。最小化该教师分布与学生策略之间的散度，会在行为策略处作为切线恢复策略梯度方向，同时在顺序小批量上将更新约束到一个有限目标。在实证上，TreeOPSD 在多个具有竞争力的基准和模型规模上提升了智能体能力，尤其是对于长时程任务，例如，在 Deep Research 设置下，将 Qwen3-30B-A3B-Thinking-2507 的平均准确率较 GRPO 提高了 4.97 个百分点。
cs.LG / 89 / 2610.23223
Process Reward Modeling with Reliability- and Conservatism-Aware Bayesian Belief
基于可靠性与保守性感知贝叶斯信念的过程奖励建模
large language model
大语言模型相关
Abstract
Process Reward Models (PRMs) play a central role in test-time scaling for large language model (LLM) reasoning. However, PRMs may provide inaccurate process rewards for test-time scaling, which we identify as arising from two sources of uncertainty: (1) \textit{PRM uncertainty} that arises from insufficient supervision data and imperfect PRM training; and (2) \textit{policy uncertainty} that results from the mismatch between the policy to collect supervision data and the policy for test-time scaling, leading to systematic reward overestimation. Previous studies fail to jointly consider both uncertainties. To bridge this gap, we propose BayesianPRM, which treats the process reward as a random variable within a reward hypothesis space and maintains a Bayesian belief over it. Since exact Bayesian inference over the continuous hypothesis space is generally intractable, we discretize the reward hypothesis space with a reward ensemble and introduce two complementary belief updates, each targeting one source of uncertainty: (1) to address PRM uncertainty, we introduce \textit{reliability-aware belief updating} that measures how well each ensemble member explains the supervision data and updates the belief with an amortized belief network optimized by the evidence lower bound (ELBO); and (2) to tackle policy uncertainty, we propose \textit{conservatism-aware belief calibration} which further reweights the belief toward lower rewards to suppress overestimation. These two updates yield a reliability- and conservatism-aware Bayesian belief, whose expectation determines the process reward. Experiments on multimodal reasoning benchmarks indicate that BayesianPRM consistently yields more accurate process reward estimates while enhancing test-time scaling performance.
Chinese Translation
过程奖励模型（PRMs）在大语言模型（LLM）推理的测试时扩展中发挥着核心作用。然而，PRMs 可能为测试时扩展提供不准确的过程奖励，我们指出这源于两类不确定性：（1）\textit{PRM 不确定性}，其源于监督数据不足和 PRM 训练不完美；（2）\textit{策略不确定性}，其源于收集监督数据的策略与用于测试时扩展的策略之间的不匹配，从而导致系统性的奖励高估。先前的研究未能同时考虑这两种不确定性。为弥补这一空白，我们提出 BayesianPRM，它将过程奖励视为奖励假设空间内的随机变量，并在其上维护一个贝叶斯信念。由于在连续假设空间上进行精确贝叶斯推断通常是难以处理的，我们使用奖励集成对奖励假设空间进行离散化，并引入两种互补的信念更新，每种更新分别针对一种不确定性来源：（1）为处理 PRM 不确定性，我们引入\textit{可靠性感知信念更新}，它衡量每个集成成员对监督数据的解释程度，并使用由证据下界（ELBO）优化的摊销信念网络来更新信念；（2）为应对策略不确定性，我们提出\textit{保守性感知信念校准}，它进一步将信念重新加权到较低奖励，以抑制高估。这两种更新产生一个可靠性与保守性感知的贝叶斯信念，其期望决定过程奖励。在多模态推理基准上的实验表明，BayesianPRM 在持续产生更准确的过程奖励估计的同时，提升了测试时扩展性能。
cs.LG / 90 / 2610.23265
ADT: All-Atom Diffusion Tokenizer for 3D Molecular Structures
ADT：面向三维分子结构的全原子扩散分词器
diffusion
扩散模型相关
Abstract
Discrete structural tokenization provides a promising route to autoregressive 3D molecular generation. Yet, existing tokenizers face a trade-off between compact representation and all-atom reconstruction: atom-level tokens preserve atomic details but produce long sequences, while compact residue-level protein tokens typically omit side-chain geometry. To overcome this limitation, we propose All-atom Diffusion Tokenizer (ADT), which decouples token granularity from reconstruction resolution. ADT uses one structural token per atom for small molecules and one token per protein residue that jointly represents backbone and side-chain geometry. To mitigate geometric distortion introduced by quantization, we leverage the generative capacity of diffusion models to recover fine-grained atomic geometry beyond deterministic decoding from discrete tokens. Extensive experiments across small molecules and proteins demonstrate that ADT achieves strong reconstruction from compact discrete representations while enabling high-quality autoregressive generation of both small molecules and all-atom protein structures. These results establish compact structural tokens as an effective discrete interface for all-atom molecular reconstruction and generation.
Chinese Translation
离散结构分词为自回归三维分子生成提供了一条富有前景的路径。然而，现有分词器面临着紧凑表示与全原子重建之间的权衡：原子级 token 保留了原子细节，但会产生长序列；而紧凑的残基级蛋白质 token 通常忽略侧链几何结构。为克服这一局限，我们提出了全原子扩散分词器（All-atom Diffusion Tokenizer，ADT），它将 token 粒度与重建分辨率解耦。ADT 对小分子使用每个原子一个结构 token，对蛋白质使用每个残基一个 token，该 token 联合表示主链与侧链几何结构。为缓解量化引入的几何失真，我们利用扩散模型的生成能力，在离散 token 的确定性解码之外恢复细粒度的原子几何结构。在小分子和蛋白质上的大量实验表明，ADT 能够从紧凑的离散表示实现强重建，同时支持小分子和全原子蛋白质结构的高质量自回归生成。这些结果确立了紧凑结构 token 作为全原子分子重建与生成的有效离散接口。
cs.LG / 91 / 2610.23289
Data-Efficient Classifier Pretraining Through Diffusion Models
基于扩散模型的数据高效分类器预训练
diffusion
扩散模型相关
Abstract
Machine learning with limited data is susceptible to overfitting, especially with modern deep neural networks. Prior to the introduction of AlexNet, reconstruction-based pretraining with auto-encoder and denoising auto-encoder were found to be beneficial, partially due to the limited size of the training sets at the time. Motivated by these traditional approaches, we propose DiffPT, a simple diffusion modeling approach to pretrain a given classifier backbone for limited domain-specific training data. We experiment on time-series data, standard natural-image benchmarks, and medical imaging datasets, and contrast the diffusion approach to a multitude of modern and traditional un/self-supervised methods. While contrastive pretraining performs better on CIFAR-10 and Tiny ImageNet in low-data regimes, DiffPT is particularly effective on medical imaging and time-series tasks, where it outperforms self-supervised baselines across a range of low-data regimes. Our results suggest that DiffPT is a simple and effective approach for data-efficient classifier pretraining on different modalities.
Chinese Translation
在数据有限的情况下，机器学习容易过拟合，尤其是使用现代深度神经网络时。在 AlexNet 被提出之前，使用自编码器和去噪自编码器进行基于重构的预训练被发现是有益的，部分原因是当时训练集规模有限。受这些传统方法的启发，我们提出 DiffPT，一种简单的扩散建模方法，用于在有限的领域特定训练数据上预训练给定的分类器骨干网络。我们在时间序列数据、标准自然图像基准和医学影像数据集上进行实验，并将该扩散方法与大量现代和传统的无监督/自监督方法进行对比。尽管对比预训练在低数据情形下的 CIFAR-10 和 Tiny ImageNet 上表现更好，但 DiffPT 在医学影像和时间序列任务上尤其有效，在这些任务中，它在一系列低数据情形下优于自监督基线。我们的结果表明，DiffPT 是一种简单而有效的方法，可用于在不同模态上进行数据高效的分类器预训练。
cs.LG / 92 / 2610.23404
Information Estimation with Discrete Diffusion
基于离散扩散的信息估计
diffusion
扩散模型相关
Abstract
Information-theoretic measures, such as Mutual Information (MI), play a crucial role in understanding non-linear relationships between random variables and are widely used across scientific disciplines. Yet, their use on real-world discrete data remains challenging. Existing methods typically rely on embedding discrete data into a continuous space and apply neural estimators originally designed for continuous distributions. This process requires careful engineering for both the embedding model and estimator architecture, but suffers from issues related to high data dimensionality. In this work, we introduce InfoSEDD, a discrete diffusion-based approach that bridges information-theoretic estimation and generative modeling such that they can be used to compute Kullback-Leibler divergences. Backed by Continuous Time Markov Chains theory principles, the design of InfoSEDD is lightweight and scalable and allows seamless integration with pretrained models. We showcase the versatility of our approach through applications on motif discovery in genetic promoter data, semantic-aware model selection in text summarization, and entropy estimation in Ising models. Finally, we construct consistency tests on real-world textual and genomics data. Our experiments demonstrate that InfoSEDD outperforms alternatives that rely on the "embedding trick". Our results position InfoSEDD as a robust and scalable tool for information-theoretic analysis of discrete data.
Chinese Translation
互信息（MI）等信息论度量在理解随机变量之间的非线性关系方面发挥着至关重要的作用，并被广泛应用于各科学学科之中。然而，将这些度量用于真实世界的离散数据仍然具有挑战性。现有方法通常依赖于将离散数据嵌入到连续空间中，并应用原本为连续分布设计的神经估计器。这一过程需要对嵌入模型和估计器架构进行细致的工程设计，但仍会受到与高数据维度相关的问题的困扰。在本工作中，我们提出了 InfoSEDD，一种基于离散扩散的方法，它架起了信息论估计与生成建模之间的桥梁，使得二者可用于计算 Kullback-Leibler 散度。在连续时间马尔可夫链理论原理的支撑下，InfoSEDD 的设计轻量且可扩展，并允许与预训练模型无缝集成。我们通过在遗传启动子数据中的模体发现、文本摘要中的语义感知模型选择以及 Ising 模型中的熵估计等应用，展示了该方法的多功能性。最后，我们在真实世界的文本与基因组学数据上构建了一致性检验。我们的实验表明，InfoSEDD 优于依赖“嵌入技巧”的替代方法。我们的结果使 InfoSEDD 成为面向离散数据信息论分析的稳健且可扩展的工具。
cs.LG / 93 / 2610.23638
Deep LLM layers expose passage-level context in EEG during silent reading
深度LLM层在默读期间的EEG中揭示篇章级上下文
large language model
大语言模型相关
Abstract
In a previous report we showed that a decoder can recover open-vocabulary word-level information from 19-channel EEG recorded during silent reading, by aligning short EEG windows with hidden-state embeddings of the presented word taken from a large language model (LLM). That analysis contrasted two of the model's 33 layers: the context-free input embedding and one mid-depth layer. Here we ask how EEG-decodable structure is distributed over the full depth of the language model. We trained one decoder per layer of Llama-3.1-8B, with and without a causal transformer over the sequence of EEG windows. We distinguish word-level decodability from a context-tracking component. Word-level decodability peaks in early layers and declines steadily towards the last layer. In contrast, the context-tracking component of the retrieval gain rises with depth, both with and without a causal transformer. Unexpectedly, a decoder that sees a single 0.5-s window at a time tracks the text passage too. Two new controls dissect this component. A target-geometry control, a synthetic predictor made of the true target plus noise calibrated to each decoder's word-level accuracy, shows that context tracking is not a by-product of word decoding and target geometry. A control that permutes predictions across runs from either the same or from different sessions shows that most of the observed context-tracking is not caused by an EEG recording session signature. Instead, context tracking mostly follows the text. A single EEG window therefore carries passage-level information that a contextual target can absorb. The results locate word identity in the early layers of the language model, discourse state in the deeper layers, and make the choice of target layer a design dial for EEG language decoders.
Chinese Translation
在先前的一份报告中，我们表明，通过将短EEG窗口与取自大型语言模型（LLM）的所呈现词的隐藏状态嵌入对齐，解码器可以从默读期间记录的19通道EEG中恢复开放词汇的词级信息。该分析对比了该模型33层中的两层：无上下文的输入嵌入和一个中等深度层。在这里，我们探讨EEG可解码结构如何在语言模型的整个深度上分布。我们为Llama-3.1-8B的每一层训练了一个解码器，分别在有和没有对EEG窗口序列施加因果Transformer的情况下进行。我们将词级可解码性与一个上下文跟踪成分区分开来。词级可解码性在早期层达到峰值，并向最后一层稳步下降。相比之下，检索增益中的上下文跟踪成分随深度增加而上升，无论有或没有因果Transformer均如此。出乎意料的是，一次只看到一个0.5秒窗口的解码器也能跟踪文本篇章。两个新的对照分析剖析了这一成分。一个目标几何对照——由真实目标加上根据每个解码器的词级准确率校准的噪声构成的合成预测器——表明上下文跟踪不是词解码和目标几何的副产物。一个将来自相同或不同会话的各次运行的预测进行置换的对照表明，观察到的大部分上下文跟踪不是由EEG记录会话特征造成的。相反，上下文跟踪主要跟随文本。因此，单个EEG窗口携带着上下文目标能够吸收的篇章级信息。这些结果将词身份定位在语言模型的早期层，将话语状态定位在更深的层，并使目标层的选择成为EEG语言解码器的一个设计调节旋钮。
cs.LG / 94 / 2610.23703
Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2
Blitz：加速 AlphaFold3 与 Boltz-2 的扩散模块
diffusion
扩散模型相关
Abstract
Diffusion-based protein structure prediction models produce all-atom structures in about 200 denoising steps, and that cost has driven a number of few-step samplers. We present Blitz, a few-step distillation framework that uses a geometric noise schedule to distribute steps more evenly across noise scales, then learns those updates through multistep consistency training. Physical steering and target-free candidate selection help reduce the remaining geometric defects at inference. We apply Blitz to AlphaFold 3 and Boltz-2, reducing denoising from 200 steps to eight. Because published few-step results use different benchmarks and scoring procedures, we compare both distilled models with their teachers and other few-step methods under a common evaluation protocol covering structural accuracy, interface quality, ligand placement, and stereochemical validity. Blitz retains teacher-level structural accuracy, with AF3-Blitz outperforming all evaluated few-step baselines and exceeding its teacher on several accuracy metrics.
Chinese Translation
基于扩散的蛋白质结构预测模型通过约 200 个去噪步骤生成全原子结构，这一代价催生了若干少步采样器。我们提出 Blitz，一种少步蒸馏框架，它使用几何噪声调度在噪声尺度上更均匀地分配步骤，然后通过多步一致性训练学习这些更新。物理引导与无目标候选选择有助于减少推理时残留的几何缺陷。我们将 Blitz 应用于 AlphaFold 3 和 Boltz-2，将去噪从 200 步减少到八步。由于已发表的少步结果使用不同的基准和评分流程，我们在一个涵盖结构准确性、界面质量、配体放置和立体化学有效性的统一评估协议下，将两个蒸馏模型与其教师模型以及其他少步方法进行比较。Blitz 保持了教师级别的结构准确性，其中 AF3-Blitz 优于所有被评估的少步基线，并在若干准确性指标上超过其教师模型。
cs.LG / 95 / 2610.23766
Logic-Guided Graph Denoising Diffusion Models
逻辑引导的图去噪扩散模型
diffusion
扩散模型相关
Abstract
Graph diffusion models are competitive approaches for generating complex graph-structured data. However, when such distributions have strict structural constraints, diffusion models may fail to capture them exactly, resulting in the generation of invalid graphs. In this paper, we propose a method for grounding graph diffusion models in domain-specific knowledge via First Order Logic (FOL), leveraging its expressive and interpretable formalism to impose structural constraints on the generative process. We incorporate domain knowledge using classifier guidance, a technique that conditions the diffusion model's distribution on external variables to exert control over the sampled data. During the generative process, classifier guidance steers the denoising trajectory to optimize an objective function, which we define as a fuzzy and differentiable relaxation of the logical rules. Our methodology requires only the specification of constraints in FOL, whose satisfaction is encouraged at inference time to guide a pre-trained generative model. We evaluate our approach by comparing the unconditional generation of knowledge graphs with our logic-guided framework, which incorporates domain ontologies as structural constraints during inference. This approach significantly increases the proportion of generated graphs that satisfy the ontological constraints.
Chinese Translation
图扩散模型是生成复杂图结构数据的具有竞争力的一类方法。然而，当此类分布具有严格的结构约束时，扩散模型可能无法精确地捕捉这些约束，从而导致生成无效的图。在本文中，我们提出了一种通过一阶逻辑（FOL）将图扩散模型建立在领域特定知识之上的方法，利用其一阶逻辑所具有的表达性和可解释性的形式化体系，对生成过程施加结构约束。我们使用分类器引导来融入领域知识，该技术以外部变量为条件来调节扩散模型的分布，从而对采样的数据施加控制。在生成过程中，分类器引导将去噪轨迹导向以优化一个目标函数，我们将该目标函数定义为对逻辑规则的模糊且可微的松弛。我们的方法仅需要以一阶逻辑指定约束，在推理时鼓励满足这些约束，以引导一个预训练的生成模型。我们通过将知识图的无条件生成与我们的逻辑引导框架进行比较来评估我们的方法，该框架在推理过程中将领域本体作为结构约束纳入其中。该方法显著提高了所生成图满足本体约束的比例。
cs.MA / 96 / 2610.22674
Gossip Agents: Reputation through Opinion Dynamics over Evolving Networks
八卦智能体：在演化网络上通过观点动力学建立声誉
diffusion
扩散模型相关
Abstract
Sustained cooperation among self-interested learning agents is essential in many shared environments. Reputation systems are among the best-studied mechanisms for promoting such cooperation. Existing multi-agent reinforcement learning approaches, however, typically assume that reputation is either globally available to all agents or derived solely from an agent's own observations, leaving largely unexplored how reputational information propagates between agents. We study gossip, the exchange of information about third parties, as a mechanism for reputation diffusion over dynamic communication networks. Agents form private opinions from limited observations and exchange them with their neighbors following DeGroot opinion dynamics, while the communication network changes as agents interact and move. A tunable broadcast radius controls the reach of information, spanning private reputation, local gossip, and globally shared reputation. We study this framework in two \(n\)-player social dilemmas with distinct network dynamics: a Prisoner's Dilemma with randomly resampled interaction groups and Harvest, where communication neighborhoods emerge from agents' movements. Our results show that the reach of reputational information shapes cooperation in both environments. Moreover, in Harvest, reputation diffusion increases environmental rewards and leads to more peaceful and sustainable behavior.
Chinese Translation
自利的学习智能体之间的持续合作在许多共享环境中至关重要。声誉系统是促进此类合作研究最充分的机制之一。然而，现有的多智能体强化学习方法通常假设声誉要么全局可供所有智能体使用，要么仅由智能体自身的观察得出，因而声誉信息如何在智能体之间传播在很大程度上仍未得到探索。我们研究八卦，即关于第三方的信息交换，作为一种在动态通信网络上进行声誉扩散的机制。智能体从有限观察中形成私人观点，并按照 DeGroot 观点动力学与其邻居交换这些观点，同时通信网络随着智能体的交互和移动而变化。一个可调的广播半径控制信息的传播范围，涵盖私人声誉、局部八卦和全局共享声誉。我们在两个具有不同网络动力学的 \(n\) 人社会困境中研究该框架：一个是随机重采样交互组的囚徒困境，另一个是 Harvest，其中通信邻域由智能体的移动产生。我们的结果表明，声誉信息的传播范围在两种环境中都塑造合作。此外，在 Harvest 中，声誉扩散提高了环境奖励，并导致更和平、更可持续的行为。
cs.MA / 97 / 2610.22701
Emergent Discrimination and Polarization in Multi-Agent Systems
多智能体系统中的涌现性歧视与极化
large language model
大语言模型相关
Abstract
We study theoretically how discrimination and polarization can emerge in a population of learning agents. Each agent uses an adaptive linear decision layer to map a language-model embedding of an issue to a binary opinion and maintains an evolving estimate of how much it trusts every other agent. Through pairwise interactions, agents update both their decision weights and trust estimates using an online Bayesian rule derived from teacher-student learning over a noisy channel. In-context-learning experiments with large language models qualitatively reproduce the predicted opinion and trust updates. We then assign each agent one of two arbitrary group labels, A or B, and model group bias by allowing a fraction $f_b$ of agents to shift its perceived agreement with each message by a strength $b$, a bias steering prejudice, depending on whether the sender and receiver share a label. We use order parameters from statistical mechanics to connect these local dynamics to collective states and transitions. Starting from uniformly random initial conditions, simulations reveal a rich phase diagram in the $(b,f_b)$ plane, comprising four collective states: spin-glass frustrated out-group favoritism, nondiscriminatory polarization, discrimination with opinion alignment, and discrimination without opinion alignment. Which state emerges depends on both the strength and prevalence of the bias. These results show how simple local learning dynamics can generate qualitatively distinct forms of polarization and discrimination, underscoring the need for population-level safety analysis of multi-agent systems. They also suggest measurable order parameters for characterizing analogous collective states in human societies.
Chinese Translation
我们从理论上研究歧视与极化如何能够在一群学习型智能体中涌现。每个智能体使用一个自适应线性决策层，将某个议题的语言模型嵌入映射为二元观点，并维护一个不断演化的、关于其多大程度信任其他每个智能体的估计。通过成对交互，智能体使用一种由噪声信道上的师生学习推导得到的在线贝叶斯规则，同时更新其决策权重和信任估计。使用大语言模型的上下文学习实验在定性上复现了所预测的观点更新和信任更新。随后，我们为每个智能体分配两个任意群体标签 A 或 B 中的一个，并通过允许比例为 $f_b$ 的智能体根据发送者和接收者是否共享标签，以强度 $b$ 将其对每条消息的感知一致程度进行偏移来建模群体偏见，这是一种引导偏见的偏差。我们使用统计力学中的序参量将这些局部动力学与集体状态和相变联系起来。从均匀随机初始条件出发，模拟揭示了在 $(b,f_b)$ 平面中丰富的相图，包含四种集体状态：自旋玻璃阻挫的外群体偏爱、非歧视性极化、伴随观点对齐的歧视，以及不伴随观点对齐的歧视。哪种状态出现取决于偏见的强度和普遍程度。这些结果表明，简单的局部学习动力学如何能够产生定性上不同的极化和歧视形式，凸显了对多智能体系统进行群体层面安全分析的必要性。它们还提示了可用于表征人类社会中类似集体状态的可测量序参量。
cs.AI / 98 / 2610.23396
AWAM: Active Trajectory-Conditioned View Imagination for Robust Robot Manipulation
AWAM：面向鲁棒机器人操作的主动轨迹条件化视角想象
diffusion
扩散模型相关
Abstract
Rapid progress in embodied intelligence has enabled increasingly capable robots to follow language instructions and manipulate diverse objects. However, strong results in structured environments do not automatically transfer to open-world conditions, where clutter, occlusion, and changing viewpoints can obscure the spatial relationships required for action. To address this perceptual gap, a virtual active-perception paradigm is introduced. AWAM, an Active World-Action Model actively plans where to observe and imagines what will be seen along the planned trajectory before acting. Task-related virtual camera motion is predicted and used as a structured query into a geometry-conditioned world model, producing a reusable spatial prior for manipulation, which is later compressed into a fixed memory and fused with live observations through cross-attention for diffusion-based action prediction. AWAM achieves competitive success rates with 75.9% on LIBERO-Occ, 62.4% on LIBERO-Plus, 76.5% on RLBench-OG Occlusion1, and 94.4% in real-robot trials. These results support improved robustness to the evaluated occlusions and distribution shifts, addressing part of the gap between structured manipulation and less controlled environments.
Chinese Translation
具身智能的快速进展使机器人能力不断增强，能够遵循语言指令并操作多样化的物体。然而，结构化环境中的优异结果并不会自动迁移到开放世界条件，在开放世界中，杂乱、遮挡和不断变化的视角可能会遮蔽动作所需的空间关系。为弥补这一感知差距，本文引入了一种虚拟主动感知范式。AWAM 是一种主动世界-动作模型，它在执行动作之前主动规划在何处进行观察，并想象沿规划轨迹将会看到的内容。任务相关的虚拟相机运动被预测出来，并作为结构化查询输入到几何条件化的世界模型中，产生可复用于操作的空间先验，该先验随后被压缩为固定记忆，并通过交叉注意力与实时观测融合，用于基于扩散的动作预测。AWAM 取得了具有竞争力的成功率：在 LIBERO-Occ 上为 75.9%，在 LIBERO-Plus 上为 62.4%，在 RLBench-OG Occlusion1 上为 76.5%，在真实机器人试验中为 94.4%。这些结果支持了对所评估的遮挡和分布偏移的鲁棒性提升，缩小了结构化操作与较少受控环境之间差距的一部分。
cs.SE / 99 / 2610.22520
Understanding LLM Hallucinations Caused by Soft Errors
理解由软错误引起的大语言模型幻觉
large language model
大语言模型相关
Abstract
While hallucinations in Large Language Models (LLMs) have been extensively studied from a software perspective, limited attention has been paid to hardware-level factors. In this paper, we performed a case study to understand LLM hallucination caused by soft errors on hardware. We introduce LLMEI to inject soft errors into LLMs during inference, based on what we evaluate the impact of soft error across diverse LLM architectures and tasks, including mathematical reasoning and code generation. We summarize the taxonomy of different failure modes based on LLM hallucination behaviors and analyze their distributions under different error rates. Our find- ings underscore the hallucination-specific diagnostics within the reliability of LLM implementation on hardware.
Chinese Translation
尽管大语言模型（LLM）中的幻觉已从软件角度得到广泛研究，但硬件层面的因素受到的关注有限。本文开展了一项案例研究，以理解由硬件软错误导致的LLM幻觉。我们提出LLMEI，用于在推理过程中向LLM注入软错误，并基于此评估软错误在不同LLM架构和任务（包括数学推理和代码生成）中的影响。我们基于LLM幻觉行为总结了不同故障模式的分类体系，并分析了它们在不同错误率下的分布。我们的发现强调了在LLM硬件实现可靠性中针对幻觉的专门诊断。
cs.SE / 100 / 2610.22528
ForTress: Bringing Assurance and Determinism for LLM-Driven Fortran-to-Python Translation
ForTress：为 LLM 驱动的 Fortran 到 Python 翻译带来保证与确定性
large language model
大语言模型相关
Abstract
Modernizing mature scientific software from Fortran to Python can improve accessibility and integration with modern computing tools, but manual rewriting is expensive. Large language models (LLMs) make translation more practical, yet generated code can pass regression tests while violating source-language semantics, and contextual corrections may not persist across subsequent translations. We present ForTress, which solidifies recurring LLM translation decisions into guarded executable rules for deterministic reuse and provides semantic assurance for a non-trivial subset of these rules. Its LLM-assisted auditing process expresses universally quantified preservation obligations over mechanized Fortran and Python semantics in Lean. Human experts review semantic fidelity and correspondence with production rules, and Lean's kernel checks proofs and refutations under explicit assumptions. These audit records expose the conditions either supporting correctness or identifying failure cases. We evaluate ForTress on 17 benchmarks from 15 scientific software spanning diverse domains. The translated programs achieve bitwise-identical with native Fortran on all evaluated tests under the recorded execution and initialization configurations. Against these software, ForTress accumulates 193 executable rules and 188 audit records, of which 121 have checked soundness proofs, including 31 established after counterexample-guided refinement. Cross-software reuse exceeds 90% for 14 of the 15 software. Even with additional assurance and determinism, the LLM usage cost is 44.2% lower than direct LLM translation.
Chinese Translation
将成熟的科学软件从 Fortran 现代化到 Python 可以提高可访问性以及与现代化计算工具的集成，但人工重写成本高昂。大语言模型（LLM）使翻译更加实用，但生成的代码可能通过回归测试，却违反了源语言的语义，并且上下文修正可能无法在后续翻译中持续保留。我们提出 ForTress，它将反复出现的 LLM 翻译决策固化为带守卫的可执行规则，以实现确定性复用，并为这些规则中一个不可忽视的子集提供语义保证。其 LLM 辅助审计过程在 Lean 中针对机械化的 Fortran 和 Python 语义表达全称量化的保持义务。人类专家审查语义保真度以及与生产规则的对应关系，Lean 的内核在明确假设下检查证明和反驳。这些审计记录揭示了支持正确性或识别失败情形的条件。我们在来自 15 个科学软件的 17 个基准上评估 ForTress，这些软件涵盖多个不同领域。在记录的执行与初始化配置下，翻译后的程序在所有评估测试中与原生 Fortran 达到逐位一致。针对这些软件，ForTress 累积了 193 条可执行规则和 188 条审计记录，其中 121 条具有已检查的可靠性证明，包括 31 条是在反例引导的精化之后建立的。对于 15 个软件中的 14 个，跨软件复用率超过 90%。即使增加了保证与确定性，LLM 使用成本仍比直接 LLM 翻译低 44.2%。
cs.SE / 101 / 2610.22881
No One-Size-Fits-All: Adaptive Code Editing with Feature-Based Strategy Selection
并非一刀切：基于特征策略选择的自适应代码编辑
large language model
大语言模型相关
Abstract
Large language models have shown strong potential in automating code editing (CE) tasks, yet most existing systems rely on a fixed CE strategy (i.e., one-step, search-replace, and plan-apply) across diverse scenarios. In practice, CE tasks vary widely in code context, edit locality, and semantic complexity, causing no single strategy to perform consistently well. Moreover, systematic studies on the relationship between strategies and CE tasks remain scarce, limiting further optimization of LLM-based code editing strategies. To address this gap, we construct CEBench, comprising 500 real-world CE tasks curated from SWE-bench. Then we compare three representative CE strategies (one-step, search-replace, plan-apply). Our study reveals two key findings: (1) each strategy exhibits distinct advantages for specific CE tasks, with no universal optimality; (2) strategy effectiveness is significantly influenced by four core task features: total lines of code, modification type, modification scope, and task type. Motivated by these findings, we propose Ace, an adaptive code editing framework that dynamically selects optimal CE strategies based on task features. Unlike classifier-based or instruction-prompting approaches, Ace employs a fine-tuned LLM that learns feature-strategy mappings from empirical data to recommend strategies with interpretable rationales. Extensive experiments on our CEBench-verified show that Ace outperforms state-of-the-art baselines: it boosts Pass@1 by 21.7% while reducing token consumption by 27.3% compared to the best baseline (AutoCodeRover). Our strategy selector achieves 72.6% selection accuracy with 69.9% rationale accuracy (Cohen's Kappa = 0.807), providing transparent and trustworthy decision-making for developers.
Chinese Translation
大语言模型在自动化代码编辑（CE）任务中已展现出强大潜力，然而大多数现有系统在多样场景中依赖固定的CE策略（即一步式、搜索-替换和计划-应用）。在实践中，CE任务在代码上下文、编辑局部性和语义复杂度方面差异很大，导致没有单一策略能够始终表现良好。此外，关于策略与CE任务之间关系的系统性研究仍然稀缺，这限制了对基于LLM的代码编辑策略的进一步优化。为填补这一空白，我们构建了CEBench，其中包含从SWE-bench整理出的500个真实世界CE任务。然后我们比较了三种代表性的CE策略（一步式、搜索-替换、计划-应用）。我们的研究揭示了两个关键发现：（1）每种策略对特定CE任务表现出独特优势，不存在普遍最优性；（2）策略有效性受到四个核心任务特征的显著影响：代码总行数、修改类型、修改范围和任务类型。受这些发现启发，我们提出了Ace，一个基于任务特征动态选择最优CE策略的自适应代码编辑框架。与基于分类器或指令提示的方法不同，Ace采用一个微调后的LLM，从经验数据中学习特征-策略映射，以可解释的理由推荐策略。在我们CEBench-verified上的大量实验表明，Ace优于最先进的基线：与最佳基线（AutoCodeRover）相比，它将Pass@1提高了21.7%，同时将token消耗降低了27.3%。我们的策略选择器实现了72.6%的选择准确率和69.9%的理由准确率（Cohen's Kappa = 0.807），为开发者提供透明且可信的决策。
cs.SE / 102 / 2610.23035
VulnSAGE: Adjudicating SAST Reports by Connecting LLM Reasoning to Program Structure
VulnSAGE：通过将 LLM 推理连接到程序结构来裁定 SAST 报告
large language model
大语言模型相关
Abstract
Static Application Security Testing (SAST) reports flag operations whose safety depends on project-specific context, and maintainers must decide which reports describe real defects. Large language models (LLMs) can interpret this context, but adjudicating a report with them requires the decisive code, which often lies far from the reported line, and a decision basis that program analysis can check. We present VulnSAGE, an adjudication system for SAST reports that connects LLM reasoning to program structure in both directions. Its report-centered slicer recovers the dependencies of the reported operation together with the program structure that explains them, such as the caller object on which a helper acts. CLAIM, an intermediate representation for report judgments, records each model interpretation as a semantic fact that cites the program relations it assumes. Program analysis checks these premises, and the verdict follows from the facts whose premises hold. Both mechanisms share a program graph, which lets the model cite the code it reads and lets the checker query those citations. Statically supported true positives then guide proof-of-concept generation and runtime reproduction. Across 187 labeled reports from four corpora covering five vulnerability families, VulnSAGE rejects 89 of 97 false alarms and retains 85 of 90 real defects. Under a common PoC procedure, 25 of its positive verdicts reproduce at runtime, compared with three for the LLM-based adjudicator LLM4SA. Applied to 45,440 alerts raised by five SAST tools on the 200 most-starred C/C++ projects on GitHub, it reproduces 104 defects at runtime. Maintainers have confirmed 24 of the 26 defects they have responded to, including ten previously unknown defects, nine of which are already fixed.
Chinese Translation
静态应用安全测试（SAST）报告会标记其安全性取决于项目特定上下文的操作，而维护者必须判断哪些报告描述的是真实缺陷。大型语言模型（LLM）能够解释这种上下文，但用它们来裁定一份报告需要决定性的代码——这些代码往往远离被报告的行——以及一个可由程序分析检查的决策依据。我们提出 VulnSAGE，一个用于 SAST 报告的裁定系统，它在两个方向上把 LLM 推理与程序结构连接起来。其以报告为中心的切片器会恢复被报告操作的依赖关系，以及解释这些依赖关系的程序结构，例如辅助函数所作用的调用者对象。CLAIM 是一种用于报告判定的中间表示，它把每个模型解释记录为引用其所假设程序关系的语义事实。程序分析检查这些前提，而判定则来自那些前提成立的事实。两种机制共享一个程序图，这使模型能够引用其阅读的代码，并使检查器能够查询这些引用。随后，得到静态支持的真阳性会指导概念验证生成和运行时复现。在来自四个语料库、覆盖五类漏洞家族的 187 份带标签报告中，VulnSAGE 拒绝了 97 个误报中的 89 个，并保留了 90 个真实缺陷中的 85 个。在一种通用的 PoC 流程下，其正面判定中有 25 个可在运行时复现，相比之下，基于 LLM 的裁定器 LLM4SA 为三个。应用于五个 SAST 工具在 GitHub 上 200 个星标最多的 C/C++ 项目上提出的 45,440 条告警时，它在运行时复现了 104 个缺陷。维护者已确认了其所回应的 26 个缺陷中的 24 个，包括十个此前未知的缺陷，其中九个已经修复。
cs.SE / 103 / 2610.23054
Efficient Program Reduction for Library Fuzzing Harnesses
面向库模糊测试驱动的高效程序缩减
large language model
大语言模型相关
Abstract
Library fuzzing harnesses translate generated inputs into API calls. When a complex harness triggers a crash, developers must determine whether the cause is a library defect or invalid harness usage. A smaller reproducer can simplify this investigation, but obtaining one requires repeatedly building and executing modified harnesses and analyzing their crashes, making manual or even automatic minimization time-consuming. Moreover, data-dependent statements prevent the reducer to remove unnecessary code and the heavy stack trace symbolization in libraries also slows down the candidate execution. To address these challenges, we present HarnessMinimizer, a system for efficient minimization of C/C++ library fuzzing harnesses. Its optimized pipeline reuses header analysis and prepared libraries and runtimes to reduce the work required to build and execute each simplified harness during the reduction process. It records and replays input values so that unnecessary input operations can be removed without changing the values supplied to surviving calls. It also uses a fast method to check whether each simplified harness preserves the original crash, avoiding repeated stack trace symbolization. We evaluate HarnessMinimizer with four reduction engines on 200 crash cases from 21 open source libraries. Compared with the unoptimized pipeline, HarnessMinimizer achieves speedups of 3.46-4.51 times in average total wall time. Average harness size decreases from 1,352 original tokens to 143-263 remaining tokens after minimization. In crash triage using a large language model, accuracy in identifying harness bugs increases from 28% with the original harnesses to 44-51% with the minimized harnesses, while accuracy for library bugs remains at 100%.
Chinese Translation
库模糊测试驱动将生成的输入转换为 API 调用。当一个复杂的测试驱动触发崩溃时，开发者必须确定其原因究竟是库缺陷还是无效的测试驱动用法。更小的复现程序可以简化这一调查，但获得这样一个复现程序需要反复构建并执行修改后的测试驱动并分析其崩溃，这使得手动乃至自动缩减都耗时。此外，数据相关语句会使缩减器无法移除不必要的代码，而库中繁重的堆栈跟踪符号化也会拖慢候选执行。为应对这些挑战，我们提出 HarnessMinimizer，一个用于高效最小化 C/C++ 库模糊测试驱动的系统。其优化流水线复用头文件分析以及预先准备好的库和运行时，以减少缩减过程中构建和执行每个简化测试驱动所需的工作。它记录并重放输入值，从而可以在不改变提供给保留下来的调用的值的情况下移除不必要的输入操作。它还使用一种快速方法来检查每个简化测试驱动是否保留原始崩溃，从而避免重复的堆栈跟踪符号化。我们在来自 21 个开源库的 200 个崩溃案例上，使用四个缩减引擎评估 HarnessMinimizer。与未优化流水线相比，HarnessMinimizer 在平均总墙钟时间上实现了 3.46-4.51 倍的加速。平均测试驱动大小从原始的 1,352 个 token 降至最小化后剩余的 143-263 个 token。在使用大语言模型进行崩溃分诊时，识别测试驱动缺陷的准确率从原始测试驱动的 28% 提高到最小化测试驱动的 44-51%，而对库缺陷的准确率保持在 100%。
cs.SE / 104 / 2610.23546
Recall-Oriented Precise Binary-to-Source Function Matching
面向召回的精确二进制到源码函数匹配
large language model
大语言模型相关
Abstract
Binary-to-source function matching supports software composition analysis by identifying reused third-party code, but existing embedding-based approaches restrict subsequent refinement to initially retrieved candidates. When the correct source function is absent from this set, selecting or filtering the available candidates cannot recover the match. Inlining further complicates the retrieval by combining multiple source functions into one binary function. In this paper, we propose BScout to improve matching precision and uncover additional matches. BScout uses the positional information of existing matches to supplement candidates for unresolved binary functions, extending the search beyond the initial candidate set, and adopts composition-aware training to address the inlined composition matching. Large language models are adopted to select and review matches, while positional information also helps filter inconsistent results. Experiments show that BScout is effective in binary-to-source function matching. BScout achieves overall higher precision, recall and F1 scores than BinaryAI and SBridge with acceptable analysis costs. Particularly, on large stripped binaries, BScout outperforms the competitors with a precision increment of 32% and approximately 2.9 times the recall.
Chinese Translation
二进制到源码函数匹配通过识别被复用的第三方代码来支持软件组成分析，但现有的基于嵌入的方法将后续的细化限制在初始检索到的候选之上。当正确的源码函数不在该集合中时，对可用候选进行选择或过滤无法恢复该匹配。内联通过将多个源码函数组合成一个二进制函数，进一步使检索变得复杂。本文提出 BScout，以提高匹配精度并发现更多匹配。BScout 利用已有匹配的位置信息为未解决的二进制函数补充候选，将搜索扩展到初始候选集之外，并采用组合感知训练来处理内联组合匹配。采用大语言模型来选择和审查匹配，同时位置信息也有助于过滤不一致的结果。实验表明，BScout 在二进制到源码函数匹配中是有效的。在可接受的分析开销下，BScout 整体上取得了比 BinaryAI 和 SBridge 更高的精确率、召回率和 F1 分数。特别地，在大规模剥离二进制文件上，BScout 优于对比方法，精确率提升 32%，召回率约为其 2.9 倍。
cs.SE / 105 / 2610.23559
An Interactive GraphRAG Workspace for Reifying, Auditing, and Steering Conversational AI Memories
用于具化、审计与引导对话式 AI 记忆的交互式 GraphRAG 工作区
large language model
大语言模型相关
Abstract
Conversational Large Language Models (LLMs) are widely employed in software engineering for code generation, debugging, and architectural design. However, as developers engage in multi-session development tasks, technical knowledge becomes siloed in isolated, linear chat histories. Reusing context across sessions leads to three critical challenges: (1) flat-vector retrieval fails to traverse multi-hop relational dependencies between distant architectural decisions; (2) temporal drift in software packages and APIs results in silent context poisoning when reusing aged dialogue logs; and (3) raw conversation concatenation induces prompt bloat and attention degradation. We present CognitiveCanvas, an interactive tool that transforms conversational AI memories from ephemeral text logs into persistent, manipulable, and audited knowledge objects. \textsc{CognitiveCanvas} integrates: (1) a Hierarchical Semantic Fact-Graph (HSFG) that distills dialogues into Goal, Fact, and Code artifacts, achieving a 92.2\% token reduction while enabling multi-hop GraphRAG traversal; (2) a Continuous Temporal Drift Auditor coupling 60+ regex heuristics with live calendar-anchored LLM verification for 1-click self-correction; and (3) a Spatial-Symbolic Prompt Compiler (SSPC) that translates 2D direct manipulation gestures (Voronoi clustering, node resizing, synthesis edges) directly into structured attention directives
Chinese Translation
对话式大语言模型（LLM）被广泛应用于软件工程中的代码生成、调试和架构设计。然而，随着开发者参与多会话开发任务，技术知识被孤立在彼此隔离的线性聊天历史中，形成知识孤岛。跨会话复用上下文会导致三个关键挑战：（1）扁平向量检索无法遍历相距遥远的架构决策之间的多跳关系依赖；（2）软件包和 API 中的时序漂移会在复用陈旧的对话日志时导致静默的上下文污染；（3）原始对话拼接会引发提示膨胀和注意力退化。我们提出 CognitiveCanvas，一种交互式工具，它将对话式 AI 记忆从短暂的文本日志转化为持久的、可操作且经过审计的知识对象。\textsc{CognitiveCanvas} 集成了：（1）分层语义事实图（HSFG），它将对话提炼为目标、事实和代码工件，在实现 92.2\% 的 token 减少的同时支持多跳 GraphRAG 遍历；（2）持续时序漂移审计器，将 60 多种正则启发式与以日历为锚的实时 LLM 验证相结合，实现一键式自我校正；（3）空间-符号提示编译器（SSPC），它将二维直接操作手势（Voronoi 聚类、节点缩放、合成边）直接转换为结构化的注意力指令
cs.AI / 106 / 2610.23199
Validating LLM-generated political networks with survey-reported partisan exposure
利用调查报告的党派接触验证LLM生成的政治网络
large language model
大语言模型相关
Abstract
Large language models (LLMs) are increasingly used to simulate human behavior in social science research, including decisions, interactions, and social relationships. However, in network formation tasks, existing evaluations often focus on aggregate network properties, while networks with plausible structural properties may still fail to reproduce individuals' social and political environments. Here, we generate LLM-based political networks using personas sampled from the nationally representative ANES 2020--2022 Social Media Study and use respondents' survey-reported partisan exposure as an empirical benchmark for the political environments generated for the corresponding personas. We first characterize the structural properties and political assortativity of the generated networks, and then compare ego-level partisan exposure with the survey. We find that the structure and political assortativity of the generated networks vary across experimental configurations. More importantly, when comparing generated partisan exposure with the survey, some models generate stronger separation between Democratic and Republican egos than reported in the survey, with higher same-party and lower cross-party exposure, while others mainly underestimate Republican exposure across ego groups. Moreover, networks with similar political assortativity can still generate different partisan environments around the same personas, including among networks with plausible structural properties. Finally, generated networks show group-specific errors, under- or overestimating exposure to either Democrats or Republicans for specific social groups. These findings show that evaluating LLM-based social simulations only through aggregate outcomes can hide errors in the local environments through which those outcomes emerge, highlighting the importance of validation at both aggregate and individual levels.
Chinese Translation
大语言模型（LLMs）越来越多地被用于社会科学研究中模拟人类行为，包括决策、互动和社会关系。然而，在网络形成任务中，现有评估往往聚焦于网络的聚合属性，而具有看似合理的结构属性的网络仍可能无法复现个体的社会与政治环境。在此，我们使用从具有全国代表性的ANES 2020--2022社交媒体研究中抽样的角色画像生成基于LLM的政治网络，并将受访者调查报告的党派接触作为对应角色画像所生成政治环境的经验基准。我们首先刻画生成网络的结构属性和政治同配性，然后在ego层面将党派接触与调查进行比较。我们发现，生成网络的结构和政治同配性随实验配置而变化。更重要的是，在将生成的党派接触与调查进行比较时，一些模型在民主党与共和党ego之间生成的分隔强于调查所报告的分隔，表现为更高的同党接触和更低的跨党接触，而另一些模型则主要低估了各ego群体中的共和党接触。此外，具有相似政治同配性的网络仍可能在相同角色画像周围生成不同的党派环境，包括在具有看似合理的结构属性的网络之间也是如此。最后，生成网络表现出群体特异性误差，即对特定社会群体低估或高估对民主党人或共和党人的接触。这些发现表明，仅通过聚合结果评估基于LLM的社会模拟，可能会掩盖这些结果得以产生的局部环境中的误差，从而突显了在聚合层面和个体层面同时进行验证的重要性。
cs.CL / 107 / 2610.23446
"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws
“你”并不有毒：LLM 标注者如何改写网络的对话定律
large language model
大语言模型相关
Abstract
What the Web community believes about toxic conversations, such as that toxicity rises in longer and deeper threads, was measured mostly with Google's Perspective API, which is shutting down; large language models (LLMs) will make the next measurements. We find that LLM labelers mistake second-person address for toxicity, and that this one bias manufactures laws that human labels do not show. On 1.25 million news comments with crowd labels, Claude Haiku flags 12% of the non-toxic comments that contain no "you" and 51% of those with four or more; a single crowd rater flags 6% and 10%. Deeper replies address someone more often, so under Haiku toxicity rises with reply depth (+0.93 percentage points per level), while under the crowd's labels it slightly declines (-0.12). Rewriting "you" in the third person removes two thirds of the false rise; flagging only as many comments as the crowd does not remove it, and showing Haiku the parent comment enlarges it. All seven LLMs we test over-flag "you", and at the crowd's flag rate, without the parent, any false rise of the open ones vanishes at a fixed number of second-person words; Haiku also over-flags "you" on Reddit, Gab and Wikipedia talk pages. Agreement with human labels does not reveal the bias: we prove that a fixed Cohen's kappa leaves a law nearly unconstrained. We therefore certify the law instead of the labeler. Threadsafe uses the labeler's scores to choose which comments human raters label and to predict the others, and corrects the predictions with the human labels; linear laws are estimated without bias whatever the labeler does. In simulated audits its intervals for the laws keep nominal coverage from 250 labels on, where uniform-sample methods such as PPI++ fall short, and at 1,000 to 8,000 labels are 0.75 times as wide as those of design-based supervised learning (DSL); a thousand human labels rule out Haiku's rising law.
Chinese Translation
网络社区对有毒对话的信念，例如毒性在更长、更深的讨论串中会上升，主要是用谷歌的 Perspective API 测量得到的，而该 API 即将关停；大型语言模型（LLM）将进行接下来的测量。我们发现，LLM 标注者会把第二人称称呼误认为毒性，而这一单一偏差制造出了人类标注并不显示出的定律。在 125 万条带众包标注的新闻评论上，Claude Haiku 将不含“you”的无毒评论中的 12% 标记为有毒，将含有四个及以上“you”的无毒评论中的 51% 标记为有毒；单个众包评分者对应的比例分别为 6% 和 10%。更深的回复更频繁地称呼某人，因此在 Haiku 下，毒性随回复深度上升（每层 +0.93 个百分点），而在众包标注下则略有下降（-0.12）。将“you”改写为第三人称可消除三分之二的虚假上升；只按众包标注相同的数量标记评论并不能消除它，而向 Haiku 展示父评论则会扩大它。我们测试的全部七个 LLM 都过度标记“you”，并且在众包的标记率下、不提供父评论时，开放模型中的任何虚假上升都会在固定的第二人称词数量处消失；Haiku 在 Reddit、Gab 和 Wikipedia 讨论页上也过度标记“you”。与人类标注的一致性并不能揭示该偏差：我们证明，固定的 Cohen's kappa 会使定律几乎不受约束。因此，我们认证的是定律，而不是标注者。Threadsafe 使用标注者的分数来选择哪些评论由人类评分者标注，并预测其余评论，再用人类标注校正这些预测；无论标注者如何做，线性定律的估计都是无偏的。在模拟审计中，从 250 个标注起，其对定律的区间保持名义覆盖率，而 PPI++ 等均匀样本方法在此情形下表现不足；在 1,000 到 8,000 个标注时，其区间宽度是基于设计的监督学习（DSL）区间的 0.75 倍；一千个人类标注即可排除 Haiku 的上升定律。
cs.AI / 108 / 2610.23673
On the possibilities and limitations of LLM-based multi-agent systems for multi-criteria decision-making in healthcare operations management
论基于LLM的多智能体系统在医疗运营管理多准则决策中的可能性与局限性
large language model
大语言模型相关
Abstract
Multi-criteria models for healthcare operations need weights that express how stakeholders trade off patient waiting time, provider workload, and operating cost, but these weights are difficult to quantify. Teams of large language model (LLM) agents, each prompted to represent a stakeholder, have been used to supply them. We propose a framework for assessing the reliability of LLM-generated preferences as inputs to multi-objective optimization and quantifying how their uncertainty affects decisions. We study explicit preferences by asking agents to rate the importance of the criteria in an outpatient appointment scheduling problem, and implicit preferences by inferring trade-offs from the agents' pairwise choices between emergency department configurations. The computational experiments show that the explicit preferences respond to the clinic's operating conditions and differ across roles, but deliberation among the agents can erase these differences, whereas simple averaging preserves them. The pairwise choices used to reveal the agents' implicit preferences depend largely on which option is presented first, and we find limited evidence that a statistical model using contextual information alongside presentation order predicts the agents' choices on new comparison pairs better than a model using presentation order alone. We then formulate robust weighted-sum optimization models to account for preference uncertainty. In most cases, these models select different schedules or configurations from those obtained using average preference weights. These results indicate that LLM-generated preferences need an audit of this kind before they are used as optimization weights in healthcare decision support.
Chinese Translation
医疗运营中的多准则模型需要权重，用以表达利益相关者如何在患者等待时间、提供者工作量和运营成本之间进行权衡，但这些权重难以量化。由大型语言模型（LLM）智能体组成的团队——每个智能体被提示代表一个利益相关者——已被用于提供这些权重。我们提出一个框架，用于评估LLM生成偏好作为多目标优化输入时的可靠性，并量化其不确定性如何影响决策。我们通过要求智能体对门诊预约调度问题中各准则的重要性进行评分来研究显式偏好，并通过从智能体在急诊科配置之间的成对选择中推断权衡来研究隐式偏好。计算实验表明，显式偏好会响应诊所的运营条件，并因角色而异，但智能体之间的审议可能抹去这些差异，而简单平均则会保留它们。用于揭示智能体隐式偏好的成对选择在很大程度上取决于哪个选项先呈现，并且我们仅发现有限的证据表明，一个同时使用上下文信息和呈现顺序的统计模型在预测智能体对新比较对的选择时优于仅使用呈现顺序的模型。然后，我们构建稳健的加权和优化模型，以考虑偏好不确定性。在大多数情况下，这些模型选择的调度方案或配置不同于使用平均偏好权重所获得的调度方案或配置。这些结果表明，LLM生成的偏好在被用作医疗决策支持中的优化权重之前，需要这种类型的审计。
人工智能 (cs.AI)
76
cs.AI / 1 / 2610.22538
Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks
Abstract
Large reasoning models (LRMs) that leverage test-time computation achieve strong performance on mathematical and scientific benchmarks and are increasingly deployed in real-world applications. However, their ability to perform \textit{everyday logical reasoning}, which often departs from textbook formal logic, remains underexplored. In this work, we systematically evaluate how well reasoning models align with everyday reasoning patterns by introducing a benchmark based on linguistic and pragmatic reasoning phenomena. We also compare reasoning models with non-reasoning models under controlled prompting conditions. We find that, in many cases, (1) reasoning models show degraded performance in everyday reasoning when such reasoning deviates from formal logical rules, (2) some non-reasoning models with chain-of-thought prompting outperform comparable reasoning models in everyday reasoning, and (3) prompting with additional instructions provides limited control over reasoning behavior. These findings suggest that current reasoning models do not reliably capture everyday reasoning and point to limitations in post-training, motivating improved training objectives and evaluation across diverse forms of reasoning.
cs.AI / 2 / 2610.22561
FRACTAL: A Self-Similar Multi-Agent Orchestration Architecture with Formal Communication, Convergence, and Depth Bounds
Abstract
We present FRACTAL, a self-similar multi-agent architecture for recursive goal decomposition, in which the same five-component organizational pattern -- Orchestrator, Ensemble, Knowledge Ledger, Decision Maker, Meta-Observer -- recurs at every level of a decomposition tree. We make four formal contributions: (1) a communication bound showing intra-node cost is linear in ensemble size and round count (O(kr)) under a two-phase blind-proposal/stress-test protocol, versus quadratic cost for free-form debate, with an explicit, stated assumption on fixed evaluator count; (2) a two-mode orchestration criterion separating generative and evaluative phases via a measurable diversity-saturation threshold; (3) a supermajority convergence guarantee with explicit probability bounds at stated agreement levels, and a formal depth bound under an operationalized goal-complexity measure; (4) a Bayesian inter-layer escalation mechanism unifying hysteresis and contract-breach signaling under one posterior update. We illustrate the architecture's execution qualitatively on a worked example and a multi-layer case study; we do not report quantitative performance claims in this paper, and we explicitly discuss the formal and inductive-bias limitations of the model, including an unimplemented but architecture-compatible direction toward bounded self-improvement. We use "fractal" as an architectural metaphor for self-similarity, not a measured mathematical property. Empirical validation of the architecture's comparative performance is left to future work.
cs.AI / 3 / 2610.22603
At the Limits of Computability: Steps Towards Autonomous AI Discovery on the Busy Beaver Frontier
Abstract
Recent breakthroughs in AI for mathematics---ranging from Olympiad medal performances to resolving long-standing open conjectures---have rapidly expanded what automated systems can achieve. Yet these successes primarily concern predefined, human-formulated problems, leaving open how AI systems can autonomously uncover mathematical structure and formulate new questions. To push beyond static challenges, we explore the Busy Beaver (BB) domain, an uncomputable frontier that is fundamentally open-ended. Because no algorithm can compute the Busy Beaver number $\mathrm{BB}(n)$ for every $n$, this domain provides a natural setting for sustained exploration through mathematical conjecture and proof. In this paper, we introduce a Gemini-based agentic pipeline that automatically discovers accelerations of Turing machines, expressed as compact mathematical rules summarizing many execution steps. Our system integrates pattern recognition, automated conjecturing, informal mathematical reasoning, and formal theorem proving in Lean. With this framework, our system autonomously discovered and formally verified novel Collatz-like accelerations for eleven machines in $\mathrm{BB}(6)$. Subject to community acceptance, we provisionally classify these machines as Cryptids, namely machines whose blank-tape executions admit exact, compact mathematical descriptions but whose halting behavior remains unresolved. These results advance the analysis of $\mathrm{BB}(6)$ holdouts on the Busy Beaver frontier and demonstrate the potential of autonomous AI systems for open-ended mathematical discovery in fundamental computer science.
cs.AI / 4 / 2610.22668
From Relevant Memories to Sufficient Evidence: Rethinking Memory Utilization in LLM Agents
Abstract
Long-term memory enables LLM agents to reuse past experiences, yet retrieving relevant memories does not necessarily provide sufficient evidence for answering a query. Essential information may remain fragmented across memory records, obscured by abstraction, or disconnected from the relations required for reasoning. We identify this relevance-sufficiency gap and propose QEMem, a framework that reformulates memory utilization as the construction of query-specific evidence sets, shifting the focus from individual retrieval relevance to collective evidential support. QEMem integrates three complementary mechanisms: query-guided discovery retrieves evidence across multiple memory granularities; source-grounded reification organizes it into structured ViewCards with explicit evidential roles and source references; and sufficiency-guided refinement selectively recovers missing information. Operating on existing memory stores without modifying memory formation, QEMem achieves 91.43% LLM-judge accuracy on LoCoMo with Claude Opus 4.5, outperforming the strongest evaluated baseline, MemR$^3$, by 7.6 percentage points and NEMORI by 10.7 points under a shared memory store. Further analyses demonstrate improved evidence coverage and strong performance on questions requiring cross-session evidence integration.
cs.AI / 5 / 2610.22672
Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation
Abstract
Modern mobile GUI agents can perceive visible screens and execute local actions, but they still lack persistent app-specific navigation knowledge: where functions live, which pages contain them, and which screens are equivalent across repeated visits. We introduce TAG-Nav, a Task-Agnostic Graph Navigation framework that wraps existing GUI-agent backbones with reusable destination memory. Offline exploration is distilled into compact records of app functions, target pages, and waypoint routes before any evaluated instruction is seen. When a future task's destination is covered, TAG-Nav retrieves the destination and provides route guidance without task-specific demonstrations. This separates mobile GUI destination navigation from full task completion: many tasks share the prefix of reaching the right page, while final actions may enter content, send data, make purchases, delete information, or change sensitive settings. To evaluate route finding without downstream execution confounds, we introduce DestBench, a navigation-only protocol that asks whether an agent reaches the correct pre-action destination screen. DestBench enables a controlled graph-memory study comparing no-memory agents with random, automated BFS, and clean graph memory. Graph-derived memory improves strict navigation success rate of UI-TARS-1.5-7B from 52.4% without graph memory to 58.3% with automated BFS memory, and improves Qwen-VL-Max success rate from 59.5% to 71.4% on the same conditions. Together, TAG-Nav and DestBench show that destination coverage and route cleanliness strongly shape graph-memory value, making graph quality a measured variable rather than an assumed benefit of exploration.
cs.AI / 6 / 2610.22682
FlexUMM: Flexible Representation Learning for Unified Multimodal Models
Abstract
Unified multimodal models (UMMs) can integrate understanding and generation within a single backbone. However, it is commonly known that these two capabilities can cause conflicting representations, leading to a central question: how representation and computation should be shared across functionalities? Existing UMMs often rely on a largely fixed structure, where different inputs and functionalities follow the same depth and parameter-sharing pattern. We argue that this assumption is too restrictive: more effective unification requires sharing the computation that supports cross-functional communication while specializing the capacity that becomes task- or input-specific. We introduce FlexUMM, a flexible UMM framework that dynamically decides when to preserve shared computation, when to route specialized capacity, and when to stop further processing. Evidence shows that this flexible allocation improves UMMs along two complementary axes. It improves accuracy by reducing harmful competition between functionalities, and it improves efficiency by avoiding unnecessary computation once an input representation is sufficiently mature. Beyond simple architectural unification from the pre-trained models, FlexUMM works as a model-agnostic, functional plug-in to make existing UMMs more unified, stronger, and efficient. Code is available at: https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/flexumm.
cs.AI / 7 / 2610.22698
Repair, Retry, and the Cost of Hindsight: Verifier Feedback in Formal Theorem Proving
Abstract
Verifier-guided self-repair conditions a model on feedback from failed attempts before it tries again, without weight updates. Yet two findings in prior work are in tension: verifier feedback can improve repair of failures, but self-repair itself can provide little-to-no advantage over independent resampling. We reconcile these findings in Lean theorem proving through controlled one-step experiments that hold the preceding failure fixed while varying whether the failed proof is shown, whether the compiler error corresponds to that failure, and the retry instruction. We test frozen Qwen3.6-27B, a general-purpose model, and DeepSeek-Prover-V2-7B, a specialized prover, on difficult MiniF2F problems. We find two opposing effects. Showing the model its failed predecessor reduces success and sharply increases copying of it, while on Qwen adding that attempt's compiler error recovers the loss, to the level of the diagnostic alone. Replacing the predecessor with another theorem's failed proof largely removes the degradation on both models; on Qwen, replacing the diagnostic with another problem's error substantially reduces the recovery. Without the failed proof, correct diagnostics show no resolved advantage over wrong-diagnostic controls on either model. The failed-proof penalty persists with and without the retry instruction, whereas Qwen's diagnostic recovery is weaker without it. In 11-problem sequential experiments through 64 attempts, Qwen retains a correct-diagnostic benefit within failed-proof histories, while DeepSeek-Prover shows no resolved recovery. However, no tested one-step feedback condition establishes an advantage over instruction-only resampling, nor does 64-attempt sequential repair with failed proofs and correct diagnostics at equal token budgets. Verifier feedback can thus improve repair without overcoming the performance loss from conditioning on a failed attempt.
cs.AI / 8 / 2610.22708
GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations
Abstract
Text-based world models have recently shown potential to simulate complex RL environment state transitions for a diverse set of use cases. However, no standardized benchmark exists to compare, study and improve these models. Through this paper, we attempt to solve this problem by building GlobeBench, a novel benchmark with 1,632 tasks across five separate domains and nineteen simulation skills. The tasks check whether models can predict the environment's response to a specific action given the full agent trajectory by understanding earlier state changes, following environment rules, and predicting the effects of actions on both related and unrelated objects. We compare the predicted environment response with an execution-derived reference, using a hybrid architecture with both deterministic gates and a semantic judge. Among 20 general-purpose LLMs and specialized text world models, GPT-6 Astra achieves the highest task success rate at 49.57%. We find that model accuracy decreases as actions affect a larger portion of the environment. Specifically, when the agent's changes must propagate to dependent objects, task success drops to an average of 5.1% across all 20 models. Although model strengths differ across tasks, 631 tasks remain unsolved by any model, suggesting a significant gap in general LLM capabilities as a world model.
cs.AI / 9 / 2610.22734
Graph Serializations for Reasoning LLMs: A Study Across Knowledge Extraction and Structural Reasoning
Abstract
The way a knowledge graph is serialized as text for inclusion into LLM context can substantially affect downstream LLM performance, with prior work reporting tens of percentage points variations across formats. However, existing studies consider only a limited set of representations and predate frontier reasoning models, which now offer stronger reasoning, instruction-following, and long-context capabilities. We evaluate 15 graph representation formats across nine frontier LLM configurations on ontology-based knowledge extraction and structural graph reasoning tasks. While knowledge extraction is relatively insensitive to different serialization formats, the choice of the optimal format for structural reasoning tasks significantly affects accuracy. We demonstrate that the optimal representation choice depends on model, query, and task and propose to approach graph serialization as a dynamic orchestration decision.
cs.AI / 10 / 2610.22740
R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
Abstract
In recent years, agentic orchestration has been widely applied to solve complex human tasks. However, existing orchestration learning still faces key challenges, including hand-designed curricula, advantage collapse when a task's rollouts all reach the same outcome, and unreliable evaluation signals that judge each skill by one noisy score. To address these challenges, we propose R$^2$Orch, a new paradigm of Recursive Self-Play Orchestration: an inner loop of self-play that improves the orchestrator on self-chosen tasks, and an outer loop that improves how the inner loop improves. To support this paradigm, we introduce Recursive Learnability, a training objective that rewards a Proposer for choosing tasks whose rollouts keep reaching different outcomes as skills are withdrawn. Built on this objective, we further propose Skill Withdrawal, in which the outer loop withdraws each skill once a controlled evaluation within self-play shows that the orchestrator performs as well without it. Moreover, the two loops form one cycle: failures are distilled into skills, skills into parameters, and withdrawals lead the Proposer to new failures, so the orchestrator moves toward zero-skill deployment. Experimental results on twelve datasets show that R$^2$Orch consistently outperforms baselines across question answering, math, code, and interactive tasks.
cs.AI / 11 / 2610.22760
Between Spoilers and Guesswork: Avoiding Over-Specification in Coding Benchmarks
Abstract
Benchmarks drive AI progress, but are hampered by task under-specification and over-specification. Task under-specification causes guesswork, where solutions have to make things up, or worse, resort to contamination or reward hacking. Task over-specification causes spoilers, where solutions can bypass some of the harder steps the benchmark was supposed to measure. Since future-looking benchmarks should be difficult, over-specification reduces a benchmark's shelflife of community adoption. Repository-level coding benchmarks have helped the field progress but experience unnecessary churn due to both over- and under-specification. For instance, SWE-bench Pro, FEA-bench, Doc2Feat-bench, and Multi-SWE-bench all tackle under-specification by augmenting task statements with hints derived from ground truth. We show that leaking ground-truth information into the task statement causes over-specification in some of these benchmarks. We experiment with various specification levels, trying to thread the needle between spoilers and guesswork. Based on these results, we recommend techniques for deriving parsimonious hints. We will release our hints so the community can obtain more reliable signals from popular benchmarks.
cs.AI / 12 / 2610.22768
StoBasket: Joint Upsell and Cross-Sell Optimization for Contribution Margin
Abstract
Retail recommenders rank relevance, but retailers earn contribution margin: an upsell earns the margin difference over the item it replaces, whereas a cross-sell earns its full margin. We introduce joint stochastic basket editing, a formulation in which one policy selects replacement and addition offers for an existing basket, values them by expected incremental contribution margin, and couples them through displayed-offer, accepted-offer, and joint-acceptance response channels. To our knowledge, it is the first formulation in which upsell and cross-sell responses are coupled; prior separable treatments are its zero-coupling special case. In real-data-calibrated simulations on five public basket datasets, StoBasket captures essentially all attainable offer margin: within 0.5% in every default-price setting and within 0.7% when the true response lies outside its model family. Correct economic accounting is the broadest gain: margin-weighted popularity leaves up to 11% of attainable margin unrealized, and a published next-basket recommender leaves 11-52%. Joint selection adds 10.2% of attainable margin on Instacart, where premium tiers and add-ons are strongly linked. When offers pay only together, the best heuristics lose 64% while exact search stays within 0.5%. For a structured subclass, we characterize when staged policies fail and give an exact optimizer that solves every tested real basket in 0.13 ms median and at most 111 ms. Profit saturates long before prediction accuracy: a correctly structured predictor trained on 5,000 randomized sessions comes within 0.06% of perfect prediction, while a signal-free one loses 38%. The same estimates yield a decision-aware market-basket analysis whose cost does not grow with association order.
cs.AI / 13 / 2610.22779
When Can You Trust an Agent's Own Verifier? Stress-Testing Counterfactual Rationale Verification Under Search Pressure and Oracle Access
Abstract
Automated agent design increasingly trusts a single score to decide whether a change is an improvement -- but a score optimized aggressively enough ceases to measure what it was meant to. When an agent also explains its design, that explanation is a second signal we can check without human labels: negate each stated claim, re-run, and observe whether performance drops. We ask whether such a verifier stays reliable once an agent optimizes against it. Our contribution is not a new verifier but a way to stress-test one: a pressure ratio $ρ$ measuring how much the verifier steers the agent's search, and four falsifiable probes covering four failure modes -- giving no signal, being too weak to matter, being gamed by stating fewer claims, and being queried before the agent commits. We apply the test to our own verifier, the Rationale Consistency Score (RCS), inside an LLM agent that designs machine-learning pipelines, with one generator model, six tasks, and three seeds. First, in our prior configuration the verifier barely affects the search objective ($ρ\ll 1$), partly because of an accuracy gate, so its robustness had not been tested; it still shapes the reported artifact through a final-selection tie-break that $ρ$ does not register. Second, a capability control shows the agent shortens its claim lists when instructed, yet search does not, even at $ρ$ over 100 times the prior level; since feedback omits claim volume and horizons are short, this null result is largely structural rather than evidence of robustness. Third, when the agent can probe the verifier before committing, it keeps claims that pass while held-out faithfulness falls -- a decline invisible to in-loop metrics, though partly confounded by how RCS scores small pipelines. We offer these as preliminary evidence and propose a reporting standard: audit a verifier with $ρ$ and these probes before trusting it.
cs.AI / 14 / 2610.22805
Compression Precedes, Content Decides: Emergent Misalignment in Reward Models
Abstract
Emergent Misalignment - the out of domain generalization of narrow harmful fine-tuning first shown by Betley et al. [2025] has been widely studied in language models but not in reward models that supervise them. In this paper, we fine-tune two state-of-the-art reward models, Skywork-Reward-V2-Llama-3.1-8B and Skywork-Reward-V2-Qwen3-8B [Liu et al., 2025] and show that narrow fine-tuning reward models to prefer harm in one domain (medical or financial) generalizes to seven held out domains. This is not degeneracy, as the fine-tuned reward models lose only 4.8 and 5.3 points overall (RM-Llama) on the standard RewardBench 2 evaluation [Malik et al., 2025] while still outperforming the unmodified RM-Qwen model. We then fine-tune the underlying language models, Llama-3.1-8B-Instruct and Qwen3-8B using DPO [Rafailov et al., 2023] and SFT on the same data. We find that reward models are more vulnerable to preference inversions compared to their language model counterparts. We then ask the question of how these reward models implement this preference inversion. Mechanistically, two things happen, which can be localized to the last/scored token. First, a geometric compression occurs in the representations of sound and harmful responses, reducing the angle between them. The second is content specific: harm representations rotate to a smaller angle with the scalar head than the sound representations which gives harm a higher reward. This happens only in harmful preference fine-tunes and not in control.
cs.AI / 15 / 2610.22825
World Action Critic (WAC) Model: Joint Foresight and Criticism for Policy Learning
Abstract
World models have shown considerable promise in sequential decision-making by anticipating the consequences of actions and providing predictive foresight for policy learning. However, learning to reason about counterfactual actions remains challenging, since predictive objectives may underemphasize information critical for action evaluation, while unexecuted actions lack observed futures for direct supervision. To address these challenges, we propose the World Action Critic (WAC) Model, which proposes multimodal action candidates and jointly learns candidate-conditioned future prediction and consequence evaluation. In our proposed framework, critic supervision directly shapes intermediate predictive representations during rollout, encouraging them to retain information needed to distinguish action quality. Observed futures supervise prediction for executed actions, while external task-quality targets provide evaluation signals for counterfactual candidates. Through joint optimization, critic signals from these candidates propagate into predictive representations without requiring their observed futures. Our theoretical analysis characterizes critic-guided feature retention and bounds candidate-selection regret. Applied to autonomous driving, WAC achieves 93.4 PDMS on NAVSIM-v1, demonstrating the effectiveness of jointly learning predictive foresight and consequence evaluation.
cs.AI / 16 / 2610.22861
Evo-B&B: Behavior-Aware LLM-Guided Joint Evolution of Branching and Node Selection Policies for MILP
Abstract
In branch-and-bound for Mixed-Integer Linear Programming (MILP), branching and node selection are two critical and tightly coupled decisions, yet most existing learning-based approaches focus on optimizing them separately. We propose Evo-B&B, an LLM-guided evolutionary framework that jointly evolves executable branching and node selection policies using end-to-end SCIP solving feedback. To improve search efficiency and reduce redundant evaluations of behaviorally similar policies, Evo-B&B introduces a behavior-aware multi-island evolutionary mechanism. It measures policy similarity from branching and node-selection behaviors on a fixed set of B&B decision states and uses this similarity to preserve population diversity. We evaluate Evo-B&B on four standard MILP benchmarks across multiple instance scales. Experimental results show that Evo-B&B achieves competitive solving performance and generalizes well to larger-scale instances, while ablation studies further demonstrate the effectiveness of joint evolution and behavior-aware diversity management.
cs.AI / 17 / 2610.22869
MS-Eval: Benchmarking Sub-Agent Delegation and Orchestration in MCP-Intensive Working Scenarios
Abstract
LLM agents are increasingly integrated into real-world works, yet they must handle diverse domain-specific tools and complex multi-sub-task problems that sequential tool invocation resolves inefficiently and with severe context bloat. The Main-Agent Delegation and Sub-Agent Execution (MAD-SAE) paradigm addresses these issues by delegating sub-tasks to specialized sub-agents for parallel execution in isolated contexts, but its delegation and orchestration quality remains unquantified. We introduce MS-Eval, a unified evaluation kit comprising an MCP-intensive benchmark, a controlled execution harness, tailored delegation and orchestration metrics, and the MSQ index for comprehensive assessment. Evaluating four MAD-SAE variants and eight frontier models reveals that (1) sub-agent specialization (Spec) is the best practical practice; (2) open-source models are largely competitive with few exceptions; and (3) MAD-SAE generally raises cost with contingent payoff, benefiting most open-source models in both performance and efficiency while leaving closed-source models with higher cost and degraded performance. Our case study further traces these gains to better single-agent solutions, effective context offloading, and genuine sub-task parallelism, whereas unnecessary or redundant delegation turns the same cost into pure overhead. Codes: https://github.com/LINs-lab/MS-Eval
cs.AI / 18 / 2610.22871
ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers
Abstract
Federated Prompt Tuning (FPT) adapts large vision--language models by freezing the pretrained backbone and optimizing only lightweight prompt parameters across clients. Although this design improves communication and parameter efficiency, it creates an overlooked security risk. Since the backbone is shared and frozen, a malicious client can induce backdoor behavior through the representation space while keeping its uploaded prompt updates close to benign ones. We propose \textbf{ShadowFPT}, a targeted backdoor attack that exploits this frozen-backbone attack surface. ShadowFPT first pretrains a learnable \emph{Shadow Trigger} against the frozen CLIP visual encoder, using either auxiliary public data or the malicious client's local data, so that triggered inputs are steered toward the target class in representation space. During federated prompt tuning, the malicious client adapts the trigger under the current global prompt and then optimizes its local prompt on both clean and triggered samples. Only prompt parameters are uploaded to the server, while the trigger and frozen encoders remain local. By shifting most of the attack burden from prompt updates to trigger-induced representation steering, ShadowFPT achieves targeted misclassification while preserving prompt-space stealthiness. Across multiple datasets, aggregation rules, and non-IID federated partitions, ShadowFPT increases the attack success rate from 19.81\% to 90.36\% in our main setting, while maintaining clean accuracy. It remains effective across textual, visual, and joint vision--language prompt tuning. These results identify frozen backbones as stealthy and underexplored backdoor surfaces in federated prompt tuning, suggesting that defenses based only on prompt-update anomaly detection are insufficient.
cs.AI / 19 / 2610.22872
Look Back, Correct Forward: Error Feedback Calibration for Time Series Forecasting
Abstract
Time series forecasting models are increasingly deployed in continuously evolving environments. After a forecast is issued, future observations gradually arrive and reveal how the model has failed. Existing methods mainly treat such observations as supervision for parameter updates. However, it remains unclear whether verified forecasting errors can themselves provide causal feedback for correcting future predictions. To address this, we introduce Error Feedback Calibration (EFC), a fully test-time framework for deterministic post-hoc correction of time series forecasting. EFC constructs a Causal Error Diagonal (CED) from revealed errors of overlapping historical forecasts, providing a fresh view of the frozen forecaster's recent failure behavior. Then Cumulative Error Transport (CET) maps this error state to future residuals using cumulative estimation, while temporal out-of-sample validation controls the correction strength. Temporal Residual Learning (TRL) further predicts only the residual structure left after CET from recent CED trajectories, rather than relearning the entire forecasting error. Extensive experiments across forecasting backbones and datasets show that EFC broadly improves forecasting performance while substantially controlling harmful over correction.
cs.AI / 20 / 2610.22883
ToPA: Block-wise Toeplitz Adaptation for Expressive and Efficient Fine-Tuning
Abstract
Parameter-Efficient Fine-Tuning (PEFT) has emerged as a prevalent strategy for adapting large pre-trained models. Among PEFT techniques, reparameterization-based methods, particularly those that employ low-rank or sparsity structures, have attracted significant attention. However, these approaches often underperform full fine-tuning. Specifically, low-rank methods constrain the update rank, which is misaligned with the inherently high-rank nature of full fine-tuning. In contrast, sparsity-based techniques limit the flexibility of the parameter space and compromise connectivity across weights. To overcome these limitations, we explore the use of Toeplitz matrices, whose entries are constant along each diagonal, therefore providing a compact parameterization without imposing explicit rank or sparsity constraints. A straightforward approach is to update weights via a product of Toeplitz matrices. However, similar to RNNs, long multiplicative chains often lead to gradient instability. To address this issue, we propose \textbf{ToPA}, a more refined version of this idea that replaces standard Toeplitz matrices with block-wise Toeplitz, greatly increasing each factor's capacity and shortening the chain, thereby stabilizing training and improving expressiveness. We theoretically show that products of block-wise Toeplitz matrices can approximate arbitrary matrices under shorter chain length, justifying the structural design of ToPA. Furthermore, we demonstrate that ToPA offers greater expressivity than existing low-rank and sparse parameterizations. Empirical evaluations across 21 NLP and CV datasets, spanning 5 model architectures, consistently validate the effectiveness of ToPA.
cs.AI / 21 / 2610.22897
When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents
Abstract
Long-horizon agents learn from failure, but the value of a failure record can change as execution progresses. We study when retaining, compressing, or removing the same history block improves subsequent execution relative to full-history KEEP. We introduce RECAP, a framework that combines matched context interventions, execution-relational prediction, and independent policy certification. Matched environment states and rollout seeds measure success and continuation prompt-cost differences; prefix-only relations between a failure block and the current execution propose promising edits. Independent task bundles then select a policy subject to harmful-intervention and non-positive-benefit risk constraints under a fixed measurement protocol. A five-seed longitudinal case holds both the failure block and its summary fixed: compression fails at an early state and becomes a cost-saving intervention at a later state. Across five backbones, RECAP produces positive paired-success saving estimates with backbone-specific reliability-efficiency tradeoffs. On Hunyuan-A13B, it edits 43.24% of retained checkpoints while matching KEEP's observed 91.89% continuation success and reducing unconditional mean continuation prompt cost by 4,619.8 tokens (2.13%). The framework provides an execution-level reference for deciding when to forget: retain history by its value to what comes next.
cs.AI / 22 / 2610.22915
Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents
Abstract
Open-ended tool agents can complete the same objective through different tool paths, and their quality depends jointly on the interaction trace, environment state, and final artifacts, so such tasks rarely admit an executable ground-truth reward. The central challenge is therefore to construct a reward from existing experience that exposes the current policy's failures while remaining stationary during optimization. We introduce Reward Engine, which, like a self-motivated student, first works out what success requires and then optimizes toward it. From the task and initial trajectories, it builds a RewardPack of testcases for required correctness and rubrics for graded quality, each executed as a check that returns a judgment with evidence. The frozen RewardPack then guides the agent to revise and select its outputs, and can also serve as a reward for reinforcement learning. On WorkBuddy Bench Office and Claw-Eval, the constructed reward largely agrees with the benchmark graders, not reversing their ordering on 82.3\% and 80.5\% of candidate pairs. Used to revise executions, it raises three-run consistency ($\mathrm{Pass}^{3}$) over Baseline from 45.0\% to 54.0\% and from 78.4\% to 83.7\%.
cs.AI / 23 / 2610.22987
From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents
Abstract
Coding agents use execution harnesses to constrain the consequences of prompt injection by enforcing permission boundaries. These systems deterministically intercept any out-of-sandbox action and prompt the user for approval. To reduce approval fatigue, users can approve allow-rules proposed by the agent, after which any future action matching the rule is allowed without a prompt. In this work, we show that this mechanism introduces a new risk: untrusted context can influence the permissions presented for approval, with consequences extending beyond the immediate task. We investigate privilege escalation targeting this permission-update process. Specifically, adversarial context planted in a repository can reliably steer an agent performing a routine task into requesting a plausibly justified, yet dangerously broad allow-rule. If the user approves the request, the saved rule can effectively blind the harness, enabling adversaries to execute malicious actions that would otherwise require approval, without triggering human oversight. Across production platforms, including Claude Code and Codex CLI, we demonstrate that agents successfully elicit the target allow-rules in 58% of experimental instances. Once acquired, these rules exhibit a massive semantic blast radius, covering 56.9% of category-specific capability test cases. To further quantify the downstream risk, we show that conditional on rule approval, our exploit achieves verified, silent goal completion in 47 of the 58 eligible instances (81.0%), undermining the safety promise of human-in-the-loop oversight.
cs.AI / 24 / 2610.23006
Efficient CPU-GPU collaborative MoE inference via Adaptive Expert Assignment
Abstract
Mixture-of-Experts (MoE) models reduce inference computation through sparse expert activation, but their large expert memory footprint often exceeds GPU capacity, making expert offloading necessary for memory-constrained inference. However, existing systems typically handle GPU-missed experts by either transferring them to the GPU, incurring substantial PCIe transfer latency, or executing them on the CPU, making the slower CPU computation the bottleneck. To address these limitations, we present \method, a CPU-GPU collaborative MoE inference system that adaptively assigns these missed experts to the CPU or GPU for efficient execution. By exploiting differences in expert token-loads, \method selectively transfers high-load experts to the GPU while executing low-load experts on the CPU. To minimize inference latency, \method adaptively adjusts expert assignments to balance the CPU computation path against the GPU transfer-and-computation path, allowing both execution paths to complete as early as possible. Extensive experiments on three representative MoE LLMs show that \method improves prefill speed by up to 51.9\% and decode throughput by up to 39.5\% over the strongest baseline across the evaluated batch sizes, with consistent gains under different GPU-memory budgets.
cs.AI / 25 / 2610.23012
Adaptation of Computer Use Agents to New Software with User Manuals
Abstract
Recent advances in multimodal foundation models have significantly improved the ability to operate real-world software through graphical user interfaces (GUIs) using visual observations and natural language instructions. However, maintaining strong performance on unseen applications remains challenging, as practical tasks often require long-horizon procedural knowledge and application-specific operation patterns. In addition, constructing executable tasks and collecting corresponding trajectories for training still heavily rely on human-annotated data, incurring substantial cost. In this paper, we propose MAGE (Manual-augmented computer use AGent Execution and learning), a framework that leverages user manuals as a scalable source of procedural knowledge to improve both inference and training. During execution, MAGE retrieves task-relevant manual content and uses a planner to generate explicit procedural guidance for CUAs. To further account for the discrepancies between manuals and actual GUI environments, such as version differences or user settings, MAGE also adapts the agent policy through environment interaction. During training, MAGE converts manuals into executable task-plan pairs and uses the plans to guide environment rollouts toward task-relevant states; the resulting trajectories are then used to optimize the policy without human-authored demonstrations. This unified design enables effective utilization of manuals for both task execution and training, reducing reliance on human-authored training data. Experimental results across six applications demonstrate improvements in task success rates over strong baselines, highlighting the effectiveness of manual-guided planning and policy adaptation in new software environments. We provide the code and reproduction scripts at https://github.com/ku-autoai/MAGE.
cs.AI / 26 / 2610.23025
Causal Value Attribution of Marketing Labels in LLM-Generated Ads: A Deconfounder Approach
Abstract
As AI-generated content becomes commoditized, economic value increasingly shifts from final artifacts to the upstream expertise used to specify inputs that steer LLM generation. Yet organizations lack methods for valuing this expertise and identifying which individual skills or configurations causally improve performance. This study examines generative marketing, where structured expert labels serve as skill primitives conditioning LLM-generated advertisements. Estimating their causal effects is difficult because outcomes reflect stochastic LLM mediation, high-dimensional label interactions, and multimodal confounding. This study proposes D-VAE (Deconfounder Variational Autoencoder), a counterfactual-inference framework integrating a multi-treatment deconfounder with causal representation learning. D-VAE recovers causally disentangled expert-labeling logic and derives PAC-based bounds on omitted-variable bias from single-treatment confounders. Applied to 29,177 foundation-makeup ad creatives, D-VAE outperforms state-of-the-art benchmarks. A field experiment shows that recommended configurations increase engagement by 32.7% and perform statistically indistinguishably from human experts, while a simulation study confirms causal attribution under complex confounding and treatment interactions. Overall, the study shifts value assessment from AI-generated products to skills, providing a foundation for pricing and optimizing AI-augmented skill portfolios.
cs.AI / 27 / 2610.23027
ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting
Abstract
External text can supply predictive information absent from numerical histories, but its value depends on how it is represented, combined and deployed. We present Attribution-guided Semantic Evolution (ASE), a modular framework for constructing textual factors, evaluating their conditional contribution, and deploying validation-selected numerical--textual predictors. ASE supports task-specific implementations with lexical measurements, semantic primitives or evolved event/state factors, coupled to suitable numerical experts. Its feedback-based implementation uses conditional refitting diagnostics to guide factor revision and combination search. We evaluate ASE on the four text-associated TESS datasets and dynamic CSI 300 stock ranking. The four forecasting implementations obtain test MSEs of 0.001041, 0.022439, 0.333115 and 1.8150, each below the corresponding published TESS score. On CSI 300, the frozen text-fusion implementation improves mean RankIC by 0.000642 over its numerical expert and by 0.000701 over matched shuffled text, with positive differences in all five seeds. Controlled text perturbations support functional reliance on text correspondence, while incumbent retention prevents regressions in the measured development criterion. Together, the experiments characterize ASE at two levels: forecasting performance of task-specific implementations and the conditional contributions of individual modules.
cs.AI / 28 / 2610.23044
Where Does the Performance Go? Understanding and Recovering Performance Regressions in C-to-Rust Translation
Abstract
Rust is increasingly adopted as a memory-safe successor to C, and transpilers such as \crust{} can migrate entire C projects while preserving their behavior. Preserving behavior, however, does not guarantee preserving performance. We study this gap on 20 real-world C projects and their \crust{} translations. On the eight projects used for analysis, the translated Rust is more than 3\% slower than the original C on 12 of 19 operations, by up to 62.2\%. By comparing execution-cost profiles, optimized LLVM IR, and compiler optimization remarks across each C--Rust pair, we trace these slowdowns to nine recurring performance issues, spanning residual runtime work, missed compiler optimizations, and C-style source patterns, and derive an optimization direction for each. We then present \ourtool{}, an LLM-based approach that encodes these issues as reusable optimization rules to recover the performance of translated Rust. \ourtool{} profiles each project to locate hot functions and matches them against rule-specific patterns. The LLM then checks each rule's rewrite preconditions against the project context before rewriting, and a rewrite is kept only if it passes both functional and performance validation. On 41 operations from twelve held-out projects, \ourtool{} preserves C outputs on all tested workloads and speeds up 40 operations over the \crust{} translations, 18 of which end up faster than the original C. It outperforms the general-purpose coding agent \codex{} on 31 of the 37 operations where both preserve C outputs, and outperforms its own variant without rule guidance on 34 of the 41 operations.
cs.AI / 29 / 2610.23048
Distill Until Divergence: Error-Truncated Hindsight Distillation for Agentic RL
Abstract
Long-horizon interaction requires agents to make multiple decisions in a changing environment. Outcome-based GRPO lacks fine-grained guidance for intermediate decisions. Recent methods distill completed trajectories into reusable skills and use hindsight-conditioned OPD to provide dense token-level credit. However, skill-induced probability shifts do not necessarily represent reward-aligned credit; especially after an erroneous action changes subsequent interaction states, skill-conditioned supervision may become mismatched with the current state. To address this mismatch, we propose \methodname{}, which routes two complementary hindsight skills according to trajectory outcomes and error boundaries: Positive skills capture successful workflows to reinforce successful trajectories and pre-error decisions, while Negative skills capture failure-specific corrections to target decisive errors. The post-error suffix is optimized exclusively by GRPO, retaining outcome-based learning without extending potentially mismatched hindsight supervision. Extensive experiments demonstrate competitive task performance and support effective fine-grained credit allocation.
cs.AI / 30 / 2610.23051
Agentic In-Context Learning: What Algorithms Can Transformers Learn in Closed-Loop Interaction?
Abstract
The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples. Language model agents extend this setting through interaction: their actions determine what evidence enters context next. We call this regime Agentic In-Context Learning (Agentic ICL). Existing studies largely depend on externally supplied contexts or reward-bearing interaction histories, while work on general tool-using agents lacks a learning algorithm view. We represent an interactive learner by its state, update, and decision rule, covering active learning, bandits, and online reinforcement learning (RL). For expressivity, we construct causal Transformers that approximate these algorithms over a finite horizon, and bound the resulting loss in return. For training, we characterize when optimizing process rewards preserves behavior that maximizes the final outcome. Distillation experiments show that Transformers can learn to execute three representative recursions. We then train models from outcome or process rewards without oracle action or state targets, as in RL post-training. Different rewards yield distinct decision policies. Improving the alignment of process supervision with outcome-optimal action preferences shifts learned decisions accordingly. Such behavioral differences persist in a pretrained language model across textual paraphrases.
cs.AI / 31 / 2610.23085
Rethinking GRPO for Large Reasoning Model Safety Alignment
Abstract
Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals. In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.
cs.AI / 32 / 2610.23104
WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning
Abstract
Real-world frontend development requires agents to compose reproduction, generation, modification, and repair over an evolving application state. Learning a general policy for this process requires translating heterogeneous execution feedback into task-aligned rewards and effective credit assignment across long sequences of interdependent actions. We introduce WebCoder, a general frontend development agent trained with artifact-grounded agentic reinforcement learning. Our key idea is to assess diverse development activities through their effects on a shared executable artifact and its satisfaction of task requirements. We build Webdev-World, an executable environment that integrates code editing, application execution, browser inspection, and on-demand skills while preserving project state across interactions. An independent verifier evaluates the final application against task requirements to provide terminal rewards, while changes in intermediate artifact progress redistribute the resulting rollout advantage across development segments. Context compaction and message-tree sample construction support training on trajectories beyond a single context window while avoiding repeated loss computation on shared generated history. We further introduce Webdev-Bench to evaluate composed development requirements within a single project. On Webdev-Bench, WebCoder-27B raises requirement satisfaction from 64.8 for its Qwen3.8-27B base model to 72.8, exceeding Claude-Opus-4.8 (65.1) and GPT-5.4 (59.6); on Vision2Web, it improves the average visual and functional score from 41.1 to 62.4, on par with Claude-Opus-4.8. Ablation studies confirm the contributions of artifact-grounded credit assignment and long-horizon training.
cs.AI / 33 / 2610.23151
CIIR - Class-Indexed Interface Representation Learning
Abstract
Multimodal learning commonly assumes paired or synchronized observations, yet independently collected datasets rarely provide such correspondence. We introduce CIIR, a modular encoder--decoder framework that connects heterogeneous datasets through a class-indexed address space while preserving each dataset's native input format and output module. A known local index selects address $k$, and an input-dependent 128-dimensional address representation Z_k(x) is written to that address. To demonstrate CIIR, we learn this address space from handwritten digits, clothing images, spoken-digit spectrograms, and hand signs without synchronized or instance-matched samples. Newly trained label-free receivers recover the requested addresses with 94.7% macro-average Address Realization Accuracy. We then attach the unseen SVHN dataset after freezing the base address space and train only lightweight adapters, achieving 90.2% MNIST>SVHN and 82.3% SVHN>MNIST end-to-end Address Transfer Accuracy without target-side decoder labels. To demonstrate generalizability of CIIR, we reuse the ten addresses as code elements in collision-free arithmetic and hierarchical address codes for all 100 CIFAR-100 classes and evaluate direct arithmetic-code realization through the existing receivers. These results establish CIIR as a compact interface for modular cross-dataset communication and post-hoc dataset attachment without exhaustive paired multimodal supervision.
cs.AI / 34 / 2610.23153
QuAD: Quantitative Assessment via Distribution-Alignment for Qualitative Data Quality Enabled by a Content-Aware Multi-Agent System
Abstract
Assessing whether qualitative responses collected from semi-structured questions are ready for a particular analysis goal is a persistent challenge. Manual qualitative analysis is interpretive and labor-intensive, while existing automated methods often lack a dataset-level perspective for deciding whether a response set supports a researcher-selected analytical lens. We introduce QuAD, a framework for quantitative pre-analysis assessment of qualitative response datasets. QuAD compares a reconstructed intrinsic cluster distribution reconstructed from responses with an extrinsic, researcher-defined analytical lens derived from metadata. QuAD is structured into a four-quadrant assessment whose metrics are read jointly, from lens-level relations to within-cluster variation, to provide fitness-for-purpose signals for revealing general patterns and unique insights. Because this assessment relies on lens-conditioned inter-response dissimilarity, we propose a content-aware multi-agent system with two-stage instruction generation and adaptive anchors to improve scoring consistency. We validate QuAD through controlled experiments on synthetic data, including robustness and baseline comparisons, and a real-world social survey case study. The case study shows how QuAD quantitatively assesses whether a selected analysis axis is promising and also surfaces representative or unusual responses for human interpretation.
cs.AI / 35 / 2610.23161
Algorithm Modelization: Transforming Optimization Algorithms into Evolvable Intelligence
Abstract
Can algorithms be learned and evolved as models, rather than repeatedly redesigned as programs? We explore this question by introducing Algorithm Modelization, a new paradigm that turns iterative algorithms into trainable models. Its central idea is to regard algorithm execution trajectories as computational experience: by learning how algorithmic states evolve over iterations, a model can internalize the knowledge embodied in existing algorithms, reproduce their computational behaviors, and further evolve beyond them through training. As a first realization of this paradigm, we modelize iterative optimization as an autoregressive process, where preceding generations constitute the computational context for generating the next population. We develop a dimension-independent population-temporal architecture to capture both population interactions and iterative dynamics, together with an imitation-to-surpassing learning framework that first acquires algorithmic knowledge from heterogeneous teachers and then improves the resulting optimizer directly through optimization feedback. The modelized optimizer consequently goes beyond reproducing individual algorithms: it consolidates heterogeneous search knowledge into a single trainable model, achieves stronger optimization capability through subsequent improvement, and generalizes its acquired knowledge across problems and substantially larger unseen dimensions. More fundamentally, Algorithm Modelization makes algorithmic capability accumulable through learning, where new algorithms and problem-solving experience can become training resources for progressively improving the model. Our results establish a promising path from manually designed algorithms toward general algorithm models that learn, generalize, and evolve from computational experience.
cs.AI / 36 / 2610.23167
LongForge: Measuring and Improving Long-Context Instruction Data via Evidence-Assertion Reasoning Graphs
Abstract
Existing quality control for long-context instruction-tuning data is a one-shot judgment inside the generation pipeline that leaves only a scalar, revealing neither which document parts an answer uses nor how many reasoning steps it writes down; the data can thus be neither audited nor controlled. We formalize data quality as a property of the alignment structure among document, question, and answer, and build LongForge, which pairs an analyzer that measures long-context data with a synthesizer that generates it. The analyzer reverse-reconstructs each sample into an evidence-assertion reasoning graph and programmatically computes four dimensions on it: breadth, depth, correctness, and completeness. Every reading traces back to specific evidence spans and reasoning steps, making data auditable. Measuring seven mainstream datasets reveals two previously unquantified defects: breadth barely grows with document length, and depth is generally low, the two being nearly orthogonal. The synthesizer then embeds the analyzer into generation, gating on correctness and completeness and deferring breadth and depth to a training-time mixture over measured values, which makes the training distribution controllable. On two backbones, its two mixtures, LongForge-1 and LongForge-N, outperform the strongest baseline by 2.2-4.7 points on average over six long-context benchmarks, and in a controlled comparison that holds the gate and the candidate pool fixed, raising breadth and depth alone is worth a further 1.16 points, making the structural distribution of training data a lever on downstream quality rather than only a description of it.
cs.AI / 37 / 2610.23194
OfficeTown: Generative Organizations for Evolving Agent Intelligence
Abstract
AI agents increasingly use files and software tools to produce professional deliverables. Such work draws on materials, decisions, and feedback accumulated across assignments. We introduce OfficeTown, a generative simulation framework for professional organizations in which tasks and their context arise from continuing organizational work. From a natural-language description of an organization's purpose and activities, OfficeTown initializes an organization whose members coordinate work, delegate bounded assignments to assistant agents, and review their deliverables. Artifacts, decisions, and unresolved matters persist and shape subsequent work, while responsibilities and permissions determine what each member can see and do. We curate work requests, their context, and review feedback into executable resources for evaluation and learning. We construct OfficeTown Bench with 406 tasks and 4,836 evaluation criteria from 23 simulated organizations across eight professional domains, incorporating domain-expert feedback. Frontier-model evaluations reveal substantial remaining headroom. Supervised fine-tuning and on-policy distillation on OfficeTown data substantially improve Qwen3.5-9B on external professional-work benchmarks, including GDPval. Supplementary evaluations show that the principal--assistant harness achieves higher scores than the single- and multi-agent baselines tested. These results highlight the value of organizational simulation in generating work-grounded data and environments for the continued development of generalist agents.
cs.AI / 38 / 2610.23195
FuzzInsight: Navigating the Efficiency-Effectiveness Trade-off in Coverage-Guided DNN Fuzzing
Abstract
Collecting and labeling diverse test datasets for evaluating Deep Neural Networks (DNNs) is costly and time-consuming. This has motivated coverage-guided fuzzing techniques that generate defect-revealing corner cases from limited test data. However, their effectiveness depends heavily on configuration choices, particularly coverage criteria and seed-selection strategies. We present a configuration-oriented analysis and decision guidelines for scalable DNN fuzzing targeting two bottlenecks: inefficient seed utilization and the cost of coverage guidance. Our analysis shows that existing seed-selection strategies repeatedly prioritize a small subset of inputs, leaving many collected seeds unexplored. To address this issue, we propose Equalized Sampling, a constraint-based strategy that increases mean seed queue utilization to 77%, improving on prior approaches by 13 percentage points. We further evaluate 84 fuzzing configurations on four benchmark datasets: MNIST, SVHN, GTSRB, and CIFAR-100. Results show that the non-structural criteria LSCD, LSC, and MDSC achieve coverage outcomes comparable to structural and alternative non-structural criteria while requiring substantially less computation. These criteria also promote broad behavioral exploration of the DNN under test, reducing the need to monitor multiple coverage metrics. Among them, LSCD provides a modest but consistent runtime advantage. Finally, we show that retraining with generated corner cases improves generalization robustness by up to 77 percentage points on average without degrading performance on the original test distribution. Together, these findings provide a practical blueprint for selecting fuzzing configurations that balance efficiency, behavioral exploration, and model hardening.
cs.AI / 39 / 2610.23213
Not All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning
Abstract
Test-time scaling methods implicitly assume that additional computation from the same model provides independent evidence. We show that this assumption fails within a single reasoning model: errors are correlated across repeated samples, across reasoning depth, and in the model's own self-evaluation, so more compute from the same model is not more evidence. We introduce DIA (Disagreement-triggered Independent Arbitration), which treats disagreement between an intermediate-stage and the final-stage prediction as a per-question signal of reasoning instability and resolves it by asking a small panel of heterogeneous models to solve the question independently, without seeing either candidate, and aggregating their answers directly. Across seven benchmarks and two base models, DIA outperforms full chain-of-thought in 12 of 14 settings, including a 7.00% gain on MuSiQue, while activating on only $\sim 17\%$ of examples and using up to 47.2% fewer tokens than a multi-agent baseline. Ablations show that both independent solving and solver heterogeneity are necessary for these gains.
cs.AI / 40 / 2610.23256
Beyond Reward Scoring: Online Multi-Agent Preference Learning for Multimodal Response Generation
Abstract
Multimodal preference alignment often follows a score-based RLHF pipeline: a reward model is first trained to learn human preferences, then used to score on-policy responses and optimize the response model's generation. This turns preference supervision into a scalar signal for generation: a model may learn to produce answers with higher preference scores, but is rarely trained explicitly to learn the reasons underlying human preferences. We therefore ask whether making preference understanding an internal capability of the response model can improve its generation. To investigate this question, we propose \textbf{PrefMind}, an online dual-agent reinforcement learning framework with two trainable agents: a response Generator and a preference Critic. Given a pair of candidate responses, the agents analyze and discuss which response better aligns with human preference. Human labels supervise their judgments and corrections. For pairs on which both the Generator and revised Critic correctly identify the human-preferred response, the Generator produces an improved answer, and the Critic compares it against the original preferred response to provide a generation reward. Both agents are jointly updated from interactions sampled with their current policies. Through this process, \textbf{PrefMind} trains the Generator to understand human preferences via explicit judgment and correction, and to translate this understanding into better responses. Experiments across three model scales show improvements in both standalone preference judgment and downstream response generation.
cs.AI / 41 / 2610.23257
Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To
Abstract
Chain-of-thought (CoT) controllability, the ability of reasoning models to control properties of their CoT, is consequential for both their usefulness and reliable oversight. Recent work evaluates this capability by measuring whether models follow CoT constraints given in the user prompt, and reports very low controllability across models. We argue that this low controllability may not reflect an inherent inability of reasoning models to control their CoT, but instead result from naively specifying constraints in the user prompt. To test this, we introduce Self-Instructed CoT Controllability: a complementary evaluation that places the constraint as a self-instruction inside the model's CoT and measures its ability to follow it. We show that, under self-instruction, models achieve significantly higher controllability than with standard user instructions. Even a generic reminder to follow the user's constraint substantially improves controllability. Moreover, effective self-instructions can be bootstrapped into the reasoning through a simple instruction in the user prompt, enabling improved control even without access to the reasoning channel. Finally, using self-instruction, we demonstrate in a realistic monitoring setup that CoT controllability can be leveraged for monitor evasion, posing a risk to CoT monitorability.
cs.AI / 42 / 2610.23284
Annotrace: A Workbench for Step-Level Annotation of Agent Traces
Abstract
Every study of why LLM agents fail begins by labelling traces by hand, and every such study builds its own throwaway tool. Observability platforms record agent traces but offer no structured labelling; annotation tools label text but have no notion of a step, a delegation or a tool call, leaving it unclear what the smallest unit of execution to label should be. Annotrace reads six trace formats, including OpenTelemetry GenAI, LangSmith and Langfuse, and turns each run into an ordered, nested list of steps. The annotator writes their own labelling codebook; a frontier model proposes a starting set from a sample of their own traces, answers first, and is asked exactly the questions the human is asked, making annotation a joint human-AI process. Annotrace measures agreement between any two annotators, human or AI. On a multi-agent question-answering corpus it reproduces a previous manual effort's 235 annotation units exactly, and two models labelling the same traces agree 91% of the time at $κ= 0.82$.
cs.AI / 43 / 2610.23299
Optimist Deep Research: An Iterative, State-Centric Agent for Autonomous Literature Research
Abstract
Most retrieval-augmented systems treat literature search as a single retrieval step followed by summarization. Real literature research works differently: the investigator's understanding of the question changes as they read, new subquestions appear, and evidence is weighed, not just collected. We present Optimist Deep Research, an open-source agent that runs literature research as an iterative evidence-acquisition process over an explicit, mutable research state. A research question is broken into a tree of subquestions, each served by search intents that are translated into source-specific queries over arXiv, OpenAlex, Crossref, PubMed, Semantic Scholar and the web. Results are merged into a canonical, provenance-preserving document registry through a conservative four-tier deduplication hierarchy. The registry is expanded by citation snowballing until a saturation criterion is met, and then screened, ranked and allocated to subquestions round-robin. Documents are read in parallel into a structured evidence store of claims with support, location, type and strength. A critic reviews the state along eight dimensions and returns research actions rather than a binary verdict: it can add, split, rephrase, deprioritize or close subquestions, so the research plan is an output of the investigation as well as an input. The final report is written section by section from the evidence store, and every citation resolves to a retrieved document. We describe the design, the deterministic machinery that keeps a language-model pipeline auditable, and a case study that read 28 of 237 retrieved documents in about three and a half minutes. We also report weaknesses the case study exposed.
cs.AI / 44 / 2610.23306
SimOPD: Relational Knowledge Transfer for On-Policy Distillation
Abstract
A key challenge in on-policy distillation (OPD) is extracting effective supervision from teacher knowledge. Supervision from token probabilities may underutilize internal knowledge, while direct hidden-state matching forces the student to shift to the teacher's feature space and complicates cross-architecture transfer. Inspired by contrastive learning's use of relationships between instances, we propose \textbf{SimOPD} to transfer teacher internal knowledge through response similarities. Specifically, SimOPD uses paired memory banks of student-generated responses to extract supervision from the teacher's internal representations as similarity distributions. The student then matches these distributions within its own representation space, retaining representational freedom and removing the need for cross-model projection. Extensive experiments show that SimOPD improves the overall average over the strongest baselines by up to 2.6 and 3.1 percentage points in same- and cross-architecture settings.
cs.AI / 45 / 2610.23313
From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory
Abstract
Long-term agent memory requires recovering answer-relevant evidence dispersed across turns and sessions. Existing systems increasingly use structured memories and adaptive retrieval, yet often pass retrieved or expanded units directly to the reader, coupling memory addressing with evidence presentation and leaving the final context redundant or incomplete. To address this gap, we propose MemCompile, a query-conditioned evidence compilation framework for long-term conversational memory. Given a query, the Address Plane locates relevant support through compact source-linked descriptors and capacity-aware Cells, while a normalization boundary resolves heterogeneous retrieval hits to stable SourcePacket identities. Its Evidence Plane then allocates a bounded evidence budget across complementary packets, admits limited additional support, and renders the retained claims associated with each selected packet into an answer-ready context. MemCompile achieves 80.52% Judge Accuracy on LoCoMo and extends the empirical quality-cost Pareto frontier to a higher-quality regime. Its strong performance on LongMemEval-S further suggests that the benefits of query-conditioned compilation generalize across long-term conversational-memory benchmarks.
cs.AI / 46 / 2610.23317
AgentOnto: Ontology-Guided LLM Agents for Enterprise Data Analytics
Abstract
Enterprise data analysis requires agents to connect analytical requests with knowledge distributed across databases, business documents, and existing example tasks. Without an explicit representation of this knowledge, an agent may produce incorrect analytical results and cannot reliably incorporate feedback. We present AgentOnto, an ontology-guided LLM agent for enterprise data analytics. AgentOnto organizes heterogeneous enterprise knowledge into an ontology that directly guides analysis and incorporates confirmed feedback. Iterative Ontology Preparation further improves the ontology through self-generated analysis tasks. We evaluate AgentOnto on two public data sources. AgentOnto substantially outperforms representative methods, while module-level evaluations confirm the effectiveness of ontology construction, analysis, and updating. A real-world deployment further demonstrates its practical value in enterprise analytics. These results demonstrate that explicitly constructing, using, and updating enterprise knowledge supports reliable and adaptable data analysis.
cs.AI / 47 / 2610.23342
Robust Function Calling for LLM Agents Under Interface Drift
Abstract
Tool-augmented language-model agents rely on interface text: tool descriptions, argument schemas, and the candidate tool list shown at inference time. These interfaces often change after deployment even when the underlying tool semantics remain fixed, creating a practical reliability problem for function-calling systems. We study this failure mode as interface drift and evaluate whether a lightweight, training-free repair layer can recover robustness without retraining the base model. Our method, SchemaShield-Lite, combines schema validation, candidate-set retry for missing tool names, and one-shot constrained repair. On Qwen3.5-9B, it raises DICE-300 accuracy from 0.717 to 0.813 and BFCL-200 accuracy from 0.580 to 0.805, recovering 19/22 and 33/45 drift-induced failures with no observed repair harms. Naive retry does not explain the gains. Ablations show that candidate-set drift, especially distractor tools, accounts for most observed damage, while description and schema edits alone are less harmful. The results position validation-and-repair as a practical robustness layer for deployed LLM-agent tool use.
cs.AI / 48 / 2610.23345
OliO: ODE-based Linear Transition Operator for Self-Supervised Time Series Forecasting
Abstract
Self-supervised representation learning has advanced time series forecasting by capturing robust features from unlabeled data. However, existing contrastive and masking-based methods often struggle to explicitly model the underlying temporal flow, either by treating time lag as noise or by disrupting the inherent temporal dynamics essential for forecasting. To address these limitations, we propose OliO (**O**DE-based **Li**near Transition **O**perator), a plug-and-play self-supervised learning method that learns bidirectional temporal flows as a continuous and structurally consistent dynamical system. OliO introduces a transition operator derived from linear Ordinary Differential Equations (ODEs) to align latent representations across arbitrary time shifts. By imposing a strictly upper triangular constraint on the transition matrix, we mathematically ensure numerical stability and enforce a robust polynomial boundary that effectively captures long-term dependencies while preventing the exponential divergence typically found in unconstrained ODE-based models. This structural constraint induces a temporally coherent representation space that preserves the underlying flow essential for accurate forecasting. Extensive experiments across various backbone architectures and benchmarks demonstrate that OliO achieves significant performance gains, providing Mean Squared Error (MSE) reductions of up to 5.75% over the most competitive grid-searched SOTA baselines. The code is available at *https://github.com/eai-lab/OliO*
cs.AI / 49 / 2610.23362
Monitor-Validity Failures: LLM Safety Evidence Paths at Evolving Framework Boundaries
Abstract
Modern LLM applications rely on runtime monitors that inspect tool calls, traces, streams, detector outputs, and verifier inputs. This creates an interface-reliability problem: a framework path may remain operational while policy-relevant evidence never reaches its decision point. In one representative case, a regex detector returned a positive match, but the framework stranded the result before the block decision. This differs from jailbreaks and detector errors because the evidence is already recognized yet disconnected from enforcement. We study this repair-boundary signature with canary contracts: parent/fixed confirmation units that place a known payload only on the changed evidence surface and record monitor visibility together with the downstream decision for action-boundary cases. Applying the protocol to public framework repairs, we confirm 18 evidence-path failures across 11 LLM or agent safety project groups and four implementation ecosystems. Nine are established-boundary repairs, eight are new-surface policy-coverage gaps, and one is a boundary-only evidence case. Fourteen reach E1 enforcement boundaries: seven native action outcomes, six configured enforcement boundaries, and one synthetic-consumer consequence. The remaining cases affect verifier, safety-prompt, or observability channels. The cases form carrier, adapter, authority, and surface families and show how ordinary application behavior can remain intact while a safety-specific evidence path misses its decision. Exact revisions, deterministic outcomes, negative controls, and a frozen prospective audit document the confirmation rule. Monitor assumptions therefore need regression tests at the schemas, streams, tools, and adapters that frameworks evolve.
cs.AI / 50 / 2610.23436
Generative Policy Improvement from Observational Data
Abstract
We study how to improve decision policies from observational data to generate better actions when actions are complex, structured objects, such as text messages or images. This task is subject to two major challenges: (i) action-level overlap is often too weak for reliable off-policy evaluation and improvement, and (ii) generating new structured actions often requires expensive training or fine-tuning (e.g., of an LLM) and can be ill-defined due to information loss in action representations. To address both, we propose PolicyLift, a three-stage framework for generative policy improvement with structured actions from observational data. In Stage 1, PolicyLift learns a small set of discrete action codes that capture high-level strategies in the observed action space and are tailored to support policy improvement under limited overlap. In Stage 2, it learns a context-dependent policy over these action codes. At deployment time, Stage 3 "lifts" the learned code-level policy back to the original action space by generating actions from the historical behavior system in a code-conditioned way. We show theoretically that improving the policy over action codes guarantees improvement of the induced policy in the original action space. We demonstrate PolicyLift across experiments in synthetic and real-world text generation tasks with consistent policy improvements over the behavior policy.
cs.AI / 51 / 2610.23443
ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals
Abstract
Efficient expert routing is a fundamental challenge in Mixture-of-Experts (MoE) systems, particularly for heterogeneous collections of independently trained Large Language Mod els (LLMs). Existing routing approaches typically rely on jointly trained gating networks, prompt-based predictors, or computationally expensive multi-model inference, limiting their flexibility and efficiency. This paper presents ScoutRoute, a lightweight post-hoc routing framework that predicts expert competence using internal language model signals extracted from a single forward pass through an inexpensive scout model. Hidden-state representations are used to estimate cross-model competence through lightweight logistic probes, while logit-derived confidence measures are incorporated within a domain-aware routing strategy for efficient expert selec tion without modifying the underlying language models. ScoutRoute isevaluatedon fiveopen-weightlanguagemodels across fourbenchmark datasets covering mathematical reasoning, code generation, commonsense reasoning, and multidis ciplinary knowledge. Experimental results show that ScoutRoute consistently outperforms individual routing signals while requiring only a single scout-model inference. The proposed competence probes also achieve superior calibration compared with conventional confidence measures, demonstrating that internal language model representations provide transferable information for accurate and efficient expert routing across heterogeneous language models
cs.AI / 52 / 2610.23449
GatewayBench: Unified Capability and Security Evaluation Platform for MCP Gateways
Abstract
AI agents have increasingly interacted with external environments by invoking tools, with Model Context Protocol (MCP) emerging as a standard interface for connecting agents to heterogeneous tools and services. As organizations deploy agents at scale, MCP gateways have become a critical middleware layer for aggregating, governing, and shaping the tool surface exposed to agents. Despite the growing importance of MCP gateways in agentic systems, no prior work has systematically studied how gateway design affects agent capability, efficiency, and security. To address this gap, we introduce GatewayBench, the first unified evaluation framework and benchmark for MCP gateways. We characterize gateway architectures into two major paradigms: flattened gateways and hierarchical gateways. GatewayBench comprises 524 tasks across six domains, together with a pool of 11,118 real-world tools from 251 authentic MCP servers. To emulate large-scale deployments, we systematically inject task-irrelevant tools sampled from this pool into each agent environment, enabling controlled evaluation as the accessible tool surface scales. Based on GatewayBench, we conduct extensive experiments across seven widely deployed MCP gateways and diverse agent configurations, covering both benign and adversarial tasks. Our results reveal substantial differences across gateway designs in task completion, tool-use efficiency, and resilience to attacks, and further demonstrate that these effects interact strongly with the underlying model and agent harness. These findings highlight the MCP gateway as a consequential yet underexplored component of the agentic AI stack. By enabling controlled evaluation of gateway architectures under realistic tool ecosystems, GatewayBench provides both practical guidance for MCP gateway design and a rigorous testbed for developing more capable, efficient, and secure agentic systems.
cs.AI / 53 / 2610.23598
RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory
Abstract
Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory. Existing methods focus on lexical or semantic similarity, leading to target memories being crowded out by temporally mismatched yet semantically similar candidates. Even time-aware approaches inject temporal cues only during memory organization or answer generation, overlooking temporal constraints at the retrieval stage, where candidate selection occurs - the root cause of the displacement. Inspired by the temporal contiguity effect in human episodic memory, we propose TIDER (Temporal Interval-driven Dual-channel Evidence Retrieval), a time-aware agentic retrieval framework. Central to TIDER is a reinforcement learning (RL)-trained temporal interval locator that infers the temporal windows in which the evidence lies. RL optimizes window localization as a discrete, non-differentiable retrieval decision, using composite rewards to convert sparse answer supervision into fine-grained retrieval feedback. These temporal windows first define the temporal retrieval channel; the retrieved candidates are then merged with results from the semantic channel and reranked by proximity to the windows. This temporal grounding enables TIDER to retrieve temporally aligned evidence that would otherwise be drowned out by the volume of semantically similar yet temporally mismatched memories. With a lightweight 1.7B locator, TIDER achieves state-of-the-art accuracy on both long-horizon memory benchmarks - 61.52% on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%) - which indicates that temporal cues are crucial for pinpointing target evidence amid a vast pool of semantically similar historical events in long-term daily-life memory.
cs.AI / 54 / 2610.23617
GEMS: Molecular Structure Elucidation as Learned Search over Isomer Graphs
Abstract
Identifying small-molecule structures from molecular formulas and tandem mass spectra remains a bottleneck in untargeted metabolomics and natural-product discovery. Even when database retrieval or de novo prediction misses the exact structure, the resulting candidates can provide starting points for exploring nearby structures through a few bond edits. We introduce GEMS (Guided Editing of Molecular Structures), which refines candidates through learned 2-switch bond edits that preserve the molecular formula and each atom's total valence. GEMS uses MIST, a pretrained spectrum-to-fingerprint model, to predict fingerprint probabilities once per spectrum. A neural policy compares each candidate's fingerprint with these predictions, links present features to the atoms and bonds that generate them, and scores all valid edits in one forward pass. The policy is trained on shortest edit paths between structures, so it can be pretrained on PubChem isomers without spectra and then fine-tuned with spectrum-predicted fingerprints. At inference, GEMS searches from multiple initial candidates matching the supplied molecular formula and ranks candidates by fingerprint agreement and multi-start consensus. With the target structure excluded from PubChem initialization, three edit rounds achieve Hit@1/Hit@10 of 29.64%/50.81% on NPLIB1 and 30.22%/41.33% on MassSpecGym, exceeding published baselines under comparable settings. On NPLIB1, GEMS achieves higher identification accuracy in about 3.3 minutes end to end, compared with 26.2 hours for our FRIGID run. GEMS also improves identification from generated candidates, including those refined by FRIGID. Experiments with exact target fingerprints identify fingerprint prediction error as a bottleneck in target recovery and final ranking. Deeper search recovers additional targets without retraining, with larger gains from weaker initial candidates.
cs.AI / 55 / 2610.23670
OwlSight: A Panoramic Monitor for Chain-of-Thought Reasoning
Abstract
Chain-of-Thought (CoT) monitoring offers actionable evidence for oversight AI systems, with the potential to enable early detection and intervention. However, existing approaches are focused on specific tasks, whereas the regulation of AI, particularly reasoning models, requires monitoring a broader range of tasks. We introduce Factor for CoT monitoring, a strategy that supports customizing monitoring requirements in the prompt, which enables a single monitor to cover different tasks by relying on different factors. Furthermore, we introduce two types of factors, fine-grained and coarse-grained (e.g., monitoring whether a specific email address has been leaked versus monitoring privacy leakage), to accommodate diverse user monitoring needs. Based on this, we develop a CoT monitor named OwlSight that relies solely on factor to monitor behaviors related to multiple dimensions, including capability, safety, and trustworthiness. Experiments across 15,971 reasoning trajectories show that OwlSight surpasses general-purpose LLMs by 1.32-30.57 and 11.71-27.81 points at fine and coarse granularity, respectively, and achieves state-of-the-art results on OpenAI's CoT monitorability benchmark. It also generalizes to unseen models' CoTs and remains robust to Factor rephrasings, CoT compression, and obfuscation, offering a promising paradigm for CoT monitoring that can contribute to the oversight of frontier AI.
cs.AI / 56 / 2610.23791
DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery
Abstract
Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction. Its environment integrates over 80 scientific operations, biomedical resources and task-specific code, preserving associations among candidates, results and artifacts. From execution trajectories and outcome feedback, DrugHarness induces action hypotheses with observable applicability conditions, recommended actions and supporting evidence. An independent review agent evaluates candidate skills derived from these hypotheses through source verification and, when needed, targeted execution trials. Admitted skills guide subsequent tasks and are revised through feedback. Evaluation on MolBench and SMDD-Bench demonstrates capabilities in molecular screening, optimization and constrained design. Without evolution, DrugHarness exceeds reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. Retrospective evaluation on skill-development tasks shows that evolution increases Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%. In a prospective HPK1 study, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates. A synthesized compound achieves a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference, with activity competitive with two selected compounds designed by experienced medicinal chemists. These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation.
cs.AI / 57 / 2610.23796
FuTransformer: FP8 prefill and INT4 decode for Llama-style inference on Hopper
Abstract
Transformer prefill reuses weights across tokens and is dominated by matrix multiplication and activation data movement; single-token decode has limited reuse and is dominated by weight and KV-cache reads. We present FUTRANSFORMER, a CUDA implementation of a Llama-style Transformer block for NVIDIA Hopper GPUs. It applies CODA's GEMM-plus-epilogue approach to inference, adding direct KV-cache placement, folding the fixed RMSNorm gain into weights, writing FP8 output directly with delayed scaling, and supporting prefill without padding. Decode uses INT4 projection weights and an INT4 KV-cache copy. Weights and keys follow Tensor Core fragment layouts for direct dequantization into MMA operands without full intermediate tiles. A block-diagonal Hadamard transform reduces cache quantization error from outlier channels. On one H100, prefill is up to 1.48x faster than compiled PyTorch FP8 and reaches 1.29 PFLOP/s. For gate/up projection with SwiGLU at the 8B shape, our BF16 path is 1.13x faster than CODA. Our FP8 path is 2.02-2.18x faster than the evaluated BF16 CODA path. Across 8B and 70B shapes at batches 1-32, INT4 decode is 1.3-2.4x faster than the faster of PyTorch's BF16 and INT4 blocks at each point: 2.1-3.3x faster than compiled, CUDA-graphed BF16 and 1.3-6.3x faster than the same block using PyTorch's INT4 projection kernel, which targets a single query row and loses to BF16 at batch 32. It is also 31-51% faster than our FP8 path. Complete Llama-3.1-8B decode takes 2.8 ms per step at batch one, versus 8.1 ms for the compiled transformers implementation. Decoding every WikiText-2 token through the INT4 weight path raises perplexity from 6.24 to 7.23 with round-to-nearest weights or 6.99 with GPTQ in the same format; the rotated INT4 cache adds 0.20. The kernels match a quantized reference computation within 0.001 perplexity.
cs.AI / 58 / 2610.23801
Beyond SAEs: Consistent Concept Discovery with Sparse Coding
Abstract
Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based analyses. We propose Consistent Sparse Dictionary Learning (ConSDL), which replaces the standard linear encoder with per-sample sparse coding. This is motivated by our synthetic analysis showing that coefficients from a single linear encoder pass are suboptimal compared with those obtained by per-sample optimization, suggesting that sparse coding is a reliable mechanism for dictionary learning. Across 16 synthetic settings, ConSDL achieves near-perfect dictionary and coefficient recovery. These gains translate to state-of-the-art cross-seed consistency across four vision and language settings, where ConSDL improves dictionary and coefficient consistency by 17.6% and 35.9%, respectively, over the average of six SAE baselines.
cs.AI / 59 / 2610.23826
Self-Building Labs
Abstract
Self-driving laboratories (SDLs) accelerate scientific discovery by automating experimental execution and optimization. AI agents can increasingly propose new experiments, but existing SDLs can only run the procedures they were built for. Adapting them to a new experiment still takes extensive manual engineering, which makes laboratory flexibility the bottleneck for scaling discovery. We introduce the self-building laboratory (SBL), which designs, configures, and operates its own laboratory setup given a high-level scientific goal. SBL starts from whatever already exists, from an empty room to a working SDL, and builds the missing capabilities: the lab layout, its digital twin, robot skills, and experimental workflows. We demonstrate the system working across different robots and setups, scaling from a single workstation to laboratories with several robots working together. We validate SBL's end-to-end capabilities through a real-world quantum-dot synthesis experiment. By shifting the role of human researchers from manually configuring and programming device automation to specifying scientific objectives, SBL makes scientific automation more accessible, flexible, and scalable. Project website: https://self-building-labs.github.io/
cs.AI / 60 / 2610.23845
AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery
Abstract
Reproducible and analytically robust research findings are essential to scientific progress, but evaluating these properties typically requires costly manual effort. We investigate whether LLM agents can help automate reproduction and robustness analysis, using a medical case study grounded in the Surveillance, Epidemiology, and End Results (SEER) database, a widely used U.S. cancer registry underlying tens of thousands of published studies. First, we introduce SEER-Repro, a benchmark of 20 published SEER-based papers together with the data and resources required to reproduce their findings. Second, we create an AI agent that successfully reproduces 77.4% of 2,190 atomic claims, while providing executable baselines for further analysis. Third, we develop an agentic analysis pipeline to test the robustness of the medical findings by identifying analytical choices underlying each finding, proposing defensible alternatives, and rerunning the analysis to determine whether those choices affect the conclusion. Across published papers in SEER-Repro, our pipeline identifies potentially consequential analytical choices in 9.5% of fully reproduced findings and flags them for expert review. These results show how LLM agents can support scientific review by reproducing findings, systematically testing their robustness, and directing expert attention toward potentially consequential analytical choices.
cs.AI / 61 / 2610.22606
Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition
Abstract
FOR-species20K is the largest public LiDAR benchmark for tree-species classification, pooling 25 independently collected terrestrial, mobile, and airborne laser-scanning datasets across 33 species. We introduce a multi-resolution, multi-view ConvNeXtV2 model that improves official hidden-test accuracy from 79.5% to 85.0%, ranking first at the time of submission. We then examine what this score measures. Because the hidden test is species-balanced, its overall accuracy is more comparable to a macro metric than to accuracy under the public set's long-tailed distribution. Moreover, the official random split places every source dataset in both training and evaluation. A logistic probe recovers source identity from out-of-fold model probabilities with 69% accuracy, while geometric features identify the source above 90% within two widely shared species; 18 of the 33 species occur in only one source. Under a source-species-cell holdout, evaluated on 5,566 trees from eight eligible species that retain the same LiDAR-type training support in other sources, the unweighted species mean falls from 82% to 36%, and tree-weighted accuracy falls from 95% to 60%. Count-matched retraining and experiments with two alternative backbones indicate that this gap is not explained by reduced training support alone, while an external PureForest experiment shows a smaller but consistent grouped-split effect. These results establish the 85% leaderboard score as strong in-distribution performance, but not as an estimate of source-disjoint generalization, motivating balanced metrics and source-aware evaluation. The code is available at: https://github.com/JaeLee18/FOR20K-lidar-tree-species
cs.AI / 62 / 2610.22795
STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases
Abstract
In autonomous driving, perceiving surrounding objects is not enough: failing to understand how they move and interact over time can lead to dangerous decisions. Evaluating vision-language models (VLMs) therefore requires testing whether they connect past spatial observations to future consequences. Yet existing benchmarks often assess edge cases through basic perception questions or test temporal understanding in simple scenarios, leaving complex spatiotemporal reasoning underexplored. We introduce STRIDE (SpatioTemporal Reasoning In Driving Edge Cases), an object-centric benchmark for long-tail, safety-critical driving scenarios. STRIDE covers six task families: spatial perception, spatial understanding, temporal memory, temporal extrapolation, trajectory prediction, and scene-context awareness. It pairs multiple-choice questions probing individual spatiotemporal judgments with related free-form visual question answering (VQA) requiring detailed scene explanations and justifications of driving decisions. We prioritize scenes with unusual object behaviors and complex traffic interactions, generate reference answers from structured dataset annotations through automated procedures, and use limited language-model assistance and human verification to ensure annotation quality. Across 12 proprietary, open-weight, and driving-specialized models, even the best performer, GPT-6-Astra, achieves only 43.4% average multiple-choice accuracy, versus an approximately 20% random-choice baseline. Further analysis reveals that strong performance in describing driving scenes does not necessarily translate to accurate safety-critical judgments, and that models struggle more with reconstructing object motion than recalling past positions. STRIDE thus exposes spatiotemporal driving limitations that aggregate VQA scores can conceal.
cs.AI / 63 / 2610.22895
HAPACT: A Benchmark For Human-Centric Physical Impact Localization in Movies
Abstract
Physical reasoning about humans in videos has largely focused on everyday interactions, where contact, pressure, pose, and motion support embodied perception and contact-aware applications. Still, short and forceful properties like impact events remain relatively underexplored, despite their relevance to broader understanding of how abrupt physical events affect the human body. To fill this gap, we introduce human-centric impact localization, a new task that localizes target physical impacts temporally within a video and spatially on the human body. We further present HAPACT, a HumAn-centric imPACT dataset and benchmark built from movie clips, containing over 2.8K video-query pairs with more than 20K fine-grained human-annotated frames. We built HAPACT around viewer-grounded annotation, text-conditioned localization, and supervision at frame-level temporal and SMPL-X-based spatial granularity. To demonstrate the effectiveness of our benchmark, we conduct extensive evaluations of existing methods, employing video moment retrieval models for temporal localization and human-scene contact models for spatial localization. For the spatial task, we propose a text-conditioned baseline by reformulating localization as query-guided body-region selection. Benchmark results show that physical impact localization remains challenging for existing methods, motivating HAPACT as a dedicated benchmark to support physical impact reasoning in videos.
cs.AI / 64 / 2610.22907
COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning
Abstract
Vision-language models (VLMs) such as CLIP retain zero-shot recognition ability, but their open-world use often requires task-agnostic continual learning: tasks arrive sequentially, past data are unavailable, task identity is unknown, and the backbone cannot be retrained. We argue that this setting fails VLMs through three coupled pressures beyond the stability-plasticity tradeoff: (1) modal-consistency drift, (2) representation-space interference, and (3) per-task capacity shortfall. We propose COMET (Collaborative Mixture-of-Expert Transfer) to resolve these three issues respectively as follows. A frozen teacher provides only a feature reference on a shared image-text pool; Tri-Affinity Distillation preserves image-image, text-text, and image-text geometry; and Student-Aware Cross-Expert Distillation aligns new experts with compatible prior experts so routing errors degrade gracefully. We design COMET to keep capacity and inference separate: a contextual bandit keeps or merges full per-task experts using post-training signals, while a reconstruction router selects sparse experts at test time without teacher semantics. On the X-TAIL benchmark, we show that each proposed component improves continual learning performance. COMET shows that continual vision-language adaptation can preserve multimodal geometry, coordinate accumulated experts, and allocate capacity jointly when feature reference, expert shaping, routing, and capacity control are separated by construction.
cs.AI / 65 / 2610.22637
Hard Instances on Demand
Abstract
The conclusions of an experimental study of algorithms are only as good as the instances they are run on. Benchmark libraries for well-studied problems took decades to curate, while evaluations of new problems often fall back on random instances, which are known to be misleading: they may say more about the random generator than about the algorithmic question at hand. We treat instance generation as an optimization problem: parameterize the instance space, take the solver's running time as the objective, and let a black-box optimizer search for hard instances. We develop a widely applicable workflow built on the Nevergrad optimization platform and demonstrate it on four NP-hard problems: satisfiability, traveling salesperson, Hamiltonian cycle, and non-permutation flow-shop scheduling. The search needs only the solver, an instance encoder, and a compute budget. Our evolved instances beat the hardest instance of a budget-matched random draw in all four case studies, by an order of magnitude or more in three. They are often competitive with long-curated benchmark sets: our SAT champion is harder (for a three-solver portfolio) than 78% of the 498 SAT Competition formulas in its size class; our 20-city TSP champions take several times longer than the hardest measured classical TSPLIB instance (229 cities); and our flow-shop champion takes 238.6 seconds of CP Optimizer time to solve to proven optimality (versus 4.5 seconds for the hardest of 100,000 random draws) while denying CP-SAT an optimality proof. The optimizer even recovers known structure from running time alone: it locates a SAT clause/variable ratio near the phase transition and finds structurally different champions for the two scheduling engines.
cs.AI / 66 / 2610.22882
Cooperation Without Adaptation: LLM Agents in a Collective-Risk Dilemma
Abstract
Language-model agents are increasingly deployed to act on behalf of people in settings where individual decisions shape shared outcomes. We examine whether five production models adapt their contributions to changing risks and partner behaviour, or instead follow a relatively fixed strategy. In a collective-risk dilemma game (CRG), six players contribute to a shared pool over ten rounds; if the pool falls short of its target, a catastrophe may destroy their remaining resources. Four models contributed approximately an equal share at every positive risk level, and four of the five also contributed when risk was zero, even though contributing could not improve their payoff. Explicitly stating that only the agent's own payoff mattered reduced zero-risk contributions for most models, while a more neutral prompt did not. Several models correctly identified which action offered the higher expected payoff at different risk levels, yet their contributions changed little. Against fixed partners, no model consistently chose the payoff-maximising response, and in mixed-model groups, a single agent's shortfall in the final round could cause an otherwise successful group to fail. Payoff-based selection also depended on population composition: in a single six-agent group, it favoured the model that contributed least and reduced group success, whereas in larger, randomly mixed populations, it favoured an exact fair-share model at moderate and high risk. These findings show that cooperative self-play alone is not evidence of strategic adaptability. Consistent with a machine-behaviour approach, evaluations of agents in multi-agent systems should test how they respond to incentives, partner actions, and population composition.
cs.AI / 67 / 2610.22990
PersonaCUABench: Benchmarking Computer-Use Agents Serving Diverse Personas
Abstract
Computer-use agents (CUAs) are commonly evaluated against fixed task specifications, but users may omit requirements, add constraints, or require approval before consequential actions. Completing the benchmark instruction does not guarantee serving the person who asked. We introduce PersonaCUABench, which holds 74 tasks across 11 domains fixed and varies only who asks. A persona agent, playing one of 44 personas in four groups, rewrites each instruction in that user's own words and answers the CUA's questions, while the CUA sees only the dialogue, never the profile or the original instruction. Across nine backbones (3,330 runs), mean task completion falls by 36% relative to the original requests. The losses have two sources: in 59.9% of cells that lose points, the agent engages with what the user asked for and still gets it wrong, and 11.2% are stops the user requested that the checker scores as failures. Even when a persona request is fully completed, the user gives the top satisfaction rating in only 62.7% of cells, against 84.1% for the original requests. CUAs should therefore be evaluated not only on task completion but also on whether they meet the requirements and constraints users express. Code: https://github.com/MatrAIx-ai/PersonaCUABench. Project page: https://matraix.ai/research/persona-cua.html
cs.AI / 68 / 2610.22703
NoveltyScope: A Multi-Agent Research Assistant for Evidence-Grounded Novelty Assessment of Scientific Papers
Abstract
NoveltyScope is a prototype that compares a scientific paper with retrieved prior work and produces a structured report and a five-level novelty score. The system extracts the core task and at most two author-claimed contributions, then searches a closed local corpus using BM25 and dense retrieval with reciprocal rank fusion. Contribution comparisons request paired passages for potential refutations, and quotations are checked against cached source text. Four scoring roles (Prosecutor, Defender, Judge, and Critic) use the resulting evidence, with at most one revision of the Judge's score. Evaluation uses 50 class-balanced RINoBench samples whose idea-level labels are retained for full-paper input. Mean absolute error is 1.48 and 1.26 for single-LLM and multi-agent scoring with Gemini, respectively, and 1.22 and 1.16 with Qwen3. The vector-similarity baseline has an error of 1.68. A HiCLIP case study illustrates the contribution comparisons and paired evidence passages. Overall, the results demonstrate the feasibility of evidence-grounded novelty assessment and show that the multi-agent configurations achieve lower scoring errors than their single-LLM counterparts across both evaluated models.
cs.AI / 69 / 2610.23470
QD-Rec: Open-Ended Evolution of Recommendations via Quality-Diversity
Abstract
Open-ended search combined with foundation models has produced advances such as faster algorithms, new results on open mathematical problems, and self-improving agents. We extend this paradigm to recommendation for decision-making, focusing on decisions where the set of possible choices is fixed but the reasons for choosing them are not. In high-stakes decisions, e.g., which company to acquire or which customer to target, comparing diverse, well-reasoned options can be more useful than a single ranked list. We formalize open-ended recommendation as building a diverse, high-quality set of recommendations, each an item together with a hypothesis, a natural-language reason why the item fits the needs of a decision maker. We propose QD-Rec, a quality--diversity method that selects items with a contextual bandit and generates hypotheses with a foundation model from earlier ones kept in an archive as stepping stones. A new recommendation enters the archive if it is novel or better than its nearest neighbor. Across M&A, customer targeting, and paper and book recommendation, QD-Rec gives the highest mean quality and builds the largest archive in most settings.
cs.AI / 70 / 2610.23252
Preventing Data Hallucinations in Agentic AI for Optical Network Troubleshooting
Abstract
An agentic AI system for automated troubleshooting in optical communication networks is demonstrated that integrates data references into the agent architecture to prevent data hallucinations, improving the interpretability and trustworthiness of root-cause identification and reasoning.
cs.AI / 71 / 2610.22852
Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors
Abstract
Generalist robot policies offer strong behavioral priors, but achieving deployment-level reliability and throughput remains difficult. Reinforcement learning (RL) can improve these policies, yet its effectiveness is constrained by the cost of online interaction. Existing methods adapt frozen flow-matching policies by treating noise space as a continuous action space. We show that this behaviorally unstructured space makes exploration inefficient, often preventing these methods from extracting meaningful learning signals on challenging tasks. We introduce AnchorQ, which enables coarse-to-fine policy improvement through a structured, refinable latent action representation. At each decision step, AnchorQ uses flow reversal to map reference behaviors to state-conditioned latent anchors, samples local particles around those anchors, and applies value-guided sampling to explore promising refinements. Our analysis shows that anchors reduce identification complexity at the cost of support bias; particle refinement trades this bias against action-class complexity; and critic accuracy determines the reliability of KL-regularized, value-guided selection. Across simulated and real-world manipulation tasks, AnchorQ outperforms state-of-the-art post-training baselines in sample efficiency, delivering a 37% improvement in real-world success rate and 97% higher simulated task throughput under the same interaction budget. Additional visualizations are available on our website https://anchorq-rl.com/.
cs.AI / 72 / 2610.23053
Unifying Embodied Foundation Models via a Model-to-Policy Formulation
Abstract
Vision-language-action (VLA) models and world-action models (WAMs) have emerged as promising approaches to robot learning and control. These models integrate multimodal perception and action generation, but differ in their architectures, output representations, and learning objectives. In this paper, we present a unified, architecture-agnostic probabilistic formulation for embodied foundation models based on stochastic kernels. Our formulation has two key components: (i) we characterize embodied foundation models through conditional output distributions and action-realization kernels, which together induce task-conditioned robot policies; and (ii) we connect pretraining and post-training to policy deployment, accommodating model-specific pretraining objectives and offline or online post-training strategies. We instantiate the formulation for OpenVLA and DreamZero, covering autoregressive action-token prediction and joint video-action generation.
cs.AI / 73 / 2610.23550
Reason-guided Human Reward Modeling for Preference-based Reinforcement Learning
Abstract
Reinforcement learning (RL) policies optimized for task rewards may produce behaviors that are efficient but misaligned with human expectations, safety preferences, or intent. This paper presents a reason-guided preference-based reinforcement learning (PbRL) framework that learns a human reward model from trajectory-level feedback. In addition to preference labels, the framework incorporates reason tags explaining why a behavior is preferred, enabling the learned reward to capture both human choices and their underlying rationale. Compared with preference-only reward learning, the reason-guided formulation provides additional supervision for learning human-aligned rewards and improves the interpretability of the resulting policy. The learned human reward is combined with the environment reward during policy optimization, allowing the agent to improve task performance while producing more interpretable, human- aligned, and preference-sensitive behavior
cs.AI / 74 / 2610.23615
Beyond Action Prediction: Learning Search-Free Planning in Latent Space
Abstract
Latent world models learn action-conditioned state transitions in a compact representation space, providing a predictive model for visual planning and control. However, planning with these models typically requires costly test-time optimization over action sequences. We propose to replace this online search with a goal-conditioned inverse model that directly predicts actions from current and goal observations. A key challenge is that action prediction alone neither ensures that goal representations encode action-relevant information nor that predicted actions are consistent with the learned latent prediction model. We introduce Hindsight Inverse JEPA Planning (HIJP), a search-free framework that addresses both aspects during training. HIJP relabels future observations along offline trajectories as goals, providing action supervision across multiple temporal horizons. By retaining gradients through the goal representations, this supervision directly shapes the visual representation. HIJP further feeds predicted actions into the latent predictor model and aligns the resulting next-state representations with encoded successor observations, enforcing action-transition consistency. At inference time, the inverse dynamics model and latent predictor models are applied alternately to generate multi-step action sequences without online search. Across four visual control tasks, HIJP achieves 86.8% overall success and a 1.9-16.5x planning speedup over search-based planning.
cs.AI / 75 / 2610.22665
Diverse Reasoning Paths Matter: Symbolic Equivalence Augmentation for Enhancing RL Exploration
Abstract
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for improving the reasoning capabilities of LLMs. However, the role of supervised fine-tuning (SFT) data in enabling effective downstream RL remains underexplored. In this paper, we investigate a key hypothesis: the diversity of reasoning steps in chain-of-thought (CoT) during SFT is a critical enabler of RLVR performance. To test this hypothesis, we introduce EquivAug, a data augmentation framework that uses symbolic equivalence transformations to generate diverse yet provably correct reasoning paths. By converting CoT solutions into formal symbolic representations, applying equivalence transformations, and reconstructing natural-language explanations, EquivAug produces training data that differs in reasoning steps rather than surface-level lexical variation. Experiments on GSM8K and FOLIO across model scales demonstrate that EquivAug consistently achieves the highest reasoning path diversity after SFT and the strongest accuracy after GRPO. Our results establish that CoT diversity during SFT is positively correlated with downstream RLVR performance.
cs.AI / 76 / 2610.23196
Framing Spatial Hearing within the Free Energy Principle: A Hierarchical Model of Perceptual Learning
Abstract
Spatial hearing relies on dynamic internal representations to decode spatial information from ever-changing acoustic conditions. A fundamental example is sound localisation, where listeners infer the direction of an invisible sound. Bayesian inference provides a computational account of such a localisation process; however, modelling efforts are largely limited to static listening scenarios. Instead, the free energy principle (FEP) describes perception, learning, and action as joint minimisation of variational free energy (VFE), providing a principled framework to account for dynamic conditions. Here, we take a first step toward applying the FEP to spatial hearing. We characterise how listeners' beliefs evolve across trials, in which they estimated sound direction and received subsequent visual feedback on each trial. Listeners are described by a hierarchical generative model where visual feedback drives gradient descent on VFE, updating beliefs over three independent quantities: lateral and polar localisation precisions and hemifield disambiguation. The model captures individual learning trajectories and reveals components with different dynamics: lateral precision improves, as does hemifield disambiguation, consistent with increased sensitivity of the auditory pathway in decoding vertical direction. Notably, polar precision also increases, most for listeners with the lowest initial precision, suggesting that the observed vertical improvement is driven by both hemifield disambiguation and polar precision. This work establishes spatial hearing as a tractable domain for FEP and motivates future closed-loop experiments with full action-perception cycles.
机器学习 (cs.LG)
144
cs.LG / 1 / 2610.23783
Sampling the XY model with an attention-based neural network
Abstract
We present a transformer-based variational autoregressive network for sampling lattice systems with continuous degrees of freedom and apply it to the two-dimensional XY model near the Berezinskii-Kosterlitz-Thouless transition. The architecture is combined with physics-informed ingredients: locality-informed attention and approximate probabilities. The resulting sampler significantly outperforms existing neural samplers. We train the model on lattices up to $96\times96$, verify with a mode-dropping estimator that no parts of the configuration space are missed, and reproduce the energy, magnetization and helicity modulus obtained with the traditional Markov Chain Monte Carlo algorithm.
cs.LG / 2 / 2610.23643
Scalable long-range machine learning interatomic potentials with electrical response for atomistic material modeling
Abstract
Accurate atomistic simulations of materials under electric fields require long-range electrostatics and electrical-response modeling at length scales beyond the reach of conventional first-principles methods. Yet these capabilities remain difficult to reconcile with the scalability of machine-learning interatomic potentials (MLIPs). Strictly local models enable efficient spatial parallelization but neglect long-range interactions, whereas existing electrostatic extensions introduce global dependencies that hinder distributed computation. Here, we overcome this limitation by analytically separating the long-range electrostatic forces and Born effective charges (BECs) into a global electrostatics calculation and strictly local derivatives of learned latent charges. This formulation preserves the spatial decomposability of equivariant MLIPs while enabling distributed evaluation of energies, forces, stresses, and BECs across different boundary conditions. Implemented as DPA4C-LR on the \texttt{nvalchemi} platform, the framework scales to 5.2 million atoms on four GPUs at approximately twice the computational cost of its short-range backbone. Trained on energies and forces, DPA4C-LR accurately infers BECs and reproduces the field-dependent infrared spectra of water. We further demonstrate its capabilities by recovering the characteristic slim polarization hysteresis of relaxor \ce{Pb(Mg_{1/3}Nb_{2/3})O3}, which is obscured by finite-size effects in smaller simulations; predicting a room-temperature field-driven ionic conductivity of $6.8 \pm 0.7$~mS/cm in amorphous \ce{Li_{0.9}NbO_{0.9}Cl_{4.1}}; and modeling interfacial thermal transport across Si/diamond interfaces with slab-corrected electrostatics. By combining long-range physics, electrical-response inference, and distributed computation, this framework extends predictive atomistic modeling toward device-relevant length scales.
cs.LG / 3 / 2610.22514
Towards Reconstructing Geographically Diverse Architecture with 3D Foundation Models
Abstract
Internet-scale photo collections of real-world landmarks have driven progress in 3D computer vision yet remain highly challenging for modern 3D foundation models (3DFMs) that estimate scene structure in a single feed-forward pass. In this work, we introduce ArchWorld, a benchmark of 3D architectural landmarks with high geographic coverage and rich metadata. We use this benchmark to conduct a thorough error analysis of current 3DFMs to understand where they fall short and how they can be improved. We examine socially relevant disparities in model error and find that model performance varies by geographic region; however, unlike typical fairness regimes, the trends only partially match training data representation. Our investigation reveals that the observed geographic disparities reflect not only differences in data representation and scene size but also differences in region-specific architectural features and background distractors that make scenes from different regions inherently vary in difficulty. Leveraging insights from these model failings, we introduce an intervention scheme that improves 3DFM performance while simultaneously reducing geographic disparities.
cs.LG / 4 / 2610.22523
Learning the Shared Latent Structure of Cryo-EM Inverse Problems
Abstract
Understanding biomolecular function requires characterizing three-dimensional structure and conformational variability, for which cryogenic electron microscopy (cryo-EM) is a powerful tool. Cryo-EM produces noisy two-dimensional particle images in which molecular structure, particle pose, and imaging conditions are jointly encoded. Recovering biological information from these observations spans inverse problems including pose estimation, heterogeneity analysis, and reconstruction. These tasks are usually addressed independently despite shared protein structure and image formation, motivating a common representation across inverse problems. Since biologically meaningful differences may be spatially localized, we study self-distillation with complementary global and masked patch objectives. Because experimental dataset discrimination can reflect acquisition-specific cues, we develop controlled synthetic and experimental evaluations and find that structural discrimination benefits from dense spatial features. We introduce the EMber family of vision transformers, pre-trained on >50 million experimental particle images from 214 EMPIAR datasets together with simulated projections from 268,000 molecular assemblies. EMber improves upon existing methods at conformational and compositional discrimination by 14.85 and 14.26 percentage points, estimates conformational populations to within 1.59pp error with a linear probe, and produces particle assignments yielding reference-quality reconstructions in fewer epochs. We demonstrate how to learn representations that capture biological information for use across multiple cryo-EM inverse problems.
cs.LG / 5 / 2610.22610
Shift-Reliance Alignment: Ranking Vision-Language Models by the Cues They Trust
Abstract
Vision-language models (VLMs) are usually chosen from leaderboards, which give one ranking for every application. On questions about objects that a scene suggests but does not contain, however, the model with the lowest error on the POPE hallucination benchmark is not the one with the lowest error on such questions. For yes/no questions about object presence, we show that the best model for a target domain can be predicted without running any candidate on that domain, given 650 to 1,200 labeled target items that calibrate a detector. We measure each model's reliance on a cue, a feature such as scene context or text in the image, by editing 100 labeled source images so that the cue suggests the wrong answer, and comparing with a matched control edit. Reliance predicts which VLMs are misled by physical sticker attacks (Spearman .93 on nine models not used to design the score) and which hallucinate on unedited COCO photographs, beyond what their tendency to answer yes explains. Shift-Reliance Alignment (SRA) scores a model by its source error on the labeled source questions plus its reliance multiplied by the shift. The shift is the proportion of target questions on which the cue misleads, and a detector estimates it on the unlabeled target images. SRA selects the model with the lowest score. In pre-registered tests on new images where most questions mislead, it lowered the error of the selected model from 30.4% under selection by source error to 18.8% on Objects365 and OpenImages but, against its prediction, not on COCO. On CueDial, a benchmark we release that varies how often a cue misleads, selection by model confidence falls more than 40 percentage points behind the best model when overlaid text denies a visible object, and SRA stays within 3 points.
cs.LG / 6 / 2610.22981
How Murky Before a VLM Knows It Cannot See? Accuracy, Confidence and Abstention on a Physical Visibility Axis
Abstract
Robustness studies of vision-language models (VLMs) grade degradation on nominal severity scales ("level 1-5") that carry no physical units, and they rarely ask whether a model's confidence follows the loss of visibility. We measure accuracy, internal (token-probability) confidence, verbalized confidence and an explicit abstain option on a visibility axis in metres: fog follows the Koschmieder model, an underwater arm scales the absorption spectrum of pure water, both are applied with a monocular metric depth map, and the visibility range V is swept over eight log-spaced levels from 250 m to 1.95 m. On 1000 COCO scenes and five small open VLMs (2-5 B parameters), accuracy on a boxed-object identification task falls from near its clean value to chance within a narrow band (half-way visibility V_acc of 12.1-15.9 m in fog and 20.5-38.6 m under water) while, for the four models that are confident on clean images, the confidence stays high: at V=1.95 m in fog their accuracy is at chance but their mean internal confidence is 0.68-0.84 and their verbalized confidence 0.61-1.00, a blind-spot width W=log10(V_acc/V_aware) of at least 0.79 decades, i.e. a factor of 6.2 in visibility. The remaining model is under-confident throughout and its internal confidence follows accuracy. An explicit abstain option follows accuracy when it names the reason (W from -0.21 to -0.06 in fog) but, for one model, lags with a generic "I don't know" (W up to 0.18). ImageNet-C fog severities 1-5 correspond to only 55-110 m of physical visibility, outside the failure band, and leave the accuracy of the models within 5 percentage points of its clean value. Blank-image and shuffled-image controls, a prior-matched answer set and a 200-image pilot that preceded the study are reported. Findings are descriptive: one draw of 1000 images, depth of assumed scale, and models of 2-5 B parameters.
cs.LG / 7 / 2610.23248
Can the Parts Fool the Test? Counterfactual Pairing Cycles for Relational OOD
Abstract
Relational out-of-distribution errors are easy to describe and hard to evaluate cleanly. An image may be natural, a caption may be fluent, and all mentioned objects may be common, yet the caption may describe the wrong relation in the image. Many evaluations intended to test this behavior can be solved without checking the pairing at all, because negatives differ in caption style, template, length, object frequency, or image statistics. We propose counterfactual pairing cycles, a blocked evaluation primitive that reuses the same two images and the same two captions on both sides of the comparison, changing only which image is paired with which caption. This makes image-only, text-only, and additive component shortcuts cancel exactly in finite samples. The same contrast also gives a training loss for relational outlier exposure and a diagnostic for bilinear vision-language scores. Across synthetic controls, GQA, and COCO, cycles separate coarse image-caption matching from directional role-binding failures, while also identifying geometric size relations that remain strongly visible in global features. In a shortcut-contaminated COCO outlier-exposure stress test, a model selected by a near-perfect shortcut validation AUROC retains substantial relational signal but remains below cycle training on exact cycles; cycle training improves CycleAcc from 0.882 to 0.927. The main lesson is simple: to test whether a model understands a pairing, the benchmark should keep the parts fixed, change the pairing, and ask only about coupling. The official implementation is available online.
cs.LG / 8 / 2610.23302
Look, Reason, Draw a Region: Recursive Refinement for Spatial Context Prediction
Abstract
Spatial context prediction requires visual grounding and spatial reasoning to identify vacant regions relative to objects in a scene. Spatial relationships and physical constraints remain challenging for large vision-language models (VLMs). Recursive reasoning has demonstrated strong performance on discrete and abstract reasoning tasks, motivating its use for improving spatial context prediction in natural images. Weight sharing enables higher accuracy with a small trainable parameter count. We introduce Spatial-TRM, a compact recursive model with frozen vision-language features, a convex region decoder, and just 11.2M trainable parameters. This weight-tied reasoner supports six spatial relationships, three reference frames, and open-vocabulary anchor objects. On the RoboSpatial-Home benchmark, Spatial-TRM achieves 29.51% context accuracy, a gain of 9.02 percentage points (44.0% relative) over a separately trained single-step model with the same parameter count. Including the frozen encoders, the model has 877.8M total parameters, approximately $15\times$ fewer than the reported 13B size of RoboPoint. These results show that repeated refinement improves spatial context prediction from pretrained vision-language features at a fixed trainable parameter count on this benchmark.
cs.LG / 9 / 2610.23653
A Language-Informed Loss for Visual Generation
Abstract
End-to-end visual generation with pixel-space loss has recently achieved competitive performance without the need for learned image tokenizers. In this work, we study whether language-informed supervision can provide a richer training signal than pixel loss for visual generation. We introduce LILIE (pronounced "lily"), a Language-Informed and Lossless Image Encoder. LILIE is an invertible image transform trained such that a vision-language model can recover captions from noised features. The encoder uses a normalizing flow transformer trained jointly with a vision-language model via next token prediction. We then use the encoder to extract a simple language-informed loss for text-to-image generation. We train JiT models with LILIE loss, reducing gFD by 14 points (66.84 $\to$ 53.01) against a pixel-only baseline on the GPIC dataset. Our code, encoders, and generative models are available at https://suchiragarwal.com/lilie.
cs.LG / 10 / 2610.23661
An Efficient Approach to Beetle Damage Detection of Banana Using Deep Learning
Abstract
Visible surface damage of banana fruits by Basilepta subcostata (the banana fruit scarring beetle) can reduce their market value. Here, manual inspection of Banana fruits is subjective and difficult to scale. This study presents a lightweight framework based on YOLOv8n deep learning model for locating individual bananas and classifying their visible surface condition as healthy or unhealthy. The term beetle damage in the paper title is used operationally for the locally observed scar-like surface condition; the model does not identify a specific beetle species or biological cause. A custom dataset of 2,473 images containing 5,827 annotated banana instances was prepared from local-market imagery in Bangladesh. The reported experiment used an 80/20 protocol with 1,982 training images and 491 test images. A pretrained YOLOv8n model was trained at 640x640 resolution for 50 epochs with a batch size of 4. On the test partition, it achieved 81.36% precision, 74.30% recall, 77.67% F1-score, 87.04% mAP@0.50 and 59.33% mAP@0.50:0.95. A YOLOv8s comparison gave slightly lower F1 and mAP while using a substantially larger checkpoint. The YOLOv8n checkpoint occupied 5.96 MB and required 6.54 ms/image for inference on an NVIDIA Tesla T4. The results support YOLOv8n as a compact baseline for instance-level banana surface inspection within the studied acquisition conditions.
cs.LG / 11 / 2610.22705
Data Provenance Aggregation and Structural Descriptors for Query-Template Identification
Abstract
Runtime query similarity identification supports several activities, including workload analysis, identifying recurring behaviour, auditing, debugging, and anomaly detection. Existing methods for evaluating query similarity mostly rely on SQL text, execution plans, or result values but do not account for input data or how results are derived during query execution. This paper proposes a novel approach using data provenance for query-template identification. We construct query-level profiles that capture explicit data dependencies, occurrence frequencies, and structural properties by aggregating tuple-level Lineage and Why-provenance across entire query results. We also derive Provenance Structural Descriptors (PSD), which summarise these profiles using a small set of numerical features. We evaluate our proposals using these profiles to identify query templates over \(12{,}500\) distinct query instances generated from parametrised TPC-DS query templates. The experimental evaluation shows that query templates can be identified solely from their provenance footprints, and that accuracy improves as more historical data becomes available. It also shows that PSD achieve high classification accuracy with limited training history and low test-processing cost.
cs.LG / 12 / 2610.22769
One Communicator, Three Vendors: Extending the MoA Predictive Theory to a Real NVIDIA+AMD+CPU MPI Ensemble
Abstract
The companion MPI-extension paper validated the first communication-bearing machine-shape ensemble: four NVIDIA A100 ranks and four CPU ranks on one NCSA Delta node under one Open MPI job, reassembled by MPI_Gatherv. This paper extends it to a multi-vendor communicator. A Slurm heterogeneous job spanning Delta's A100 and AMD MI100 partitions places A100, MI100 and CPU ranks (nine in all) under one communicator for the first time. We give the theory (a w_i proportional to 1/t_i weight split for three concurrent realizations; a het-job specification as an admissible directive) and report real results, including what the run exposed: the 1/t split assumes time linear in rows, and this fails for the MI100 rank (6 rows took 597 ms against a linear prediction of 1.6 ms). A launch-geometry model, with predictions recorded before measuring, succeeds for the A100 (time flat from 128 to 3456 rows, stepping at exactly 3520 and 6976) and places the MI100 floor in the key loop; changing its launch settings cut about 600 ms to 14.6 ms and the energy per call at 6 rows by 32x, though the MI100 remains about 95x slower than an A100 at 8192 rows. Controlled reruns show that the companion paper's 8.1x A100 result and 22 ms Gatherv were artifacts of four ranks sharing one GPU under Open MPI's default transport. A key/value-split partition whose message sizes are derived from shapes alone is then run: the ring Allreduce of the 8.5 MB message took 11.3, 16.0 and 15.3 ms against a predicted 11-12 ms, the small Allreduce missed its prediction in every run, and compute times reproduce to about 1% while communication times vary up to 4x between runs. Each idea is introduced with a figure and closed with worked examples.
cs.LG / 13 / 2610.22781
Anytime Contracts for AI Services
Abstract
AI services are prone to moral hazard, as model providers can silently route requests to cheaper, lower-quality models, and the inherent stochasticity of model outputs makes this degradation hard to detect. Contract theory addresses moral hazard by designing payment rules that reward high-quality outputs and prevent strategic overcharging. However, existing frameworks assume interactions of fixed duration, whereas AI services are typically consumed through subscriptions and pay-as-you-go APIs that users can terminate at any time based on observed performance. We show that this flexibility can break standard contracts, opening them up to abuse by either the providers or users. To address this, we extend contract theory to account for unknown exit rules. We formalize the problem as a Principal-Agent game, and show that budget-optimal contracts in this setting are proportional to anytime-valid hypothesis tests based on likelihood-ratio e-values. We give an efficient algorithm for computing anytime contracts and show how betting e-processes can be used to approximate them when the models are unknown. Finally, we evaluate our findings on synthetic data and data from deployed LLMs.
cs.LG / 14 / 2610.23193
Proportionality in Ranking Compression
Abstract
Motivated by applications in reinforcement learning from human feedback, we consider a setting where a given collection of rankings needs to be compressed into a smaller collection of rankings. This can also be seen as a combination of two fundamental social choice frameworks: ranking aggregation and multiwinner voting. We propose three proportionality notions that capture different types of representation in this setting: positional proportionality, pairwise proportionality, and proportionality for solid coalitions (PSC). On the one hand, we show that positional proportionality and PSC are always satisfiable for any input rankings and target compression size, and a desired output can be found in polynomial time. On the other hand, we prove that pairwise proportionality cannot be satisfied in general, but can nevertheless be attained when the input rankings are single-peaked or single-crossing.
cs.LG / 15 / 2610.22497
Generative Value Classifier
Abstract
Q-learning is premised on learning a function to predict the value of a state-action pair, with the goal of using such a function to guide policy improvement. Most approaches learn this function from scratch using scalar or distributional return targets. This discriminative approach leaves another source of supervision for learning representations unused: the actions themselves. Rather than directly predict the return of a state-action pair, we ask the reverse question: how likely is the action under each possible return? This perspective provides a critic whose generative component is a return-conditioned policy, which can share the actor's architecture. Training this policy supervises the critic with every action rather than with returns alone, and a squared-error loss anchors the critic's Q-value to the bootstrapped target. We evaluate our approach on both offline and online RL tasks and show gains over leading discriminative methods, including distributional RL methods.
cs.LG / 16 / 2610.22519
Learning smooth dynamics via Kinematic Interpolants
Abstract
Inferring dynamics from real-world experimental data is inherently ambiguous, as many plausible dynamical systems can explain the same positional observations at measurement times. For higher fidelity modeling of applications ranging from single-cell RNA to vortex dynamics and even flocking of birds, scientific observables increasingly also include velocity estimates that serve to resolve this issue. However, incorporating this rich information in a scalable manner, free of simulation or unnecessary approximations, has remained a difficult challenge. In this paper, we introduce Kinematic Interpolants (KI), a novel framework that extends the seminal methodology of stochastic interpolants to infer nonlinear dynamics across multiple snapshots in phase space, i.e., observables with position and velocity measurements. Empirically, we demonstrate that across several scientific datasets, KI achieves superior reconstruction of the underlying dynamics compared to existing baselines while requiring only a fraction of the training time.
cs.LG / 17 / 2610.22539
Don't Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks
Abstract
The web is not static or uniform. Websites that serve the same purpose can look and navigate very differently. Even within a single site, successive redesigns can change how familiar features are presented and accessed. Humans adapt to this by recognizing familiar interaction patterns and reusing what they already know. We give web agents a similar ability with WebMix, a mixture hypernetwork that adapts them to unseen interfaces by composing previously trained low-rank adapters. WebMix begins with a bank of adapters trained on source interfaces. At deployment, a scout agent explores the target interface and summarizes its observations into a compact manual. A lightweight router uses this manual to predict mixing coefficients over the adapter bank, which combine the existing adapters into a new one for inference. Adapting to a new interface takes only a scouting run and a forward pass, without gradient updates on the target. With Qwen and Llama models, we evaluate domain transfer on TimeWarp and WebArena-Lite-v2, and version transfer on TimeWarp. In all three settings, WebMix matches or outperforms in-context learning and LoRA baselines, often matching the average success of specialists trained directly on those interfaces. On TimeWarp domain transfer, it gains about 8 percentage points on average over the best baseline with both models. These gains do not come at the cost of seen interfaces, where WebMix keeps its lead, and extend to version transfer. Adapting to a new interface can therefore rely on exploration and composition instead of retraining.
cs.LG / 18 / 2610.22564
Deterministic Bounds on Accuracy and Stability of Metric-Space Regression and Classification
Abstract
Data-driven approaches often face ambiguity in datasets due to sensor or target noise, insufficient information, and noise in the input data. While best-fit machine learning methods can accommodate such noise through uncertainty estimates, they are prone to overfitting and it is challenging to quantify ambiguity of the relation between inputs and targets. We introduce a geometric framework that quantifies this ambiguity through the average kernel size, computed with data-dependent, i.e. model-independent, feasible sets derived from worst-case input and target noise. The framework provides computable, deterministic, non-asymptotic bounds on the best achievable empirical error and accuracy for a finite, constructed dataset, independent of the predictive model class, and applies to regression and classification with pseudo-metric input spaces and metric target spaces. The ambiguity of the input-target relation implies three consequences: an empirical error floor, an accuracy-stability trade-off, and, under standard probabilistic assumptions, a strictly positive distribution-training error gap when training below the accuracy bounds. Under standard Borel assumptions for Euclidean norms, the distributional analogue of the average kernel size is related to Bayes risk as it reduces to the optimal distribution root-mean-squared error times the square root of two. Experiments spanning tabular regression, image regression and classification datasets yield nontrivial bounds on the best achievable empirical error and accuracy on a constructed dataset, alongside descriptive comparisons with model performance on the original datasets. The bounds provide a computable, model-independent reference scale on the constructed dataset and comparisons with model performance on the original dataset are descriptive.
cs.LG / 19 / 2610.22565
Coverage Cliffs in Off-Policy Quantile Certification
Abstract
We study when logged contextual-bandit data can certify an upper tail quantile for a policy selected from the same sample. We construct a PAC-Bayes Bernstein lower-confidence envelope for the target cumulative distribution function (CDF), simultaneous over a predeclared threshold grid and all data-dependent posteriors. The resulting procedure, TailCert, uses the envelope's first crossing of level p to return an atom-safe upper confidence bound on the selected posterior's pooled p-quantile (or inconclusive). It can also be used to select a policy under a certified tail constraint. The concentration ingredients are standard; the main theoretical contribution is a matched finite-class certification boundary comparing full logs with their joint weighted CDF indicators at a fixed threshold. On an explicit rare-action class, the required sample size grows linearly with the overlap factor W, logarithmically with policy-search and confidence complexity, and quadratically with the inverse CDF safety margin; the upper and lower bounds agree up to constants in the stated regime. On this class, the CDF reduction hides which actions exceeded the budget; retaining action counts restores the full-log rate. Simulations verify the predicted search-complexity shift, winner's-curse failure of naive reuse, and the power cost of splitting.
cs.LG / 20 / 2610.22575
Structure-Aware Spatial Kernels for Gaussian Process Distributed Port-Hamiltonian Systems
Abstract
Gaussian process distributed port-Hamiltonian systems (GP-dPHS) enable data-driven modeling of distributed dynamical systems while preserving their physical structure. However, standard stationary kernels can be inefficient for spatially heterogeneous systems because they neither exploit known input invariances nor reflect structured spatial dependencies. We address this limitation with a structure-aware kernel design that removes irrelevant state coordinates and decouples state and spatial dependencies through a tensor-product covariance. This modular construction allows qualitative knowledge of spatial heterogeneity to be incorporated by selecting an appropriate spatial kernel without requiring an analytical description of the underlying profile. Numerical experiments on a nonlinear string with periodic, smoothly varying, and semi-piecewise stiffness profiles demonstrate that structural decoupling consistently improves reconstruction accuracy, while profile-matched spatial kernels yield substantially larger gains in force-field and trajectory reconstruction.
cs.LG / 21 / 2610.22583
Noise-Robust Forced Alignment for Spontaneous Speech
Abstract
We show that forced aligners degrade on spontaneous, noisy real-world speech, and that fillers, laughter, and disfluencies corrupt word boundaries, impacting many downstream applications. We revisit the classical GMM-HMM aligner with upsampled WavLM features instead of MFCCs, explicit phone units for fillers and non-lexical vocal events trained on verbatim-annotated spontaneous speech, and a discriminative projection fitted on clean and noise-corrupted feature pairs to improve robustness. The pipeline is self-bootstrapped from 149h of curated audio without manual or external boundary targets. On conversational (Buckeye) and disfluent (FluencyBank) speech it reaches word-boundary MAEs of 20.6 and 76.9ms, 19% and 17% lower than the strongest of ten baselines. Catastrophic misalignments (>200ms) are 2.7 times rarer than Montreal Forced Aligners (MFA) best model even after widening MFA's search beam, which we show removes a large pruning artifact from its conversational error; event units halve the error next to laughter, while being meaningfully more robust than existing aligners.
cs.LG / 22 / 2610.22584
Initiating Removal from Operational Symptoms in Federated Vehicular Networks
Abstract
Connected vehicles train a driving model through federated learning (FL) and drive with it, so a faulty contribution often surfaces only as a delayed symptom on the road. Federated unlearning (FU) assumes that a removal request names what to forget. A symptom names no vehicle or round, and aggregation has already spread its cause across the fleet. We propose \underline{A}ggregation-\underline{T}ime \underline{O}perational-evidence \underline{N}eutralization of \underline{E}ffects (\method{}), which makes unlearning a term of the FL aggregation rule. \method{} projects everyday driving errors and reported symptoms onto a subspace spanned by recent global increments. It propagates the surviving effect of every contribution with curvature estimated from these increments and scores all contributions in one backward pass. A two-tier correction suppresses suspected effects in every round and eliminates confirmed ones after a one-round probe. We prove that backward scoring matches forward scoring and that the elimination is exact when training is flat outside the subspace. On three driving datasets, \method{} leaves 43\% to 61\% less residual harm than the strongest baseline in eight of nine settings while staying within 0.36 points of full retraining in utility.
cs.LG / 23 / 2610.22593
Flat Loss, Evolving Features: Grokking in Modular Addition
Abstract
Can a network keep learning while its training cross-entropy stays nearly flat at a positive level? We show that it can. On modular addition, a standard testbed for grokking, ReLU networks trained with supported-Polar hidden updates, normalized output-layer updates and weight decay can keep 100% training accuracy while their hidden features change and held-out accuracy rises from zero to above 90%, at relatively flat training cross-entropy. Correct classification need not minimize cross-entropy, which still depends on prediction confidence. We prove this coexistence for one-hidden-layer networks in two finite-window theorems from constructed initializations: one gives a constant gain in the Fourier fraction of the hidden features, and the other holds at arbitrarily low positive loss, where a readout refitted on the training data generalizes before the network. Experiments from random initialization show the same coexistence on a window fixed in advance, and in deeper networks held-out accuracy also rises over post-hoc windows with a net increase in training cross-entropy. With a last-layer rate set per training fraction and head decay scaled with the prime, deeper networks at p=47 and 61 meet a criterion on post-hoc windows, fixed in advance, in nearly all fresh runs. A one-block transformer trained with the same updates and a slowly decaying output gain keeps every training pair correct at every logged checkpoint while its training cross-entropy rises and its held-out accuracy approaches one; with two blocks this holds through the held-out rise, after which training pairs begin to flip. A plateau in training loss need not be a plateau in learning. Code: https://github.com/akashkumar-d/flat-loss-evolving-features.
cs.LG / 24 / 2610.22595
Diagnosing GNN Failure Modes with Layer-Wise Probe Entropy
Abstract
Graph neural networks (GNNs) are typically evaluated only at their final layer, obscuring how task-relevant information evolves during message passing. We introduce a post-hoc diagnostic that freezes a trained GNN, fits an independent linear probe to each layer, and tracks probe accuracy alongside normalized predictive entropy. Across real and synthetic graph settings, layer-wise probe entropy distinguishes beneficial, neutral, and harmful smoothing, a capability not demonstrated by existing geometric measures. Moreover, by comparing shallow and deep models, we show that the onset of entropy degradation separates the two main GNN failure modes: structural mismatch under heterophily, which appears from the earliest layers even in shallow models, and excessive-depth oversmoothing, which emerges only as propagation depth increases. These results establish probe-entropy trajectories as a tool for analyzing information flow in GNNs and motivate future work on adaptive depth selection and early-exit mechanisms.
cs.LG / 25 / 2610.22604
Rarity of Singletons Matters & Supplementary Materials
Abstract
Estimating missing-mass and unseen-species counts from finite samples relies heavily on rare-event occupancy statistics, particularly singletons. In small-sample regimes, these statistics can be highly variable, making classical estimators sensitive to sampling fluctuations and potentially unreliable. We propose a fragility-aware (FA) framework that explicitly quantifies this sensitivity and adaptively regularizes conventional missing-mass and unseen-species estimators. Using singleton-based fragility measures derived from probability mass and occupancy counts, we formulate an entropy--Kullback--Leibler (KL) optimization that adjusts the reliance on base estimates according to their fragility. The resulting objectives are strictly concave, admit unique closed-form solutions, and require no Poissonization or asymptotic approximations. The framework provides a lightweight correction to classical estimators, including Good--Turing and Good--Toulmin, while reducing reliance on potentially unstable rare-event statistics. Across 360 experimental configurations spanning synthetic distributions and real-world language corpora, our method remains competitive in stable regimes while improving estimation accuracy and stability in fragmented, long-tailed, and small-sample settings. The findings demonstrate that accounting for the fragility of rare observations can improve the reliability of inference about unobserved species and their associated probability mass.
cs.LG / 26 / 2610.22621
Noise-shaped Sampling and Annealing for Scalable On-the-edge Training of Binary and Ternary AI Models
Abstract
The training loss of binary and ternary neural networks, expanded as a multilinear polynomial over the discrete weights, is structurally equivalent to the energy of a hierarchical high-order Ising model. We exploit this equivalence to apply Ising solvers to discrete-weight network training, while accounting for a fundamental difference in objectives: an Ising solver seeks the ground state of a fixed energy function, whereas training seeks solutions that generalize to unseen data. Our key contribution is $ΣΔ$-Sampler, an optimizer that reconciles these objectives by embedding a stochastic sampling-and-annealing step as a noisy quantizer inside a first-order $ΣΔ$ noise-shaping loop. The loop pushes noise from weight quantization, high-order frustration, and mini-batch sampling out of the low-frequency adaptation band, so that discrete weight updates follow the smooth descent trajectories embedded in the rugged loss landscape. We validate the framework on a planted quadratic unconstrained binary optimization (QUBO) benchmark with a known minimizer; on CIFAR-10, where a $13$M-parameter ternary model comes within $1\%$ of full-precision accuracy; on ImageNet-1k, where a ternary ResNet-18 comes within $0.4\%$ Top-5 accuracy of a ternary baseline; on class-incremental continual learning, where prior-task accuracy is preserved; and on MatMul-free language models up to $2.7$B parameters, where it outperforms a straight-through-estimator baseline under matched continued pre-training. Compared with standard quantization-aware training, $ΣΔ$-Sampler reduces the training memory footprint by over $50\%$, delivers an order-of-magnitude speed-up, and reduces weight-update memory traffic by four orders of magnitude, making it attractive for on-device training and fine-tuning of large binary and ternary models.
cs.LG / 27 / 2610.22623
Understanding Stability of Outer Nesterov Momentum on Stochastic Pseudo-Gradients
Abstract
Large neural networks are commonly trained across many workers, and communication between them is a major cost. Workers therefore take several local optimization steps before sending their net parameter displacements, or pseudo-gradients, to an outer optimizer, which often applies Nesterov momentum. We study how the outer momentum, the number of local steps and the averaging over workers control the training dynamics, and in particular their stability. For random quadratics, we derive an exact mean-square stability criterion in any dimension, which reveals that the inner batch size determines whether increasing the number of local steps eventually stabilizes or destabilizes the outer iteration. Moreover, while averaging the pseudo-gradients of independent workers can make larger outer momentum values admissible, the number of workers required for stability can grow exponentially with the number of local steps. Finally, we show that the mean-square dynamics governing stability can be probed directly at intermediate neural-network checkpoints using shared-minibatch perturbations. We test our theoretical findings across networks, and measure their consequences for training loss.
cs.LG / 28 / 2610.22635
Alert-List Concordance for Evaluating Imputation Strategies in ICU Sepsis Prediction
Abstract
Preprocessing is an under-measured reliability risk in clinical prediction: a model can keep a similar AUROC and still change who receives an alert. Imputation quality is commonly evaluated using reconstruction error or aggregate discrimination, neither of which directly measures agreement on which patients are flagged. At a matched alert rate, pipelines with similar aggregate performance can produce different patient lists. This question is especially relevant under the frequent, structured missingness of real ICU records. We therefore introduce alert-list concordance, which asks whether two pipelines identify the same high-risk patients, measured by the Jaccard similarity of their top-k% alert sets with patient-level bootstrap confidence intervals. On PhysioNet/CinC 2019 (40,336 patients, 71.9% missingness), we compare nine imputation strategies and a native-missingness baseline using a fixed XGBoost configuration, fitted and Platt-calibrated separately for each strategy. After Bonferroni correction, 22 of the 28 reconstruction-method pairs showed no statistically significant difference in patient-level AUROC, yet their alert lists overlap at Jaccard 0.518 to 0.750. Similar aggregate discrimination can therefore accompany different patient selections. ForwardFill is a separate outlier, pairing higher discrimination with the lowest overlap in the matrix, 0.40 to 0.45. The broad pattern holds within each hospital system and under six patient-level aggregation rules. At the row level, eight of ten methods fall within an AUROC range of 0.011. In this row-level pipeline, deep imputers did not outperform mean imputation on discrimination, and the sequential implementations cost 5 to 12 times mean imputation in total pipeline wall-clock time on our hardware.
cs.LG / 29 / 2610.22650
The Ephemeral Cost of Switching Optimizers Across LLM Training Stages
Abstract
Recent work suggests that switching optimizers between pretraining and subsequent training stages incurs a performance cost, but systematic investigation of this phenomenon is lacking. We study the mismatch between AdamW and Muon optimizers by pretraining 1.2B Llama-3 models from scratch while varying the proportion of updates taken with each optimizer, and similarly interpolating between the two optimizers during continued pretraining and supervised fine-tuning (SFT). Surprisingly, we find little evidence of a meaningful optimizer mismatch penalty: final performance is largely explained by the independent effects of the pretraining and post-training optimizers, with downstream performance generally improving as the proportion of Muon updates during SFT increases. Although matching the pretraining optimizer can reduce forgetting during SFT, this does not translate into improved final benchmark performance. We further evaluate AdamW and Muon for task-general SFT of open-weight models, across the Llama and Qwen families up to 14B parameters, where Muon remains competitive with AdamW, despite these models having been pretrained with AdamW. Together, our results suggest that the choice of optimizer for post-training can be made largely independently of the optimizer used during pretraining.
cs.LG / 30 / 2610.22661
Edit Discovery in Human-AI Interactions
Abstract
Human-AI interactions are becoming a rich source of data for improving AI systems: while many model responses are already satisfactory, others expose failures that could become valuable supervision if identified and corrected. We study this setting through the notion of an editor: a more capable evaluation process that can inspect a response and either confirm it or modify it, but whose invocation is costly. The editor may be a human expert, a stronger AI system, or the same model equipped with additional computation, context, tools, or information. The key challenge is to decide which interactions to send to the editor before knowing whether an edit will actually be made, since invoking the editor itself consumes limited resources such as expert time, inference cost, latency, or tool-use overhead. We formulate this problem as Edit Discovery Optimization (EDO), which seeks to minimize edit opportunities missed among skipped interactions while controlling unnecessary confirmations and respecting an editor-cost budget. We characterize the population-optimal policy and show that it takes the form of a cost-adjusted threshold on the posterior probability of an edit. Building on this structure, we develop EDIT (Edit Discovery via Iterative Thresholding), an online algorithm for the partial-feedback setting in which edit labels are revealed only for selected interactions. EDIT combines an adaptive edit score, a confirmation-control threshold, and a cost-aware selection rule; it satisfies the editor budget at every round and controls the average confirmation rate up to a vanishing finite-horizon term, without distributional assumptions on the interaction stream. Across four reasoning and factual question-answering benchmarks, EDIT achieves a better trade-off between editor cost, Type-I error, and Type-II error than direct LLM-based batch selection.
cs.LG / 31 / 2610.22695
Chance-Constrained Personalized LLM Routing with Online Preference Learning
Abstract
We study personalized LLM routing, where a router selects among heterogeneous models with different costs and capabilities. The router must minimize cumulative inference cost while ensuring reliable realized response quality, even though user preferences are unknown and must be learned from feedback during deployment. We formulate this problem as a constrained contextual bandit and develop an algorithm that learns preferences while balancing cost and quality. Under a logistic preference model and a strict-feasibility condition, we establish sublinear expected cost regret relative to an optimal stationary randomized contextual oracle and a high-probability guarantee on realized cumulative satisfaction. The framework extends beyond LLM routing to broader online model-selection problems with reliability requirements.
cs.LG / 32 / 2610.22713
Spectral Limits of Polynomial Graph Filters on the Long-Range Ising Model Benchmark
Abstract
The Long-Range Ising Model benchmark (LRIM) guarantees long-range dependence by construction: each label is the node's spin times the spin field convolved with a power-law coupling, so any model's distance from the target is measurable. We prove that a readout whose filter is a polynomial in the adjacency, the linear core of isotropic message passing and spectral graph networks, leaves an error floor that no degree lowers, because the target varies along a level set of the spectrum on which every such filter is constant. Computed exactly, the floor and the best filter of each degree predict two spectral networks adapted from published architectures to within a tenth of a decade. A filter escapes the floor only if it weighs separately the offsets the adjacency lumps together, and must still resolve the coupling across the whole lattice. A strided filter of about fifty coefficients does both: fitted to the training labels, its test error is 52 times below the best published network's. A dilated convolutional network beats the best published result at 256 nodes, matches it at 1,024, and gives the first in-domain results at the three larger sizes, while a Fourier neural operator with its standard modes falls further behind as the lattice grows. At its standard budget the dilated network stays two to four orders of magnitude short of a larger strided filter fitted to the labels. The analysis carries over to the benchmark's second coupling.
cs.LG / 33 / 2610.22749
Native Geometry vs. Linear Recoverability: How Alignment Changes Platonic Scaling
Abstract
Do larger models learn increasingly similar representations of the same world? The Platonic Representation Hypothesis (PRH) predicts that they should, but similar can mean native cosine geometry---nearest-neighbour structure induced by cosine similarity before any learned alignment---or linear recoverability: correspondence revealed by a supervised linear map. Using matched Legacy Survey and Hyper Suprime-Cam (HSC) embeddings across five model families, we find that Ridge alignment strengthens the relationship between model size and cross-dataset neighbourhood correspondence relative to native cosine geometry. The effect appears in all five families and persists when both representations are reduced to dimension, so it is not simply that wider embeddings give alignment more dimensions to work with. The learned maps remain far from rotations or uniform scaling, and Sparse Autoencoder (SAE) and Block-Sparse Featuriser (BSF) features give no advantage over the same alignment applied to dense embeddings. Model scale thus improves linear recoverability more clearly than native cosine-neighbourhood correspondence, without establishing convergence toward a shared native metric geometry.
cs.LG / 34 / 2610.22754
Refusal Is Not Robustness: Measuring Harmful Reachability in Autoregressive Language Models
Abstract
LLM safety evaluations via jailbreaking typically index on attack success rate (ASR): whether a generated response violates safeguards. However, ASR only evaluates sampled outputs and provides little information on the underlying generation process. Autoregressive language models define distributions over possible outputs, and a sampled refusal can coexist with substantial probability mass on harmful generation paths. We study this hidden structure through bounded continuation analysis (BCA), which models autoregressive generation as a stochastic process over token prefixes and computes the finite-horizon probability that generation reaches a harmful continuation. BCA constructs a bounded continuation tree, labels terminal continuations with a semantic harm evaluator, and estimates harmful reachability using exact probabilistic model checking. BCA complements ASR by measuring how much harmful behavior remains accessible beneath a sampled safe response. This makes it useful for distinguishing models that look identical under measured ASR values, and allocating red-team effort toward regions of the generation process with greater hidden risk. Across four instruction-tuned LLMs and 100 JailbreakBench behaviors, we show that BCA effectively quantifies hidden harmful behavior and provides a useful signal to guide red teaming.
cs.LG / 35 / 2610.22771
Reusing Calibration Gains across Model Representations
Abstract
Scale calibration learns an update for one model representation. We study how much of that update can be reused by another representation of the same pretrained model. Our transfer procedure extracts positive group-wise gains and installs them in the recipient's weights or native quantization scales without recipient-side optimization. On a quantized recipient, it preserves integer codes, zero points and storage layout. For a Qwen3-1.7B floating-point/GPTQ INT3 pair, bidirectional transfer retains 76.9-90.5% of the text-loss improvement from recipient-specific calibration on fresh WikiText and C4 cohorts excluded from project development. Shuffling the same gain values within each projection worsens both recipients, showing that their learned group assignment matters after transfer. A Llama-3.2-3B control keeps the initial native tensors at all four calibrated projections identical and changes nine projections elsewhere. Cross-installation retains 98.6-99.3% of the text-calibration benefit, alongside a small, resolved preference for each recipient's own gain. These results quantify both the adaptation carried by a reusable calibration update and the remaining benefit of recipient-specific training.
cs.LG / 36 / 2610.22788
Next Token or Vector? Learning How Far to Decode
Abstract
Standard autoregressive decoding advances one token at a time, even when several upcoming tokens can be predicted together. We propose CADENCE (Context-Adaptive DEcoding), a generation method that learns both what to draft and how far to draft. Gaussian prediction heads represent future spans as continuous latent vectors, and a shared decoder converts the selected vector into token predictions in parallel. A learned router selects the prediction length from the model's own hidden states and predicted Gaussian parameters, adapting generation granularity to the current context with single-token generation as a fallback. The backbone verifies each proposed continuation and retains the prefix that matches its greedy predictions. Experiments with Llama-3.1-8B-Instruct on Alpaca, GSM8K, and HumanEval show that router-selected positions yield higher draft acceptance than applying the corresponding prediction length everywhere. CADENCE achieves up to 1.3x higher generation throughput than standard greedy decoding.
cs.LG / 37 / 2610.22794
Ori: Drift-Aware Streaming $k$-PCA for Very Low-Rank LLM Training
Abstract
A prominent family of low-rank LLM optimizers, pioneered by GaLore, performs gradient-subspace projection: at each step, they project the gradient into a low-dimensional subspace and map the update back to the full model. Most existing work either recomputes the gradient subspace at certain intervals, or performs post-hoc adaptation to catch up to the moving subspace. In this work, we explicitly identify in practice the slow, stable manner in which LLM gradient subspaces drift across time. To this end, we propose Ori, a novel low-rank optimizer for training LLMs that captures the shifting principal gradient directions over time to achieve state-of-the-art accuracy under stringent low-rank constraints. Ori employs a streaming $k$-PCA algorithm inspired by Oja++, whose low per-step cost enables high-frequency, on-the-fly subspace tracking. In practice, Ori reduces the performance gap of the state-of-the-art low-rank optimizers to full-rank optimizers by up to 69.3\% on a 1B LLaMA pretraining task, and achieves the best performance on 10 out of 16 task scenarios in Roberta-Base fine-tuning. Together, these contributions allow Ori to further push the performance and efficiency frontiers of gradient-projecting optimizers, narrowing the gap to full-rank training for LLMs. Code can be found at https://anonymous.4open.science/r/Ori-Submission-2132.
cs.LG / 38 / 2610.22831
Partial-Revelation Panel Forecasting: Predictive Information Gain, Allocation, and Coarsening-Consistent Operators
Abstract
We study partial-revelation panel forecasting: given $k$ of $n$ annotator votes on the previous dialogue turn, predict the full count distribution of the next turn's panel. This cross-turn predictive information differs from same-item annotation value. Under any nested revelation mechanism, the population decrement satisfies $V(k)-V(k+1)=I(Y;R_{k+1}\mid X_k)$; a fitted model also contributes adjacent excess-risk differences. Coarsened-completely-at-random (CCAR) masking supplies the exchangeable random-rater interpretation used in our operators. We define coarsening-consistent panel operators with a closed-form Dirichlet-multinomial realization, learned completion, and rater-count transfer. Against per-level lookup, cold-start count negative log-likelihood differences are -0.422/-0.166/-0.014 nats on IEMOCAP/M3ED/EmoWOZ. A held-out rater experiment measures same-item value: first and second additional votes reduce NLL by 0.583/0.221, 0.578/0.184, and 0.180/0.069 nats, respectively. Cross-turn and same-item gains are not interchangeable, and CCAR masks are controlled interventions rather than observed arrival processes. Consistency is auditable but does not guarantee predictive improvement; per-level marginalization can outperform a shared table. Target-free allocation beats uniform only where row-level signal is predictable: at budget one it recovers 83--93% of retrospective feature class headroom on IEMOCAP/M3ED but does not improve EmoWOZ. Gains are modest and corpus-dependent, with several small or null partial-revelation differences. Code and aggregate result artifacts accompany the preprint.
cs.LG / 39 / 2610.22846
Benchmarking and Enhancing Agentic Causal Discovery
Abstract
LLM agents can now run causal discovery algorithms on a dataset and report what comes out. Whether they do more than that, such as noticing where an algorithm went wrong and fixing it with further tests, is hard to tell from current benchmarks. Most of them give the variables meaningful names, and an LLM shown only the names of the Tübingen cause-effect pairs already gets 99% of the directions right. With the names removed, the data are usually clean enough that picking a better algorithm than PC closes most of the gap. We introduce Heuresis, a benchmark of anonymous simulated datasets whose mechanisms range from linear Gaussian to heavy-tailed. For each family of datasets it separates what a better existing algorithm can gain from what only a closer look at the data can gain, and it traces every edge an agent changes. Every family leaves a gap that none of the algorithms, the router and the two causal-discovery foundation models we ran can close, so a score above the best chosen algorithm has to come from using the data. We instantiate the benchmark with an agent harness and three LLMs, GPT-6.1 Sol, Claude Opus 5.5 and Qwen 3.8 Flash. GPT-6.1 Sol surpasses the best chosen algorithm on all six families, by 14% adjacency F1 on the most difficult one, and Claude Opus 5.5 does so by 7.4% there. A third of the true edges the two agents restore on the two hardest families appear in no algorithm's output. The code and data will be publicly released to facilitate future work.
cs.LG / 40 / 2610.22887
Batched Contextual Reinforcement II: Does RLVR Need Long Rollout Windows?
Abstract
Long rollout windows make reinforcement learning with verifiable rewards (RLVR) expensive, yet simply shortening them can remove the correctness signal: when every single-problem rollout truncates before an answer, group-relative accuracy advantages vanish. As a technical continuation of Batched Contextual Reinforcement (BCR), we study how concurrent rewards relax this dependence on long windows. Concurrent Constraint Reinforcement (CCR) places three independently verifiable problems in one capped training completion. Correct answers completed before truncation retain credit, while a shared budget encourages shorter useful reasoning without an explicit length penalty. We connect this mechanism to conditional reward- and update-variance analyses and study 4B and 8B mathematical reasoning models. Against single-problem GRPO with a 32k window, CCR with an 8,000-token window reduces tokens per training step from 975,464 to 174,957 on 4B and from 793,832 to 165,788 on 8B, saving 82.1% and 79.1%, respectively. An 8B training trace shows increasing accuracy reward alongside shorter completions. Single-problem evaluations provide supporting evidence of retained reasoning quality. These results identify reward-preserving training structure as a practical route to lower per-step cost in short-window RLVR.
cs.LG / 41 / 2610.22891
What Accuracy Hides In Learned Transformer Programs
Abstract
Transformer Programs (TP) (Friedman et al., 2023) constrain a Transformer's architecture so that its learned computation can be extracted as readable Python code. TP expose learned computations, but token accuracy can hide incorrect complete answers. In this work, we examine how program inspection provides insights beyond accuracy scores and helps guide training. We first evaluated TP on nine algorithmic tasks. We compared token accuracy with accuracy on complete output sequences and checked for failures missed by sampled tests. Further, we assessed training data improvements informed by program inspection and failure analysis, finding gains in sequence accuracy on several tasks. We examined a counting task where a program achieved 100% accuracy on sampled test sequences but failed on another input using the symbols and lengths allowed during training. The program counted correctly, but the required answer was missing from its output labels. After adding training examples with larger counts, the retrained models passed exhaustive evaluation over the bounded domain. In the frequency ranking task, inspection showed how correct counts contributed to incorrect answers in programs trained before and after changes to the training data. Finally, we studied constraints imposed by TP training and operations. The TP training process can lead to accuracy losses late in training. Our reference constructions also show how composing the restricted TP operations can use many layers or large lookup tables.
cs.LG / 42 / 2610.22898
A Feature Selection Layer Approach for Multi-Class Classification on Neural Networks
Abstract
In the age of data-driven decision-making, deep learning models have become powerful tools for uncovering patterns and making predictions that influence many aspects of society. However, their broader adoption in scientific and business applications is often limited by a lack of interpretability. Feature weighting methods, which assess the relevance of individual features, are among the most widely studied approaches for improving interpretability. While post hoc techniques have advanced significantly, especially in modeling nuanced relationships between features and labels in multiclass problems, embedded methods typically fall short in this regard. Embedded techniques integrate feature relevance learning directly into the model training process, but they are often unable to capture label-specific relationships. To address the limitation, we propose multiple extensions to the Feature Selection Layer, an embedded feature weighting mechanism for neural networks. Through extensive experimentation, we identify the most effective extension and benchmark it against established feature weighting methods. Our results show that the proposed method matches existing techniques in identifying globally relevant features and outperforms them in capturing label-specific relevance. By leveraging the internal dynamics of the trained neural network, it reveals how individual features contribute to the prediction of specific labels, requiring no post-processing beyond the training process.
cs.LG / 43 / 2610.22925
Scalable Levenberg-Marquardt Training Framework for Deep Learning
Abstract
The Levenberg-Marquardt (LMA) algorithm is a widely used nonlinear least-squares solver known for its rapid convergence. For small-scale neural networks, LMA provides better convergence and architectural robustness than first-order optimizers like Adam. However, LMA imposes prohibitive memory and computational costs. These costs are primarily from the Jacobian matrix, whose dimensions scale with both the number of samples and parameters. This bottleneck prevents its use in large-scale neural networks. Recent efforts scale LMA using preconditioned conjugate gradient (PCG) solvers to avoid materializing the Jacobian. However, PCG only approximates the solution, requiring many network propagations and a task-specific preconditioner. Other studies optimize LMA to reduce costs, but they restrict their methods to specific tasks or network architectures. As a result, LMA remains impractical for training general large-scale neural networks. In this paper, we propose LeMA, a distributed LMA algorithm for large-scale neural network training on multi-GPU platforms. Our contributions are twofold: First, an exact dual formulation of LMA removes the memory dependence on the parameter dimension. Second, a Jacobian sharding strategy computes Jacobian shards on the fly and in parallel to avoid materializing the full Jacobian. For large-scale neural networks, LeMA reduces the memory footprint of LMA from quadratic in the number of parameters to quadratic in the batch size. This makes exact LMA steps feasible for billion-parameter models without approximate PCG solves. Furthermore, for a convolutional neural network with 1,000,022,632 parameters on an 8-GPU node, LeMA achieves a 7.97x speedup over a single GPU, corresponding to a 99.7% strong scaling efficiency.
cs.LG / 44 / 2610.22955
On the tradeoff between robustness and geometric accuracy in adversarial training
Abstract
For non-parametric binary classification in the presence of an adversary, we prove an optimal quantitative estimate for adversarial classifiers controlling geometric accuracy, as measured by the Hausdorff distance to the Bayes classifier, in terms or their robustness, as encoded by the adversarial budget $\varepsilon$. This rigorously shows that robust classifiers sacrifice at most $O(\sqrt{\varepsilon})$ or $O(\varepsilon)$ in accuracy, depending on the regularity of the underlying Bayes classifier. In contrast to prior work on the incompatibility of robustness and accuracy over low-complexity hypothesis classes, our work shows that high-complexity hypothesis classes allow for robust classifiers with precisely quantified accuracy. Our results apply for adversaries determined by any norm on Euclidean space and in both the infinite and finite data settings.
cs.LG / 45 / 2610.22976
Towards Hyperparameter Transfer for Differentially Private Optimization
Abstract
Scaling laws and hyperparameter tuning remain central challenges in training large-scale models, particularly under differential privacy (DP), where optimization is further complicated by noise injection and per-sample gradient clipping. While Maximal Update Parametrization ($μ$P) allows for zero-shot hyperparameter transfer in standard training, we show that it fails under DP constraints due to the distinct spectral properties of high-dimensional noise versus low-rank gradients. In this work, we introduce DP-$μ$P, a theoretically grounded framework that extends $μ$P to the privacy-preserving setting. By analyzing the interaction between gradient clipping, noise injection, and feature learning, we derive a Signal-to-Noise Ratio (SNR) condition that governs layer-wise learning rate scaling. Combining this with our proposed spectral scaling formulation, we provide specific scaling rules for DP-SGD, DP-Adam, and DP-Muon optimizers. Empirically, we demonstrate that DP-$μ$P enables effective zero-shot transfer of learning rates across model widths on diverse architectures (MLP, ViT, GPT-2), matching or exceeding the performance of individually tuned baselines while significantly reducing computational costs.
cs.LG / 46 / 2610.22996
TALLOP: Joint Optimization of Local Training and Communication Compression for Time-Constrained Federated Learning
Abstract
Minimizing the wall-clock time required to reach a target accuracy is crucial in federated learning (FL), where many applications require frequent model updates as data distributions evolve over time. Nevertheless, in heterogeneous environments, slow clients with limited computational and communication capabilities can delay global synchronization, forcing faster clients to wait and increasing the overall training time. To mitigate this straggler effect, recent studies have introduced per-round time constraints and adapted client workloads within the available time. However, these approaches primarily focus on local computation, whereas communication can become a major bottleneck in communication-limited environments, motivating the joint optimization of training workload and communication compression. Although previous studies have explored such joint optimization, determining an effective training--communication balance under a fixed per-round time budget remains challenging, as the resulting model performance depends on both training progress and compression-induced degradation. We therefore propose TALLOP, which jointly optimizes the local training workload and communication compression under per-round time constraints. TALLOP predicts the training loss achievable under different computation budgets from the early-stage mini-batch loss trajectory (LENS), and estimates the degradation caused by communication compression using a compression-error model (CLEN). TALLOP combines these predictions to select the training-communication allocation that minimizes the post-compression loss under the given round time constraint. Experiments in heterogeneous FL environments show that TALLOP generally achieves competitive or improved time-to-accuracy compared with approaches that optimize only local training or neither, although the benefit over training-only adaptation depends on the setting.
cs.LG / 47 / 2610.22998
A Label Noise Robust Algorithm for Graph Data using Neighborhood Label Distribution
Abstract
Graph Convolutional Networks (GCNs) have demonstrated exceptional performance on graph-structured data. However, their success heavily relies on the availability of clean labels, an assumption that often fails in real-world scenarios. The learnability of GCNs is intrinsically linked to the neighborhood structure of nodes, making it critical to understand how this structure changes in the presence of label noise. In this work, we investigate how the neighborhood label distribution of a node changes under label noise. We mathematically derive that label noise smooths this distribution, which inherently increases its entropy and severely decreases interclass distinguishability. To address this challenge, we propose Leave-one-out Entropy Node Screening (LENS), an end-to-end framework that identifies noisy nodes by measuring the reduction in neighborhood entropy when a specific node is removed. LENS further incorporates a relabeling mechanism to recover and preserve valuable feature information instead of entirely discarding the identified noisy nodes. Crucially, the end-to-end computational overhead of LENS is remarkably low. Extensive evaluations demonstrate that our algorithm achieves state-of-the-art node classification accuracy while executing an order of magnitude faster than existing noise-robust approaches.
cs.LG / 48 / 2610.23007
A Minimal ReLU Classifier on Non-Linearly Separable Data: Convergence and Implicit Bias at and beyond the Edge of Stability
Abstract
Gradient descent (GD) with large learning rates can successfully optimize neural networks even though classical smoothness-based learning rate conditions are violated. Yet theory for this regime largely covers linear models or linearly separable data. We analyze a minimal setting outside both regimes. Specifically, a two-neuron ReLU network trained with logistic loss on 1D non-linearly separable data. The objective is nonconvex and nonsmooth, and GD exhibits edge-of-stability (EoS) behavior. We identify an active cone in parameter space that contains every perfect classifier, where the problem becomes piecewise logistic regression with a varying effective dataset. Within this cone, we prove a finite-time loss bound for constant learning rates and show that GD converges in direction to an explicit max-margin vector, while the orthogonal component converges to the unique minimizer of a non-strongly convex potential. We further establish catapult results where exponentially small loss is reached after a single loss spike. Furthermore, we propose an adaptive schedule that lets learning rates grow to infinity without leaving the active cone. We prove that this schedule achieves a loss rate of $O(1/(T\log{T})$, which is faster than the standard $O(1/T)$ rate. Our theoretical results extend to local convergence on more general network and are validated on MNIST dataset. Overall, we show that this setup provides a tractable neural-network model for studying large and adaptive step sizes beyond linear predictors and separable data.
cs.LG / 49 / 2610.23009
"Deep lifts": Neural networks can generalize even with the dumbest optimizer
Abstract
Given a labeled dataset $\{(x_i, y_i)\}_{i = 1}^n$, consider drawing a parameter vector $\hatθ$ uniformly at random from those that interpolate the data within a width-$W$ depth-$D$ ReLU architecture with continuous weights bounded within $[-M, M]$. We study the generalization properties of $\hatθ$ in the case where the dataset can, in fact, be explained by a smaller $w$-by-$d$ ReLU network. Despite the obliviousness of the selection rule for $\hatθ$, it adaptively generalizes at the rate $\tilde O\left(\frac{wdWD}{n}\right)$, rather than being governed by the VC dimension of the full $W$-by-$D$ hypothesis class. The reason is due to a particular inductive bias in ReLU networks: the $w$-by-$d$ parameter *witnesses* an exponential volume of $W$-by-$D$ interpolating parameters, which dwarfs the volume of poorly generalizing parameters (conditioned on interpolation). We explicitly construct this set of near-equivalent $W$-by-$D$ parameters by a proof technique called *deep lifting*, and then tie the achievable volume to stability properties of the $w$-by-$d$ witness. Examples show that these techniques allow us to convert approximation-theoretic results into generalization guarantees for nonparametric classes. We further show that generalization of uniform-random interpolators can survive label corruptions.
cs.LG / 50 / 2610.23036
Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design
Abstract
Looped language models reuse physical Transformer blocks across logical depth, but established architectures entangle recurrence placement, input injection, and recurrent state organization. We isolate these axes in models trained from scratch at matched parameter scale, logical depth, and token budgets, and then test whether the resulting design principles transfer to resident- and active-parameter-matched Mixture-of-Experts (MoE) models. The prelude-recurrent block-coda sandwich architecture consistently improves performance on context extraction and reasoning-intensive tasks. Input injection primarily enhances knowledge retrieval and context utilization, whereas random initialization and separate H/L recurrent states provide mixed or negative returns. Taken together, these results favor the Huginn-style design: at the 730M-parameter and 336B-token scale, the complete Huginn architecture outperforms Ouro across all ten benchmarks. The advantages of Huginn over Ouro transfer to MoE models. At the 8BA0.8B (32BA3B when unrolled) scale trained on 500B tokens, Huginn outperforms Ouro on nearly all benchmarks spanning knowledge, reasoning, and coding, while also exhibiting more balanced expert routing. Under matched training and inference FLOPs, the Huginn-style model uses 75% fewer resident parameters than a 112-layer feedforward MoE baseline. Despite its smaller physical parameters, it outperforms the feedforward baseline on DROP and GSM8K, matches it on MATH500, and remains competitive on the other benchmarks.
cs.LG / 51 / 2610.23042
GR-OPD: Group-Relative Sequence Credit for On-Policy Distillation
Abstract
On-policy distillation (OPD) provides token-level teacher feedback, yet common sampled-token updates assign credit locally and omit the effects of earlier actions on subsequent feedback. We introduce \method{} (Group-Relative On-Policy Distillation), which decouples feedback computation from credit assignment by aggregating token feedback into calibrated sequence-level credit. We establish a theoretical basis for this approach by showing that, under ideal on-policy assumptions, broadcasting raw summed feedback to every response token yields an unbiased estimate of the negative gradient of the sequence-level reverse-KL objective. To support effective learning in practice, \method{} discounts later feedback that may become less reliable under distribution shift and uses same-prompt rollout groups to calibrate teacher guidance. It preserves relative response preferences while retaining a controlled fraction of the group mean, which we term collective teacher pressure. Across five Qwen3 teacher--student settings and six mathematical reasoning benchmarks, \method{} outperforms vanilla OPD in macro Avg@K in every setting, including gains of 10.9 and 11.5 percentage points in two settings with matched thinking patterns. It also maintains stable training under mismatched thinking patterns, where vanilla OPD collapses. Ablations show that retaining an appropriate amount of collective teacher pressure can improve learning, while excessive collective teacher pressure or an unrestricted feedback horizon can destabilize training.
cs.LG / 52 / 2610.23073
Path to Prior-Data Fitted Relational Models via Learned Latent Multi-table Representations
Abstract
Prior-data fitted networks (PFNs) show that pretraining on large populations of synthetic tasks can enable generalization to unseen datasets with little or no downstream fine-tuning. This paradigm is effective for transformer-based tabular and relational foundation models such as RDB-PFN, but remains largely unexplored for graph-based relational models built on message passing. We pretrain the graph-based relational foundation model Griffin on large collections of synthetic relational databases and ask whether curriculum learning, which benefits transformer-based relational learners, also improves message-passing representations. Rather than fine-tune on downstream tasks, we use the pretrained encoder as a general-purpose relational representation learner and pass its latent representations to a frozen in-context predictor (\ticl{}); no component is trained on the target task. On 16 real-world relational classification tasks, evaluated over five seeds against raw-feature and untrained-encoder baselines, a purely synthetic curriculum that begins with schema-guided data and continues with schema-agnostic data gives the strongest frozen in-context performance\rev{, outperforming real-data-only pretraining and every unordered single-stage mixture of the corpora; when appended after real-data pretraining, it yields the best encoders overall}. These results provide initial evidence that PFN-style synthetic pretraining transfers to graph-based relational foundation models.
cs.LG / 53 / 2610.23074
Where Latents Belong and How They Flow: Joint Identification of Structural Roles
Abstract
Latent factors in dynamical systems can play different roles across observations and time. For example, an idea in a group meeting may be expressed by only some participants, inherit information from earlier discussions, and shape what is said next. We formalize this notion as the structural role of a latent factor, describing where it is expressed, whether it receives influence from the past, and whether it transmits influence to the future. The main challenge is correspondence, since the three structures may be identified under different permutations, while a structural role is defined only when they are aligned to the same latent coordinate. We show that, under suitable structural and variability conditions, all three roles are jointly identifiable under a common permutation at each time point. Guided by this theory, we develop RoleTrace, a framework that traces structural roles in a shared latent coordinate system. Synthetic experiments validate structural recovery, while experiments on multiparty meetings examine the identified roles in real sequential interactions.
cs.LG / 54 / 2610.23103
Towards a GraphNARX model for structural dynamics
Abstract
The current paper introduces the GraphNARX model, a graph-based nonlinear autoregressive exogenous model for structural dynamics. The model is built to account for the geometry of structures and to model dynamic behaviour as the interaction between elements of the structure. Similar to other types of NARX models, GraphNARX is used for system identification. Nodal attributes of the graph encode past displacements and nodal forces, and the output of the model is the next-timestep displacement. The graph formulation of the model allows it to generalise to structures of different dimensions and degrees of freedom. Moreover, because graph-learning is performed locally between the nodes, the GraphNARX model can also learn to infer displacements of nodes for which data are not available, i.e.\ to perform virtual sensing. The proposed model's performance is evaluated in a system-identification framework and a virtual-sensing framework, and achieves good accuracy and exhibits generalisation capabilities to structures of different sizes.
cs.LG / 55 / 2610.23110
Domain-specific language alignment of multimodal embeddings towards explainable predictions
Abstract
Multimodal data are commonly fused into latent representations that improve downstream task performance but are not interpretable, limiting their application in domains that require explainable model decisions. Vision-language alignment has made image embeddings interpretable through language, but requires vast image-caption training data. However, captioned data are typically not available for multimodal embeddings, as relevant features cannot directly be observed and annotated from embeddings alone. Here, we propose a framework to language-align any embedding space that can be linked to interpretable auxiliary data. Captioning functions convert auxiliary data into partial captions; a soft contrastive loss weights pairs by the similarity of their properties that each caption describes; and a rank-based concept retrieval task evaluates alignment against the auxiliary data rather than the captions. We apply this framework to multimodal Earth embeddings to facilitate explainable predictions for two species distribution modelling tasks: S2BMS and SatBird. AlphaEarth embeddings beat previous state-of-the-art models on 6/7 metrics, while our framework language-aligns domain-specific concepts of the embedding space, outperforming previous Earth observation vision-language models on average concept retrieval for both datasets. Finally, aligned concepts recover species-habitat associations ($ρ$=0.44) and relate habitat descriptions to species predictions, showing that abstract embeddings can be interrogated in the language of domain experts.
cs.LG / 56 / 2610.23134
MetaCurate: Evolving Curation Filters for High-Quality Agentic Coding Data
Abstract
Supervised fine-tuning on agent trajectories is an effective way to improve coding agents, but selecting useful training examples requires more than identifying successful runs. A resolved patch may follow inefficient reasoning or unnecessary tool use, while a failed run may still contain useful intermediate steps. Training pipelines therefore rely on rules and qualitative assessments to filter trajectories, yet designing these filters typically requires substantial human expertise, and validating improvements through repeated training and evaluation is expensive. We introduce MetaCurate, a training-free framework that evolves executable trajectory curation filters using pairwise LLM judgments. Disagreements between the filter's rankings and the judge's preferences provide targeted feedback for jointly revising its rules, rubric, and aggregation policy. The resulting filter scores trajectories independently, so pairwise comparisons are required only during search. Starting from a blank filter, MetaCurate autonomously discovers a curation strategy that achieves competitive downstream performance without task-specific handwritten criteria. We validate downstream utility through controlled SFT experiments on our uncurated trajectory pool and the Nemotron v3 SWE dataset, with random selection and hand-authored filtering as baselines. In our experiments, MetaCurate achieves the highest mean SWE-Verified score among the compared methods, including hand-authored rule-based filtering and a training-loop baseline whose search costs 3.5x as much.
cs.LG / 57 / 2610.23148
Grounded Credit Assignment in Reinforcement Learning for Multi-Turn Search Agents
Abstract
Outcome-reward reinforcement learning such as GRPO gives every token of a search-agent rollout the same advantage, although the rollout exercises two capabilities: finding evidence and using it. We show that accuracy improves without grounding: Search-R1 raises accuracy on all seven QA benchmarks, yet on Bamboogle the evidence-supported share of correct answers falls from 84.0% to 42.9%. We attribute this to GRPO's credit assignment: rollouts that retrieve answer-bearing evidence are favoured in aggregate only in proportion to their accuracy gain over the rest, a grounding gap that answering from memory or misusing evidence shrinks or reverses. To address this, we propose Grounded Credit Assignment (GCA), which uses verifiable sub-outcomes (evidence retrieval and answer correctness) to correct search and answer credit separately: a wrong answer is blamed more than the search that found the evidence, and a correct answer that bypasses retrievable evidence is credited less. Never negative on evidence-finding search, the correction does not shrink with the grounding gap; capped at 10% of each token's advantage, it never overrides the outcome signal. Without extra models, rollouts, or training stages, GCA improves average EM over Search-R1 by 5.7 (Qwen2.5-3B) and 6.3 points (Qwen3-30B-A3B) and raises this share from 76.9% to 93.4%.
cs.LG / 58 / 2610.23160
Nearly Linear Private Learning of Littlestone Classes
Abstract
Differentially private learning aims to learn from data while protecting individual training examples. For binary classification, finite Littlestone dimension characterizes approximate private PAC learnability (Alon et al., JACM 2022). Understanding how this dimension controls private sample complexity remains a central question. For classes of Littlestone dimension $d$, the first general upper bound was doubly exponential in $d$ (Bun, Livni, and Moran, FOCS 2020). Subsequent work improved this to $\widetilde O(d^6)$ (Ghazi et al., STOC 2021) and then to $\widetilde O(d^5)$ (Lyu, STOC 2026). We prove a nearly linear upper bound of $O(d\log d)$. Our main algorithmic contribution is a structural exponential mechanism for data-dependent candidate sets equipped with a rank satisfying suitable stability conditions. For natural Littlestone classes such as subspace indicators, this rank has a simple geometric interpretation: it is the codimension. This yields a private PAC learner for subspaces of dimension at most $d$ over any field, with tight linear dependence on $d$ and no dependence on the ambient dimension, improving the quadratic dependence of Kaplan et al. (STOC 2025). Our second contribution is combinatorial: drawing on model-theoretic ideas, we construct a suitable rank for general Littlestone classes and establish the counting bounds required by the mechanism.
cs.LG / 59 / 2610.23173
Structural Affinity: An Algorithm for Coherent Cage-Based Shape Design
Abstract
Cage-based deformation (CBD) parametrises 3D geometry through a small set of control handles, effectively enabling a variety of tasks from animation and interactive modelling to tool-based co-design, neural deformation transfer, and aerodynamic shape optimization. Many optimization approaches relying on CBD employ regularisation strategies to reduce the distortion of the deformed mesh by encouraging smooth, regular cage deformations. These techniques offer soft guidance to an optimization algorithm towards desirable deformations, but compete with the task objective rather than changing the deformation parameterization itself. In this work we introduce Structural Affinity for CBD, an explicit regularisation method which enables smooth deformations through control cages by exploiting a two-cage structure to couple cage-handle motion directly. Structural Affinity precomputes an affinity matrix using mean-value coordinates, binding the inner control-cage to a coarser outer-cage. This matrix defines how much a cage handle displacement propagates to neighbouring handles which substantially decreases and often eliminates cage self-intersections, reduces mesh distortion, and adds linear overhead at deformation time. Across robotic control--morphology co-design, source-to-target neural deformation transfer, and aerodynamic wing shape optimization, Structural Affinity produces more regular deformations and substantially reduces reliance on hand-designed structural loss terms. We report consistent gains on three complementary diagnostics: the cotangent Laplacian, a proxy for surface smoothness; cage distortion, capturing whether control handles move coherently; and surface roughness, which measures undesirable oscillation in wing geometry.
cs.LG / 60 / 2610.23185
A Stronger Teacher Within: Strengthening MoE Models via On-Policy Bidirectional Distillation
Abstract
Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved Mixture-of-Experts (MoE) LLMs, but its outcome-level rewards provide only sparse supervision. On-Policy Distillation offers dense token-level feedback, yet requires a stronger compatible teacher, while on-policy self-distillation relies on hand-designed privileged information. We observe that the MoE architecture itself offers a way to resolve this dilemma, as increasing the activation budget per token exposes latent expert capacity and yields a potentially stronger teacher within the model. However, naive expansion degrades performance because newly co-activated experts are not optimized to work together. We therefore propose MoEvolve, a bidirectional distillation framework following a recovery-then-improvement strategy: we first restore the activation-expanded high-$k$ model and further strengthen it by exploiting its additional active capacity, then reverse the direction of distillation to transfer the enhanced capabilities back into the original low-$k$ configuration, retaining its inference-time activation budget. Across diverse MoE models and benchmarks, MoEvolve improves the initial models by up to $22.2$ domain-average points and outperforms same-budget RLVR baselines by up to $6.5$ points. Notably, the final low-$k$ models often surpass their activation-expanded teachers.
cs.LG / 61 / 2610.23190
LangNCO: Compiling Natural-Language Constraints into Inference-Time Interventions for Neural Combinatorial Optimization
Abstract
Neural combinatorial optimization (NCO) rapidly produces high-quality solutions for problem structures encountered during training, but operational requirements may change after deployment and be communicated in natural language. We propose LangNCO, a framework that compiles instance-specific requirements into structured inference-time interventions for a frozen NCO solver, without changing the pretrained policy's architecture, learned input representation, or parameters. A pretrained LLM serves as a replaceable semantic front end, translating each requirement into a schema-validated domain-specific language (DSL) rule set. A deterministic Rule Executor applies the rules as nonexpansive hard action masks. Because local masking prevents immediate violations but does not ensure that a complete feasible continuation remains reachable, Critical-Offset Decoding extracts rule-referenced critical nodes and reallocates part of a fixed rollout budget to multi-strength biased rollouts. Using controlled task-specific language to isolate solver adaptation from open-ended language robustness, we evaluate a CVRP policy on road closures, time windows, vehicle assignments, visit-order requirements, all pairwise combinations, and twelve additional constraint types. LangNCO achieves 100% feasibility on the four primary datasets with mean reference gaps of 1.63--3.26%, 97.33--100% feasibility when separately generated rules are composed for pairwise constraints, and 98.67% feasibility across the additional constraints. These results demonstrate modular post-training adaptation of a frozen NCO solver through inspectable language-derived rules, without retraining.
cs.LG / 62 / 2610.23204
Non-Parametric Optimization for Scalable Learning in Stochastic Decision Problems
Abstract
Stochastic optimization (SO) plays a central role in decision-making under uncertainty. Among SO problems, time-varying stochastic optimization (TV-SO) is particularly important due to its applications in adaptive control and machine learning. Non-parametric approaches have been proposed for time-varying deterministic optimization; however, they have not been developed for their stochastic counterparts. This work addresses that gap by developing a stochastic variational framework based on Malliavin calculus. This framework yields non-parametric optimality conditions for SO problems with stochastic decisions and supports the design of a scalable deep-learning algorithm that is insensitive to the parameterization dimension. This algorithm, called the Stochastic Path Follower (SPF), is applied to two important problems under distribution drift, namely least-squares recovery and logistic regression. Experimental results show that the proposed approach outperforms state-of-the-art learning-based and gradient-based methods in both performance and scalability.
cs.LG / 63 / 2610.23239
Response-Informed Value Estimation for Cost-Aware Multi-Turn LLM Routing
Abstract
Multi-turn LLM routing selects a model at every turn before its response is available, yet task success is revealed only when the interaction ends, leaving each intermediate choice without a label of its own. Logged trajectories, however, retain the response each call produced, which exposes what each turn required and how well the chosen model served it, evidence that outcomes alone hide. We introduce RIVET (Response-Informed Value Estimation for Turn-Level Routing), which exploits this post-call evidence during training while keeping routing strictly pre-call at deployment. RIVET jointly trains two value estimators over a shared state-model representation: a deployable Routing Value Estimator that scores candidate models from pre-call information, and a training-only Response-Informed Value Estimator that additionally observes the realized response. Conditioning on the response turns response-induced return variation from residual noise into an observed covariate, and we show that this reduces the leading asymptotic estimation error of the shared representation. On the MINT benchmark, RIVET solves more held-out tasks than the strongest multi-turn baseline while using 74.7% less total call cost.
cs.LG / 64 / 2610.23242
Anisotropic Representation Dynamics in Transformers
Abstract
Small differences between internal representations can shrink or grow as they pass through a Transformer layer, affecting how information is preserved and how sensitive the model is to perturbations. Yet it remains unclear when changes in attention weights can turn a representation difference that would otherwise shrink into one that grows. We develop a \textbf{Mixture-Gated Theory of Anisotropic Representation Dynamics} that separates self-attention into value transport, query gating, and key gating. Our analysis reveals how the geometry of attended representations constrains gating and how interactions between these mechanisms can reverse contraction induced by value transport. We formalize this phenomenon as gating-induced regime switching, where a representation difference contracts under fixed attention routing but expands under full attention. We derive a necessary and sufficient spectral criterion that certifies the existence or absence of switching and identifies switching directions without exhaustive search. Beyond existence, our theory characterizes maximum relative switching strength, certifies the severity of contraction that gating can overcome, and determines the largest dimension of a subspace in which every nonzero direction switches. Experiments across four Transformer architectures validate these predictions through analytical and finite-forward comparisons, revealing distinct and joint query--key gating mechanisms and nontrivial switching subspaces. Our results provide a mechanistic and spectral understanding of how attention routing can alter local representation dynamics.
cs.LG / 65 / 2610.23251
FLARE: Physics-Grounded Evaluation of Language-Model Agents for Feeder Restoration with Degraded Observations
Abstract
Restoring service after distribution-grid faults requires decisions from incomplete and conflicting information while respecting electrical constraints. Addressing this need, we introduce FLARE (Feeder-Level Agent Restoration Evaluation), a physics-grounded benchmark for post-isolation feeder restoration. FLARE con- tains 800 AC-feasible cases across five observation settings and allows agents to gather evidence, evaluate switches, and perform restoration actions under fixed operational limits. We evaluate five instruction-tuned models over 78,750 episodes, including topology, distributed-generation, and loading shifts. Mean normalized regret ranges from 0.082 to 0.148. Although every model has zero median regret, P95 regret ranges from 0.814 to 1.000, revealing an important failure tail. Shadow replay shows that 96.4% of 15,706 guard-rejected actions would have caused a physical violation. FLARE enables reproducible evaluation of language-model based restoration support while emphasizing the need for external safety checks.
cs.LG / 66 / 2610.23255
Sample, Mix, or Verify? Allocating Test-Time Compute under Imperfect Verifiers
Abstract
A common way to make language models and agents more accurate is to spend more compute at inference time: generate many candidate solutions and submit the one a verifier prefers. Success then requires that some candidate is correct and that the verifier picks it. The first grows predictably with the number of samples; the second does not, because real verifiers (unit tests, majority votes, LLM judges, reward models) accept wrong answers, reject right ones, and make correlated mistakes. With a fixed budget, a practitioner must decide how many samples to draw, which models to draw them from, and how much to spend on verification, and today these choices are made by rules of thumb. We introduce SPLIT, which makes them from a small labelled pilot. It rests on an exact formula for the probability of submitting a correct answer under accept/reject and top-score selection that accounts for varying task difficulty, blind spots shared across models, and correlated repeated verifier calls. SPLIT fits the formula by empirical Bayes and uses it to plan budgets on unseen tasks. Across seven benchmarks spanning math, science QA, code and software-engineering agents, 26 models, 175 agent systems and six kinds of real verifiers, all from public releases, SPLIT lands within 0.02-0.8 percentage points of the best allocation in hindsight, while every fixed rule we test loses at least 2.5 points in some setting. Regret depends only on how well SPLIT ranks configurations, and it stays near-optimal under compute, price and latency costs and online. The formula also explains when asking more LLM judges lowers precision. As a forecaster far beyond the pilot, SPLIT beats power-law fitting when the verifier caps success, and we report where it loses.
cs.LG / 67 / 2610.23266
Beyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents
Abstract
LLM-agent memory selection often uses retrieval rankings based on record relevance or individual influence. Such rankings cannot identify sufficient subsets under redundant or jointly necessary evidence. We study per-decision memory sufficiency: selecting a small subset that preserves full-memory behavior. To separate search error from model error, we introduce JointCoreBench, a benchmark with nine interaction families, six workflow domains, natural-language decisions, and exact sufficient-set oracles, and JointCore, a relevance cost-ordered group-deletion baseline that tests blocks before individual records. On 6,912 exact conditions, cost-ordered JointCore recovers a minimum-cost sufficient set in 100.0% of cases, reduces subset evaluations from 30.00 to 11.13 (62.9%), and achieves 87.8% active-memory savings. Across four 12B-70B model families and 432 tasks per model, its unweighted mean is 10.28 versus 16.00 model queries and 50.6% versus 43.1% minimum-cost recovery relative to cost aware singleton deletion. Both methods preserve full-memory behavior under the deterministic primary interface. Tests with physical deletion, UNKNOWN masks, PENDING sentinels, counterfactual values, and alternative labels identify interface dependent behavior. JointCore is an efficient benchmark baseline; the study does not establish deployment-ready memory compression.
cs.LG / 68 / 2610.23272
DeDiCa: A Scalable Deconfounding Discrete Causal Generative Model
Abstract
We introduce DeDiCa, a practical framework for efficient causal inference with discrete variables and hidden confounding that revisits classical probabilistic modeling through a tensorized parameterization of causal Bayesian networks, enabling tractable inference. Given a causal graph, DeDiCa is trained once on observational data by maximum likelihood, exactly marginalizing over hidden confounders, and can then estimate any (proxy-)identifiable interventional query. Despite this deliberately simple parameterization, DeDiCa matches state-of-the-art neural causal generative models in ATE estimation, while supporting memory-efficient CPU training and exact interventional inference in milliseconds. Compared with causal foundation models, DeDiCa scales more effectively with additional observational samples and achieves competitive accuracy at substantially lower inference time and memory. DeDiCa provides a tractable, accurate, and scalable approach to causal inference in discrete settings with hidden confounding.
cs.LG / 69 / 2610.23315
Untrained CNN and coefficient-space regularization for joint reconstruction in photoacoustic tomography
Abstract
The joint reconstruction of the initial pressure and spatially varying sound speed in photoacoustic tomography from boundary measurements is considered. A regularized data-driven framework is proposed in which the unknown fields are represented in finite-dimensional wavelet spaces and fixed nonlinear features generated by an untrained convolutional neural network (CNN) are used to reconstruct the initial-pressure coefficients through Tikhonov-regularized least squares. A learned coefficient-space prior is subsequently used to infer the squared sound-speed coefficients. Numerical experiments demonstrate accurate reconstruction on held-out data and strong generalization to previously unseen phantom geometries. The frozen reconstruction map also remains stable under measurement noise not represented during training, with only minor changes in reconstruction accuracy for noise levels up to $10\%$. In comparison with the supervised trained-CNN approach considered in this work, the untrained representation yields substantially smaller reconstruction errors while requiring approximately one-sixth of the computational time. These results demonstrate that fixed random convolutional features combined with regularized coefficient-space learning provide an efficient and robust approach to joint reconstruction in photoacoustic tomography.
cs.LG / 70 / 2610.23328
Peak-Memory Schedules You Can Trust: Exact Realization, Certified Optimality, and the Limits of Recomputation for Inference
Abstract
Real-time inference on small devices needs low activation memory without extra latency, which makes operator order the most valuable memory lever: it is free in time. We show that the usual way of deploying a better order, handing it back to the compiler and regenerating code, is unreliable. In PyTorch Inductor the regenerated program misses its predicted peak in 22 of 54 CPU configurations. It is worse than the stock program in 9 of them and in 19 of 50 on a GPU, and on the GPU it also changes numerical results. We instead profile the generated program, model its memory exactly from its own statements and their read/write effects, and permute those statements. We prove that this realizes the modeled peak whenever allocation does not depend on history. It does so to the byte, with bitwise-identical results, in all our CPU runs. On the GPU we characterize why the caching allocator breaks the condition. To know when to stop, we give forward and reverse "trap" lower bounds, each computable with $O(n^2)$ minimum cuts. They certify the found order optimal for 94% of held-out wrapper graphs and all 16 microcontroller-class models we study, in under a second. So most inference graphs are already scheduled optimally. Branching networks are the exception, where reordering is free and pays, e.g. 10.0% on DeepLabV3. In compiled training we save up to 7.6%. Beyond the certified optimum we study selective recomputation with per-operator caps. It is strongly NP-hard for every fixed cap even at treewidth two, encodes knapsack exactly on series-parallel graphs, and is solvable exactly on bounded treewidth. It is useless on public tiny models whose peak is a single layer, but on architectures that hold a cheap value for a late reuse it pays: on the public ReDimNet speaker networks, whose order is already optimal, it lowers the peak by 5.2-6.1% at under 0.4% extra MACs, with bitwise-identical outputs.
cs.LG / 71 / 2610.23335
HiPredict: Predicting the missing HiRISE RED4 spectral channel with machine learning
Abstract
The High Resolution Imaging Science Experiment (HiRISE) onboard the Mars Reconnaissance Orbiter (MRO) acquires imagery of Mars at 25--30 cm/pixel, supporting studies of martian geology, climate history, and landing-site characterization. Its 14-detector focal plane assembly includes several overlapping detectors that enable the acquisition of three-channel colour-infrared images, which are important for identifying compositional variations and studying active surface processes. In 2023, a detector in the center of the focal plane assembly (RED4) began failing intermittently, creating a gap in acquired imagery and reducing the instrument's colour-imaging capability by half. The HiRISE processing pipeline provides a synthetic red (SYN4) channel obtained by linearly scaling the co-located infrared (IR10) channel. However, this approach cannot capture observed non-linear relationships between the colour channels, limiting its accuracy and applicability. Here, we present HiPredict, a machine-learning framework that reconstructs RED4 from co-located IR10 and blue-green (BG12) imagery, neighbouring red-channel observations, and acquisition metadata. The model employs a U-Net architecture to learn non-linear spatiotemporal-spectral relationships between these inputs and is trained on 81,717 training samples, representing 13,625 observations. Evaluated on 23,348 test samples, representing 3,892 pre-failure observations, HiPredict reconstructs RED4 with a Normalised Mean Absolute Error of 1.1%, substantially outperforming SYN4 (3.1%). The reconstructed products improve colour image recovery and more closely represent features of interest such as icy craters, stratigraphy exposed along ridge crests and chloride deposits. We release HiPredict as an open software tool. This work demonstrates the potential of machine learning to mitigate data loss as planetary missions age and naturally degrade.
cs.LG / 72 / 2610.23337
Minimax-optimal Scaling Laws in Ridge Regression via Transfer Learning
Abstract
Pretrained models are often reused across downstream tasks, motivating the question of whether this transfer learning procedure can provide a substantial performance improvement, namely, in terms of the scaling law. We address this question in the prototypical setting of high-dimensional regression, where (i) a pretrained \emph{source} model is trained via ridge regression, and (ii) a fine-tuned \emph{target} model is regularized via the $L_2$ distance from the source estimator. First, we establish a deterministic equivalent for the risk of this two-step procedure under both model and covariance shift. Next, by specializing the deterministic equivalent to a shared regression function and commuting power-law covariance spectra, we identify a regime where the resulting scaling law achieves the minimax-optimal exponent. Notably, in this regime, the optimal source regularization depends on source data only, indicating its universality across all targets. Furthermore, we show that the same optimal rate is achieved by performing the fine-tuning step via one-pass stochastic gradient descent. We support our theoretical analysis via experiments on both synthetic settings and language models.
cs.LG / 73 / 2610.23338
SETAREH Transform: A Sorting-Assisted Haar-Wavelet Sparsification Method for Approximate Inner-Product Computation
Abstract
Modern deep learning systems increasingly rely on high-dimensional vector operations, in which large input and parameter vectors are repeatedly stored, transferred, and processed. Although substantial research has focused on reducing neural-network model size through pruning, quantization, knowledge distillation, and efficient architectures, the storage, memory-traffic, and communication costs associated with repeatedly presented high-dimensional input vectors remain important practical considerations, particularly in distributed and edge-computing environments. This paper proposes the SETAREH (Sorting-assisted Energy-concentrating Transform for Approximate Representation using Haar wavelets) transform, a Haar-wavelet-based sparsification approach for approximating the inner product between a fixed high-dimensional coefficient vector and successive input vectors. By applying the same sequence of permutations and Haar transforms to both vectors, SETAREH preserves the exact inner product before truncation. It then retains only a subset of the transformed coefficient-input pairs, reducing their representation cost while introducing a controllable approximation error. To evaluate SETAREH, we consider two complementary experimental settings: a synthetic Gaussian-vector baseline and a real apparent-age estimation task using a convolutional neural network on the APPA-REAL dataset. The results show that SETAREH can approximate the evaluated inner products using substantially fewer transformed coefficient-input pairs than the original vector dimension. The findings support its use as a potential structured data-representation method for repeated inner products with fixed weights, while the practical benefit depends on approximation accuracy and the cost of storing or transmitting permutation metadata.
cs.LG / 74 / 2610.23343
How Finite-Rank Bottlenecks Shape the Low-Rank Adaptation Landscape
Abstract
Low-rank adaptation (LoRA) has become the standard method for parameter-efficient fine-tuning of large pretrained models, yet theoretical explanations for why low-rank updates suffice remain incomplete. Most existing analyses rely on the Neural Tangent Kernel (NTK) regime, which linearizes the network around the pretrained weights. In this work, we take a different approach that requires no linearization: our key observation is that, once the backbone is frozen, the loss sees each adapted weight matrix only through a finite-rank linear map induced by the frozen activations. Trace-norm regularization then forces the optimizer to pick a low-rank matrix among all matrices that produce the same observed output. This yields global minimizers whose ranks are controlled by the rank of this map, which we call the \emph{bottleneck rank}, rather than the sample size. For multi-head attention, we obtain tighter rank upper bounds by exploiting two additional invariances: row-wise softmax is insensitive to additive row constants in the score matrix, and value updates are observable only after attention weighting and frozen output projection. For the factorized LoRA objective with weight decay, we prove a complementary result on the nonconvex side: after an arbitrarily small generic positive semidefinite perturbation, every first-order stationary point is rank-deficient once the LoRA rank exceeds a threshold determined by the bottleneck rank. Finally, we show that global minimizers of the perturbed objective are near-optimal for the original trace-norm problem, and we provide uniform generalization bounds governed by the spectral structure of the frozen features.
cs.LG / 75 / 2610.23353
RTSM: Adapting Recursive Models to Multivariate Time-Series Classification
Abstract
Accurate multivariate time-series classification requires extracting temporal features and integrating information across a sequence. Weight-tied refinement offers a way to repeatedly process these features without allocating new parameters at each stage. Applying this principle to time series requires choosing both a temporal representation and a latent-state organization suited to classification. We introduce the Recursive Time-Series Model (RTSM), which combines a learned convolutional temporal tokenizer, single-state weight-tied refinement, and a pooled classification readout. Across all 30 UEA datasets, RTSM achieves 72.81% mean accuracy under validation-based configuration and checkpoint selection, the highest observed mean among the 17 evaluated deep-learning models and 1.60 percentage points above MambaSL. Ablations on eight datasets favor the proposed tokenizer over a simple convolutional embedding and the single-state model over the tested two-state variants. The complete classifier reaches 86.10% mean accuracy versus 84.79% for a separately trained tokenizer-only classifier. These findings identify temporal encoding and state organization as key design choices for adapting recursive architectures to time series. The resulting classifier combines competitive accuracy with a shared refinement module whose parameter count is independent of computational depth.
cs.LG / 76 / 2610.23394
Forecasting Voltage Time Series under Anomalies Using a WGAN Model With Lipschitz Control and Progressive Kernel Critic Strategy
Abstract
In time-series forecasting tasks, the use of recursive forecasting on high-dimensional data substantially increases the complexity of the prediction process. The main challenge is that long-term dynamics becomes harder to model, since the sequences show greater temporal variability. To counteract this limitation and improve the modeling of long-term dynamics, and more importantly, address high-dimensional recursive forecasting, this paper uses a Wasserstein GAN equipped with a progressive Lipschitz penalty and an adaptive kernel. This model is capable of taking a window of the data stream and forecasting the next values, subsequently reconstructing the time-series stream and dynamically adapting its kernel to each new window. The proopsed model is a deep learning framework based on a Wasserstein GAN (WGAN-CLP) with an adaptive kernel to model alterations in streaming voltage time series. To establish the methodological foundation, we review the WGAN-LP formulation another variant model in this class and extend its framework through an adaptive kernel critic model. We then evaluate this extended model using power ratios, correlation coefficient (CC), and relative root mean squared error (RRMSE). The experimental results demonstrate that the proposed model outperforms benchmark methods in both accuracy and robustness. The evaluation shows promising performance of the WGAN-CLP with an adaptive kernel framework in forecasting sequence alterations across the voltage and energy time-series datasets.
cs.LG / 77 / 2610.23402
Contracting with an Agent under Model Mismatch
Abstract
In the classical principal-agent setting, the principal delegates some task to the agent, who receives a payment conditioned on the outcome of the task, which in turn benefits the principal. To align their objectives, the principal designs a contract that incentivizes the agent to take actions beneficial to the principal. In practice, however, the agent may act under misspecified beliefs about the underlying environment, leading to systematically suboptimal decisions. In this work, we study this problem through a principal-agent Markov Decision Process (PAMDP), in which the principal repeatedly interacts with an agent whose transition model may differ from the true environment dynamics. The principal has access to the true transition model and designs contracts that account for the agent's belief and incentives. We characterize contract design under model mismatch and develop a framework that enables the principal to induce desirable long-term benefits even though the agent has a potentially inaccurate understanding of the environment.
cs.LG / 78 / 2610.23409
Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation
Abstract
The recurrent neural network transducer (RNN-T) remains a practical choice for streaming speech recognition, but training requires large intermediate tensors. We present Flash RNN-T, a joint and loss implementation combining specialized kernels with optional tiled recomputation over the full alignment space and vocabulary. Tiled and untiled Flash form the entire measured throughput-memory frontier on medium and long workloads. On H100, untiled Flash achieves 24-39% higher throughput with 26-56% less memory than the fastest baseline on these workloads wherever it fits. Tiling reduces memory further through recomputation. On medium, tiled Flash uses ninefold less memory than that baseline at 16% lower throughput when both use one backward call per batch. On long, it fits where all tested alternatives with that backward schedule run out of memory with 80 GB. These comparisons use each implementation's lowest-memory configuration reaching at least 95% of its peak throughput.
cs.LG / 79 / 2610.23421
Residual Stochastic Concept Bottleneck Models
Abstract
Concept-based models (CMs) learn human-interpretable intermediate representations that enable users to understand the model's reasoning and steer its predictions through test-time concept interventions. However, in realistic settings, their performance can be limited by concept incompleteness, where the training concepts fail to capture all task-relevant information. Adding an unconstrained residual pathway to the model's predictive process can overcome the consequences of incompleteness. However, this introduces an inscrutable, leak-prone prediction pathway that may weaken concept interventions, as these often leave residual variables unchanged. To address this limitation, we introduce Residual Stochastic Concept Bottleneck Models (Res-SCBMs), a family of concept-based models that jointly model concept and residual variables through a multivariate Gaussian distribution. Res-SCBMs learn a covariance structure that explicitly captures dependencies between the known concepts and learnt residual variables, allowing interventions to propagate to statistically dependent residual variables through Gaussian conditioning. This extends the intervention capabilities of previous CMs without requiring concept completeness during training. Our results show that, across multiple concept-incomplete benchmarks, Res-SCBMs improve predictive performance over competing baselines while retaining effective test-time concept interventions.
cs.LG / 80 / 2610.23434
Bridging Compressed Communication and Local Steps via Spectral Method in Federated Learning
Abstract
Federated Learning (FL) enables collaborative training of complex models across large-scale networks of edge devices while keeping data localized. However, data heterogeneity and high communication overhead remain fundamental bottlenecks. Recently, the spectral optimizer Muon has demonstrated remarkable advantages over Adam-like methods in centralized training. While recent efforts have attempted to transition Muon to FL, naive integrations into traditional architectures fail short. Crucially, existing approaches do not support the the communication compression and local steps --- the two foundational techniques of efficient FL. To address this limitation, we introduce a novel spectral FL method that utilizes both local updates and compression. We establish rigorous convergence guarantees under mild assumptions, notably eliminating the restrictive bounded heterogeneity condition. Empirical evaluation on standard image classification benchmarks validates the efficiency and robustness of our approach.
cs.LG / 81 / 2610.23445
RECREG: Recovery-based Shielding for Regular Safety Properties in Reinforcement Learning
Abstract
In many real-world scenarios we need to consider formal requirements, such as safety and constraint compliance, to deploy Reinforcement Learning (RL) agents. In this paper we introduce RecReg, a safe RL framework capable of enforcing non-Markovian regular safety properties via recovery-based shielding. RecReg is a unified approach that supports different levels of a priori knowledge and safety guarantees. The paper primarily focuses on building the theoretical foundations for RecReg in the tabular setting, providing rigorous safety and convergence guarantees that are not present in related work. We provide an extensive set of empirical results using the MASA-Safe-RL library, and we outline an extension of RecReg to deep RL domains using SOTA model-based RL. Our empirical results demonstrate that RecReg achieves efficient learning and matches the theory in this paper.
cs.LG / 82 / 2610.23448
Regularized free deconvolution for spectral repair of resampling distortion in high-dimensional particle filters
Abstract
Resampling is the standard remedy for weight degeneracy in particle filters, but in high dimensions it distorts the spectrum of the ensemble covariance in a way that is computable from the importance weights. This paper studies the inverse problem: recovering the pre-resampling spectrum from the observed weights and the empirical spectrum alone, and using it without biasing the filter. We first show that the forward spectral law extends to random importance weights once one conditions on the weights together with the realized offspring counts, rather than on the weights alone. We then construct a regularized free-deconvolution estimator, prove its consistency with an explicit error budget at a fixed resolution scale, and certify model mismatch through a computable forward residual. The repaired spectrum enters the filter only as the preconditioner of a Metropolis-adjusted rejuvenation move, so that spectral error affects mixing efficiency but not the invariant target. For the full filtering cycle, we derive a geometrically summable error recursion under stable scalar dynamics and an explicit law for the cost of abstaining from repair. Pre-registered simulations show that the inversion reduces the spectral error by an order of magnitude, that the corrected preconditioner recovers about 90% of the oracle's sampling efficiency in an anisotropic regime, and that the cost of abstention depends on when abstentions occur rather than on how many. The repair recovers eigenvalue laws, not eigenvectors, and is certified only within an explicitly stated exogenous-weight regime.
cs.LG / 83 / 2610.23453
Predictable Scaling: Efficient Model Exploration via Scaling-Law Prediction
Abstract
Evaluating candidate changes (hypotheses) to recommendation models requires full-scale training runs consuming thousands of GPU-hours, severely limiting the hypothesis space engineers can feasibly explore. We introduce Predictable Scaling (PS), a systematic framework that exploits empirically verified regularity conditions to predict full-scale model performance from small-scale proxies. PS constructs scaled-down variants via composable scaling functions along distinct resource axes, and projects metrics to the target scale via curve fitting with calibrated uncertainty estimates. We compare three prediction strategies of increasing expressiveness: even the simplest (raw proxy) achieves prediction error within the empirical noise floor of repeated-run variance; the most expressive (Gaussian Process) provides an indicative 15% relative MAE improvement and calibrated uncertainty. Evaluation on two ranking models of increasing complexity---a task-specific model and a foundation model---demonstrates that PS correctly ranks actionable hypotheses across diverse hypothesis types at 79-88% compute reductions.
cs.LG / 84 / 2610.23491
A Topological Signature of Learning under Noisy Labels
Abstract
How does noisy supervision change the representations learned by a neural network? Training accuracy measures agreement with training labels but provides little information about representation geometry. We use persistent homology to track loop structure across distance thresholds in neural activations. Experiments span image classification, text classification, and modular arithmetic. On a fixed CIFAR-10/ResNet-18 training probe, equally sized clean-label and corrupted-label subsets show no detectable persistence difference at epoch 10, whereas corrupted-label subsets have higher persistence at epoch 50. The contrast survives uniform-scale controls, even though predictions match the assigned noisy labels for only 7-11% of corrupted probe examples at that checkpoint. Normalisation preserves the crossover between clean and noisy training trajectories on CIFAR-10/ResNet-18, whereas CIFAR-100/ResNet-18 shows no crossover under any of the five evaluated constructions. In a separate noise-rate calibration comparison across five pipelines, mean activation norm achieves a higher mean in-sample R2 than raw persistence (0.936 versus 0.908). These findings show that noisy supervision leaves a measurable geometric signature even when agreement with corrupted labels is low. Comparing persistence under scale controls with label-fitting measurements offers a way to study how representations respond to label corruption beyond what training accuracy reveals.
cs.LG / 85 / 2610.23515
Uncertainty-aware AI-enabled framework for predicting hydrogen dispersion in confined spaces to enhance operator situational awareness
Abstract
We present a two-stage AI-enabled framework for enhancing operator situational awareness during compressed hydrogen bunkering in maritime confined spaces. In the event of a leak, the framework first estimates the unknown leakage rate from early sensor readings and then forecasts how the hydrogen cloud will disperse over the next time steps, providing operators with a predicted concentration map together with confidence bounds. The framework is based on two independent AI models building one AI system. The AI system is designed as a decision-support tool. It augments sensor readings, and thus enhances operator situational awareness and supports decision-making. This enables preventive safety barriers to be deployed before hazardous concentrations are reached, while an independent emergency shutdown system remains the ultimate safety layer. The system design incorporates safeguards to ensure the AI system operates within its competence envelope. Trained and tested on an open dataset of confined-space hydrogen releases, the framework achieves a mean absolute error of $0.052$~g/s in leakage-rate estimation and $\approx0.018$ in hydrogen-concentration prediction, approaching the sensor precision bound of $\pm0.01$. The hydrogen-concentration prediction provides well-calibrated uncertainty estimates that allow operators to judge the reliability of each forecast. These results suggest that the framework may provide early information about hazardous concentration build-up, potentially supporting preventive actions by operators.
cs.LG / 86 / 2610.23528
Out of Support, Not Out of Reach: In-Context Learning with Unsupported Instructions and Composed Tasks
Abstract
In-Context Learning (ICL) enables pretrained LLMs to adapt to downstream tasks from a small set of input-output demonstrations, without parameter updates. Yet existing theory primarily explains ICL as identifying a latent task already represented in pretraining, leaving unclear whether ICL can succeed when either the instruction or the target task falls outside the pretraining support. We develop a theoretical framework under mild assumptions that separates these two forms of distribution shift. For target tasks represented in pretraining, we show that demonstrations can identify the task and reduce sensitivity to pretrain-supported prompt templates as their number increases. For an unsupported instruction, we show that under a fixed model-induced reference extension, demonstration evidence can dominate an $N$-independent instruction shift while preserving the exponential identification rate. We then use the same ICL framework as an interpretive lens for long-horizon target tasks that are not represented in pretraining but admit a plan that decomposes them into pretraining-supported subtasks. Applying the supported-task analysis step by step yields a standard accumulated-error bound and clarifies how invoking familiar subtasks along an intermediate-state trajectory can avoid directly selecting the unsupported complete task. Together, our results characterize when demonstrations, instructions, and task decomposition can support ICL beyond the corresponding pretraining supports.
cs.LG / 87 / 2610.23565
Beyond Expected Values: Risk-Sensitive Planning with Distributional Monte-Carlo Tree Search
Abstract
Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR). We introduce Categorical Thompson Sampling with Optimism (CATSO) and Particle Thompson Sampling with Optimism (PATSO), two distributional MCTS algorithms that maintain finite-support empirical backup laws at Q-edges. CATSO represents Q-edge laws with categorical atoms and Dirichlet Thompson sampling, while PATSO represents them with adaptive particles. Both algorithms select actions using the CVaR of a Thompson-sampled Q-edge law plus a polynomial optimism bonus, and propagate scalar continuation values through visit-weighted averages of child Q-edge CVaR scores. For the finite-depth nested CVaR planning target, we prove an $O(n^{-1/2})$ bound on the root value error after $n$ rollouts, which upper-bounds the risk-sensitive simple regret. CATSO adds a fixed-grid discretization bias, while capped PATSO adds a tunable Wasserstein compression bias. Experiments on stochastic planning benchmarks show that distributional Q-edges improve lower-tail return and reduce catastrophic outcomes when mean-optimal and risk-sensitive behavior differ.
cs.LG / 88 / 2610.23571
BMuon: Budgeting Muon's Spectral Geometry for Structured Compressibility
Abstract
Muon has emerged as a strong optimizer for neural network training and admits an interpretation as steepest descent under a spectral-norm geometry. Yet we find that Muon-trained Transformers can degrade sharply when their MLP hidden channels are pruned after training. To reconcile Muon's optimization performance with structured compressibility, we introduce Budget Muon (BMuon), which modifies Muon's spectral geometry so that hidden channels across blocks compete for a shared allocation budget. The resulting oracle is exactly normalized steepest descent under a family of norms that interpolate between Muon's geometry and structured sparse geometries. Our principal sparsity control is an interpretable global channel-allocation budget, rather than a setting-dependent sparsity-penalty coefficient. For linear factorization models, we show that Muon admits minimum-cost exact representations with redundant hidden width, whereas under the BMuon oracle geometry and a sufficiently scarce budget, every minimum-cost representation is width-minimal. For practical training, we derive a minorization-maximization surrogate for the exact allocation problem. On ImageNet-1K with ViT-S/16, BMuon improves over Muon under gradual structured MLP pruning, and its learned allocations provide a more effective pruning criterion than weight magnitude. On a 1B-parameter LLaMA-style model trained on FineWeb, BMuon improves over Muon under post-training MLP channel pruning.
cs.LG / 89 / 2610.23590
Boundary Detection for High-Dimensional Point Clouds
Abstract
We introduce two new methods for boundary detection in high-dimensional point cloud data, intended to work in situations with non-uniform density, high ambient dimension, and substantial noise. Our first method fits a separating hyperplane on local neighborhoods after doing local dimension reduction---this method works best when one can reliably estimate the intrinsic dimension of the data. Our second method measures the directional uniformity of local neighborhoods, and has the best overall performance on synthetic data. In addition to pointwise numerical boundary scores, our algorithms yield an estimated normal vector at each boundary point. We validate our methodology on synthetic data as well as two real examples: we analyze spatial transcriptomics data, where we show that we can detect the boundary of a tumor, and we also study LLM activation geometry.
cs.LG / 90 / 2610.23608
When Magnitude Suffices: Reconstructing Time-Domain-Gated Microwave Signals Without Phase on a Thoracic Phantom
Abstract
Time-domain gating separates the arrivals that form a microwave S-parameter measurement. It requires the measurement of complex values, so this application almost always uses a vector network analyzer (VNA). We show that, on one calibrated measurement configuration, the gated magnitude the VNA would give can be reconstructed from a magnitude-only sweep of a spectrum analyzer with a tracking generator (SA). We select the gate center for separability of the measured states, correct each instrument once by a per-bin subtraction of a reference state, and train a fully connected network on sweeps from both instruments. The test object is a torso phantom with lung-shaped compartments whose filling takes eleven dielectric states, each held out in turn. On $S_{11}$ the reconstruction from SA input reaches a shape error of 0.08 dB$^2$ with a state ordering of 0.81. This matches the reconstruction from VNA input. On $S_{21}$ it reaches 0.82 dB$^2$ and keeps the ordered groups of the states. An explicit phase prediction performs much worse, because gating needs the phase to about 0.2 rad RMS. Across eleven measurement configurations the dependence of phase on magnitude is strong for reflection and weak for transmission. This explains the difference between $S_{11}$ and $S_{21}$. Where the magnitude carries enough of the phase, a scalar instrument gives the gated response without any phase measurement after installation. On the open UM-BMID breast data set, with whole phantoms held out, the shape error of the network is 30 percent lower than that of a per-bin linear map.
cs.LG / 91 / 2610.23611
How Many, but How Sure? Class Competition as a Confidence Proxy in Multi-Class Density-Based Counting
Abstract
Multi-class density estimators can count objects in crowded scenes, but their class-wise outputs offer no indication of how reliably density mass has been allocated. We introduce a training-free confidence proxy based on local competition between the strongest class channels. The resulting prominence measure requires only the predicted density maps, with no retraining or access to model internals. We test whether residual competition is diagnostic of cross-class leakage by attenuating contested density at inference time. On DSACA predictions for the Hicks et al. dataset, this reduces overall MAE from 0.920 to 0.846, suggesting that class competition contains useful information about count reliability.
cs.LG / 92 / 2610.23616
The Sign Code: The Hidden Binary Nature of Deep Networks
Abstract
Every neuron in a trained network outputs a number that can be separated into its sign (positive or negative) and its magnitude (absolute size). We call the set of all neuron signs the \textit{sign code}. This raises a key question: does classification depend more on the sign code or on the magnitudes? For overparameterized vision classifiers with sign-preserving activations, the answer is overwhelmingly the sign code. A $k$-nearest neighbor classifier operating only on this binary sign code achieves accuracy comparable to the full model across different architectures, activation functions, and datasets. Moreover, replacing all magnitudes in every layer with random noise while keeping the original signs barely changes residual network performance. In contrast, flipping just 20\% of the signs reduces accuracy to chance. Models generalize well when same-class inputs share similar sign codes. When this overlap is disrupted by a regularizer, training accuracy remains high, but test performance collapses. A margin decomposition helps explain these results: the sign-only contribution to each logit grows as $\sqrt{d}$ in network width, dominating the magnitude residual. The natural unit of analysis for these networks may not be through their continuous weight space but instead through their discrete sign codes.
cs.LG / 93 / 2610.23618
ThetreStageAI: Multi-View Retrieval-Augmented Generation for Structured Bengali Theatre Production Planning
Abstract
Staging a theatre performance requires coordinating dialogue, actor movement, stage directions, lighting focus, cue timing, and equipment constraints within the same production plan. This work presents ThetreStageAI, a retrieval augmented framework for structured Bengali theatre production planning. Rather than retrieving from one mixed corpus, the system maintains separate scene, actor-blocking, and lighting views. Bengali text is embedded with BGE-M3 and indexed in view-specific Qdrant collections. Retrieved references are combined with a production brief and passed to Gemini, which generates schema-constrained JSON containing characters, scenes, dialogue, stage directions, blocking, lighting, and optional sound cues. Structural and semantic checks are applied before acceptance, with at most one constrained repair attempt. We also construct a human-guided, LLM-assisted synthetic dataset of 500 short Bengali theatre records that expands to 1,500 retrieval documents. Dataset diagnostics confirm consistent identifiers and valid lighting fields while also revealing substantial dialogue repetition. A paired camera-ready evaluation compares single-combined and full multi-view retrieval under matched settings. The source code, dataset, and evaluation materials are publicly available.
cs.LG / 94 / 2610.23620
GRAFT-EEG: Gauge-Aware Physical Field Transport for Cross-Dataset Motor-Imagery EEG
Abstract
Cross-dataset EEG decoding often treats montage, reference, and active-sensor changes as representation shift, although they also alter the observation operator. We study target-statistics-blind transfer: fitting receives no target signal, label, or empirical statistic, and electrode coordinates only instantiate a fixed prediction operator. method removes the active-set common-reference gauge and transports each native mask through a template motor leadfield. A source-fitted spectral classifier anchors the decision; a bounded state-space-model (SSM) residual is secondary. We establish restricted identifiability and an operator-perturbation bound. Across five leave-one-dataset-out targets (253 participants), dynamic method achieved 60.865% equal-target balanced accuracy versus 52.684% for strict-clean 5/5 BIOT (hierarchical difference +8.182 points; 95% CI +6.382 to +10.196; positive in every target). On four EEGPT-clean targets, the descriptive means were 61.520% and 59.057%. In four post-hoc exact same-parent tests, FULL exceeded coordinate-basis, matched-spectrum, row-shuffled, and wrong-anatomy controls by 9.959, 9.698, 11.241, and 3.985 points, respectively (all 95% CIs excluded zero; Holm-adjusted P<0.001). A separate 60-participant same-paradigm evaluation retained a positive physical-operator contrast (+4.029 points; +2.338 to +5.773), whereas dynamic--static was inconclusive; its preregistered status is partial confirmation rather than unrestricted external validation. The SSM added +1.059 points (+0.013 to +1.892). Claims remain bounded to five binary motor-imagery domains and the tested operators.
cs.LG / 95 / 2610.23623
Deconstructing Multi-Task Active Learning: The Paradox of Gradient Conflict and Orthogonal Decomposition
Abstract
Existing multi-task active learning (MTL-AL) methods predominantly rely on linear scalarization of heterogeneous objectives such as uncertainty, diversity, and gradient conflict. We argue that this entangled formulation is fundamentally limited, as linear combination fails to capture Pareto-optimal trade-offs under conflicting objectives. We propose a shift from combination to deconstruction and introduce a principled framework that rethinks MTL-AL along three orthogonal dimensions. First, we identify the Conflict Paradox: although the optimization literature treats gradient conflict as interference to be suppressed, penalizing conflict when ranking candidates is measurably worse than not scoring it at all. We therefore decouple the two concerns: the acquisition function does not score conflict, and the resulting interference is resolved by gradient surgery during training. Second, we introduce Orthogonal Decomposition, separating acquisition into Informativeness and Exploration. This two-stage process, informativeness filtering followed by coverage maximization, prevents "diversity dilution" by excluding uninformative outliers before enforcing diversity. Third, we propose Cross-Task Mutual Information (CTMI), a covariance-based statistic of cross-task dependency computed from MC-dropout task entropies. Extensive experiments on COCO, Cityscapes, and NYUv2, including comparisons against recent MTL-AL and data-selection baselines, demonstrate consistent improvements over state-of-the-art methods, highlighting that structural rethinking of acquisition yields larger gains than incremental heuristic design. Code is available at https://github.com/Juho-Jung/multitask_active_learning_official.
cs.LG / 96 / 2610.23641
No-Regret Language Generation
Abstract
We introduce no-regret generation, which requires a generator to make a sublinear number of mistakes uniformly over all demonstration sequences and compatible target languages. We also formulate an agnostic variant that allows arbitrary demonstrations that may not be compatible with any languages. For both settings, we introduce tree-based mistake-forcing dimensions that exactly characterize finite-horizon minimax complexity and establish necessary and sufficient conditions for no-regret generation. Our characterizations reveal a rich spectrum of generation rates: up to constant rescaling, every admissible nondecreasing profile is attainable in both the realizable and agnostic settings. Finally, we connect generation complexity to online prediction: for language classes of finite Littlestone dimension \(d\), we obtain realizable mistake bounds of \(O(d\log(1+T/d))\) and agnostic regret bounds of \(O(\sqrt{dT\log(1+T/d)})\). The realizable bound is optimal up to universal constants.
cs.LG / 97 / 2610.23650
JAXtari: High-Throughput and Easy-to-Modify Arcade Learning Environment
Abstract
The Arcade Learning Environment (ALE) has been a cornerstone of Deep Reinforcement Learning research, providing a standardized benchmark for evaluating general agents. However, its lack of native support for object-extraction and environment modification restricts research into evaluating whether agents learn robust causal models or rely on brittle visual heuristics. Furthermore, the ALE's dependence on CPU-based emulation creates a significant bottleneck in modern GPU-accelerated training pipelines, limiting both experimental scale and iteration frequency. In this work, we introduce JAXtari, an open-source, high-performance reimplementation of Atari 2600 environments. By leveraging massive GPU parallelization, JAXtari allows PPO agents to finish training on 100M environment steps on a single A100-GPU in under an hour. Beyond speed, it provides object-centric observations and a flexible modification system, allowing researchers to, e.g. develop neuro-symbolic methods or discover novel shortcuts and subsequently train and adapt their agents rapidly, sometimes in only minutes.
cs.LG / 98 / 2610.23665
TriRouter: Sparse-to-Dense Electrode-Space Routing for Robust EEG Motor Imagery Decoding
Abstract
Motor imagery brain-computer interfaces require reliable decoding from multichannel EEG, but practical acquisition often yields sparse or corrupted electrode montages. We propose TriRouter, an electrode-space routing framework with discriminative-preservation, mask-robustness, and latent-compensation paths. A validation-calibrated router uses observable channel availability to allocate path contributions, preserving complete-channel discrimination while increasing representation recovery under severe missingness. In BNCI2014-001, TriRouter outperforms standard public backbones and a dynamic-spatial-filtering corrupted-EEG baseline. It improves accuracy over the strongest corrupted-EEG baseline by 3.68, 4.02, and 6.46 percentage points at nominal 25%, 50%, and 75% missing settings, and remains effective under structured regional missingness and additional low-channel dataset evaluation.
cs.LG / 99 / 2610.23700
Equivalence testing and run-count design for stochastic machine-learning experiments
Abstract
Claims that two training methods perform equally well often rest on a non-significant difference between two or three runs per method. We develop a protocol, with its power theory, for certifying practical equivalence in per-run metrics such as the AUROC (area under the receiver operating characteristic curve). It rests on a two-level (seed x item) variance model separating between-run dispersion from item-level noise. Equivalence within a pre-specified margin is tested by two one-sided tests (TOST), equivalently by a confidence interval inside the margin; a two-level bootstrap covers composite statistics. A closed-form formula converts a margin and a target power into runs per method; its shifted-t refinement is calibrated against the realised two-level process. Simulations on a 270-cell grid and under non-Gaussian seed laws confirm the formula within its design-relevant range, from five runs per method. For four published runs of a joint energy-based model, the failure to certify equivalence was largely decided by design: with the variance calibrated to those runs, two per method certify with probability near three per cent, whereas certifying 0.01 AUROC at 80 per cent power needs twelve even under exact equality. A second case study on a thousand-seed CIFAR-10 zoo and the MultiBERTs checks the formula against real seed laws (within 0.05 in that range) and resolves comparisons that two or three runs leave undecided. In a two-stage design (pilot, then formula), our experiments suggest that an upper confidence bound on the pilot variance substantially reduces the undersizing seen with its point estimate.
cs.LG / 100 / 2610.23707
Tool Recommendation for Multi-Step Agents: Finding Every Tool a Task Needs
Abstract
A multi-step agent task fails when a single tool it needs is missing, yet the tools an agent may need are too many to list: the roughly 8,200 tools on the platform we study have definitions that together exceed the context window of any model we evaluate. We present Tool Recommendation, a deployed system that gives an agent all of them, hundreds of times more than a typical MCP server exposes, through a fixed interface of three tools. It runs at the gateway between agent and tools, executing every call under the end user's own credentials and checking it against the user's consent for each service and the tools the agent may use. We evaluate it on both a public retrieval benchmark and billed end-to-end session cost. On ToolRet, Tool Recommendation places every tool a query needs in the top ten 57% more often than BM25, a lexical ranking that commercial tool search offers, and compares favorably with published systems up to nine times its size. In paired multi-step workflows, its first recommendation covered every required tool in 93% of Codex sessions, and against a fixed list of 100 tools it cut input tokens by 17% with Codex and 40% with Claude Code and the total bill in all 14 runs, by up to 40%, with no change in task completion.
cs.LG / 101 / 2610.23716
Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize
Abstract
Deep neural networks demonstrate remarkable generalization despite being highly over-parameterized, challenging traditional statistical learning theories. While several theoretical paradigms such as margin-based bounds, PAC-Bayesian analyses, and neural tangent kernels provide global explanations, they often lack mechanistic insight into how generalization emerges from training dynamics. We introduce Neural Expansion, a unified geometric framework that characterizes generalization through the expansion of directionally-selective ray coverage in the input space. Using Generalization Intervals (GIs), we quantify direction-specific regions around training samples where the model's predictions remain stable. Our extensive empirical studies across diverse architectures and datasets reveal that training leads to progressive expansion of ray coverage, particularly along directions with low Jacobian singular values. This mechanism helps explain a large fraction of correctly classified test inputs, adversarial examples, and even mislabeled or out-of-distribution samples through the geometry of ray coverage rather than instance-level memorization. By connecting input-space curvature to predictive behavior, our framework provides a mechanistic foundation that complements existing global theories of deep learning.
cs.LG / 102 / 2610.23729
Escaping Parameter Space: Tight Generalization Bounds via Representation Quality
Abstract
We derive a tight generalization bound for deterministic neural We derive a tight generalization bound for deterministic neural networks based on their last hidden layer representations. By evaluating representation quality with a parameter-free $k$-nearest-neighbor ($k$-NN) classifier, our certificate avoids traditional parameter-space complexity measures, KL divergences to a posterior, weight norms, and compression. The bound is tight: trained from scratch, it reaches an 8.0 percentage point gap to test error on CIFAR-10 (ResNet-18) and yields the first non-vacuous certificates we are aware of on CIFAR-100; with frozen DINOv2 representations, it certifies at 1.4\% on CIFAR-10 and 23.1\% on 1,000-class ImageNet. It decomposes additively into three empirical, independently measurable quantities: local class geometry ($k$-NN decoder error), decoder-network agreement, and training stability. This allows the bound to act as a diagnostic tool: we demonstrate that actively destroying local class geometry eliminates generalization while leaving dataset memorization intact, identifying class geometry as a causally relevant mechanism for generalization. Finally, we show that learned representations are significantly less sensitive to weight noise than linear readouts, suggesting why parameter-space certificates often struggle to remain tight.
cs.LG / 103 / 2610.23735
LagNet: Lag-Structured Operators for Long-Term Time Series Forecasting
Abstract
Long-term time series forecasting maps a finite history window to an extended future horizon. Many existing models focus primarily on the design of input representations and model structures, while paying less attention to the structural form of the history-to-future forecasting operator itself. We examine this operator directly and find that, across model families, sensitivity matrices exhibit recurring diagonal patterns whose diagonals correspond to relative temporal lags. Motivated by this observation, we develop a framework for direct multi-step forecasting based on lag-structured temporal operators. Its basic building block, LagLayer, constructs a Toeplitz matrix by tying history-to-future connections with the same relative lag and incorporates horizon-wise and history-wise scaling coefficients to relax exact sharing. This design enables LagLayer to serve as an effective replacement for temporal linear layers in existing forecasting models. Stacking LagLayer modules with activations and multiple feature maps yields LagNet. Experiments on long-term time series forecasting benchmarks show that our method is competitive with strong baselines, while sensitivity matrix diagnostics and limited-data experiments demonstrate that LagLayer provides an effective and sample-efficient alternative to a linear layer.
cs.LG / 104 / 2610.23765
Universal Semantic Enhancement for Weakly Supervised Anomaly Detection
Abstract
Weakly supervised anomaly detection on tabular data is an essential task in many label-scarce applications, since tabular data is widely encountered and it is commonly observed in practice that only a limited number of labeled anomalies are available alongside abundant unlabeled samples. Most existing methods try to learn features that capture the inherent distribution patterns in the target domain. However, such features often struggle to provide a complete description of possible abnormal characteristics and are highly susceptible to overfitting the extremely scarce labels. To address this challenge, we propose a weakly supervised anomaly detection method based on universal semantic enhancement for tabular data, which fuses the universal knowledge from a pre-trained tabular foundation model with the domain-specific knowledge learned on the target dataset. The two types of knowledge are naturally complementary: universal semantics describe the common patterns of tabular data across domains, while the domain-specific features capture the inherent distributions of normal and anomalous behaviors in the target dataset. A prototype-conditioned modulation mechanism is then designed to effectively integrate the two types of features, allowing them to reinforce each other for improved anomaly detection. Extensive experiments on public datasets show that the proposed method outperforms competitive methods when only eight labeled anomalies are available in the training procedure.
cs.LG / 105 / 2610.23811
Explainable Machine Learning Analysis of Temperature Effects on Mechanical Strength of Fiber-Reinforced High-Performance Concrete
Abstract
Predicting the compressive strength of fiber-reinforced high-performance concrete subjected to temperature exposure is challenging because to the intricate interdependencies of mixture composition, fiber properties, curing conditions, and temperature. A data-driven framework to predict compressive strength utilizing 1,204 experimental records containing a diverse array of material designs and exposure situations. This study evaluates eight machine-learning models and demonstrate that ensemble approaches routinely surpass individual learners. Both gradient boosting and extreme gradient boosting exhibit superior performance on unseen data, each attaining a R^2 of 0.903. Gradient boosting produces the minimal root mean square error (6.67 MPa), whereas extreme gradient boosting achieves the lowest mean absolute error (3.97 MPa) and mean absolute percentage error (10.60%). The models are further interpreted utilizing Shapley additive explanations and partial dependence analysis, which reveal temperature as the primary factor influencing compressive strength, succeeded by fine aggregate, superplasticizer, cement content, water-to-binder ratio, and water content. Strength diminishes significantly with rising temperature, albeit the extent of deterioration is contingent upon the specific combination design. The best performing models are subsequently implemented in a Python-based graphical user interface for expedited prediction. Our findings offer a precise and comprehensible methodology for evaluating temperature-dependent compressive strength and facilitate the practical, data-driven design of fiber-reinforced high-performance concrete.
cs.LG / 106 / 2610.23812
Learning Personalized Memory Gating from Delayed One-Sided Feedback
Abstract
Long-term memory enables language-model agents to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. Existing memory gating policies are typically shared across users and therefore fail to account for user-specific patterns of future reuse. In this paper, we study memory gating at the fact level. We first propose a retrieval-based labeling strategy that labels each fact by whether it is later retrieved and incorporated into responses. We then formulate online memory gating as contextual apple tasting with delayed one-sided feedback, and propose Warm-Start UCB, which combines offline learning from existing data with online adaptation to the current user. Our theoretical analysis establishes a high-probability regret upper bound of $\tilde{O}(Tε\wedge d\sqrt{T})$ when the delay window is at most $O(\sqrt{T})$, where $T$ is the total number of rounds, $ε$ is the initialization error, $d$ is the dimension of the feature map, and $\tilde{O}(\cdot)$ hides constant and poly-logarithmic factors. The results of our experiment demonstrate that our method outperforms all baselines under different memory budgets and history lengths, establishing it as a promising personalized memory gating policy.
cs.LG / 107 / 2610.23820
CocoRNA: Collective RNA Design with Cooperative Multi-agent Reinforcement Learning
Abstract
Designing RNA sequences that reliably fold into specific secondary structures is essential for understanding their biological functions but remains a challenging computational problem. We propose CocoRNA, a cooperative multi-agent reinforcement learning framework for RNA inverse design. CocoRNA simplifies the design task by decomposing it into smaller sub-problems, each solved collaboratively by multiple agents. This approach reduces the complexity of the problem and improves the exploration of design policies. During training, a centralized critic uses global structural information to guide the agents, enabling them to jointly optimize their design strategies. As a result, CocoRNA learns high-quality RNA design policies that generalize effectively to unseen structures without additional training. Experiments on the Rfam dataset demonstrate that CocoRNA substantially outperforms state-of-the-art methods in both success rate and design speed. Further experiments on other biological sequence design tasks highlight the effectiveness and broad potential of CocoRNA for complex design tasks.
cs.LG / 108 / 2610.23830
Waiting Buys Exactly Two Moments: Endogenous Identification in Prediction-Driven Allocation
Abstract
A planner with a fixed budget of preventive help must decide whom to help and when. Waiting reveals risk but removes intervention opportunities as higher-risk individuals leave first. We characterize this tradeoff under repeated noisy signals and risk-dependent attrition: in the canonical model, a length-$τ$ observation prefix identifies exactly the latent risk moments $m_j=\mathbb E[p^j]$ for $j=0,\ldots,2τ-1$, buying two new moments per additional period. These identify average next-period attrition among survivors but need not identify allocation welfare or attrition among those prioritized for treatment. Our observationally equivalent constructions establish unavoidable timing regret at any sample size for a fixed observation length, while exact certificates determine whether all compatible populations share a unique optimal date. Matched experiments separate the effects of estimation, bootstrap averaging, and regret weighting. Longer observation often resolves timing ambiguity, but the optimal date may already have passed. In our experiments, bootstrap averaging and early stopping improve decisions, while regret weighting adds no improvement in pooled mean squared regret: better statistical fit need not yield better allocation welfare. The code and artifacts are available at https://github.com/KurbanIntelligenceLab/waiting-buys-two-moments.
cs.LG / 109 / 2610.23294
The Optimal Ranking Tradeoffs Between Two Ranking Scores in Two-Class Classification
Abstract
The accuracy, true negative rate (negative class recall), true positive rate (positive class recall), negative predictive value (negative class precision), positive predictive value (positive class precision), class-specific intersections over unions, class-specific F-scores, Cohen's kappa, and balanced accuracy are scores that one intuitively wants to maximize when developing machine learning methods. However, they lead to different rankings of such methods. So, how can we establish a ranking that is an optimal tradeoff between any pair of these scores? It turns out that all these scores are special cases of ranking scores: a continuum of scores parameterized by the relative importance values assigned to the true negative, false positive, false negative, and true positive outcomes. Since all these scores are known to induce meaningful rankings, we seek the optimal tradeoffs among them. We consider the tradeoff between two scores to be optimal when it minimizes the Fréchet variance regarding the Kendall distance (also known as the bubble-sort distance) between rankings. For any given set of classifiers to rank, the optimal tradeoff lies at the center of the shortest (geodesic) path on the manifold of all rankings induced by the ranking scores. However, in the parameter space of the ranking scores, it is rarely achieved by averaging the importance values corresponding to the initial scores. In this paper, we present an exact analytical solution for determining a ranking score that is an optimal tradeoff for any given pair of initial scores, in the common case where the class priors are the same for all compared performances. We also experimentally investigate what happens when performance follows some distributions with fixed priors.
cs.LG / 110 / 2610.23383
Data-Driven Prediction of Optimal MPI Rank Placement and Node Count for Memory-Bound CFD on AMD EPYC Clusters
Abstract
Users who run a memory-bound OpenFOAM case on an AMD EPYC cluster must decide how many nodes to use and how many MPI ranks to place on each core complex. The fastest choice is often not the cheapest one. This paper revisits our earlier analytical model of this performance and energy trade-off with 560 measured configurations of the motorBike case on five Zen 4 and Zen 5 EPYC processors, including socket power from processor telemetry. On up to four nodes, small meshes are both fastest and cheapest on a single node on most processors. For the largest mesh the largest node count is both the fastest and the cheapest option, because its speedup is superlinear. A genuine trade-off arises only for medium meshes, where full speed costs a median of 20 to 97% more node hours than the cheapest option. The L3 miss rate and the halo traffic per cell are consistent with this pattern. For a new mesh of the same case, the configuration that was optimal on the nearest measured mesh is near optimal in 75 to 85% of the decisions. Running the two configurations it recommends for speed and for cost raises this share to 85 to 90% for runtime and node hours and determines the trade-off with median errors of 3 to 5 percentage points. Neither gradient boosted trees trained on the same data nor the earlier polynomial model improves on this simple procedure.
cs.LG / 111 / 2610.22471
Has GPT-6-Astra Achieved Embodied Intelligence?
Abstract
GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments. In this report, we ask not only how capable GPT-6-Astra is, but what its capabilities imply for the field. First, we probe the limits of standalone GPT-6-Astra. We conduct a large-scale evaluation across 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands. Under a unified protocol that allows in-context learning (ICL) from training-set demonstrations and substantially increases the number of execution steps, GPT-6-Astra achieves unprecedented performance (Fig. 1). Future research may focus on improving 1) the generality of ICL 2) cost and execution time, 3) physical-state verification, 4) recovery from irreversible failures, and 5) long-horizon preservation of progress. Second, we study the agentic harness for embodied control. We identify that close-loop coordination between a high-level "System 2" reasoner (i.e., GPT-6-Astra) and a low-level "System 1" controller (e.g., π0.5) is costly but effective. Third, we examine GPT-6-Astra on real-world deployment. Our experiments on 8 robot configurations reveal substantial gaps: 1) robot demonstrations required for ICL are expensive to collect, and 2) it is implausible to trade longer execution for higher success rates with physical constraints. Fourth, we envision the next generation of embodied benchmarks. We develop a prototype that 1) derives tasks from real-world needs, 2) provides no in-domain demonstrations, and 3) judges agents on a world clock that keeps running while they deliberate. Taken together, our results suggest that GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.
cs.LG / 112 / 2610.22556
When Plans Change, Remember: Replanning-Guided Visual Memory for Vision-Language-Action Models
Abstract
Long-horizon, multi-stage Vision-Language-Action (VLA) policies often require visual information that is no longer present in the current observation. Retaining the complete observation history is computationally expensive and highly redundant, motivating selective visual memory. We introduce RePlanMem-VLA (Replanning-Guided Keyframe Memory for VLA), which turns a policy's own plan revisions into an online memory-write signal. Under receding-horizon execution, consecutive action chunks predict some of the same future actions. RePlanMem-VLA treats unusually large disagreement between these temporally aligned predictions as evidence of a task-relevant event. We first compare 35 disagreement measures across single-task autoregressive decoder-only and Flow-Matching Transformers trained from scratch, then validate the selected signals on pretrained OpenVLA-OFT and $π_{0.5}$ policies over MIKASA-Robo-VLA tasks. The resulting detectors achieve ROC-AUC of 0.981-0.997. We then use the selected signals in RePlanMem-VLA to trigger updates to a bounded visual memory, with a robot-masked write filter suppressing redundant frames. In the evaluated memory-retention experiments, RePlanMem-VLA preserves task-relevant observations until the later decision point in 100% of episodes. Real-robot replay data further show that event selection and redundancy suppression extend beyond simulation. These results establish plan revision as a policy-internal signal for controlling what a VLA remembers. Website: https://vika787.github.io/RePlanMem/.
cs.LG / 113 / 2610.22576
Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations
Abstract
Flow-based policies such as $π_0$ are trained on a fixed set of embodiments, and transferring them to new ones is an open problem. To adapt a frozen $π_0$ to a new embodiment without additional demonstrations or parameter updates, we train a lightweight embodiment classifier to distinguish successful action chunks on the source body from those on the target body, using a few calibration tasks. On transfer tasks that $π_0$ saw on its source body but the classifier was never trained on, we nudge each denoising step of $π_0$ with the classifier's gradient, which estimates the direction from the distribution of successful actions on the source body toward that on the target body. On the LIBERO simulation benchmark with four modified controllers, embodiment-classifier guidance raises $π_0$'s average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked wrist, when the classifier lacks data or favors infeasible motions. Collecting more target-body data with guided $π_0$ can fill such gaps: on one weak-wrist arm, guided $π_0$ solved a task on which unguided $π_0$ had zero successes in the original data, and retraining the classifier on the guided successes raised success from 52.5% to 70.7%.
cs.LG / 114 / 2610.22762
Shape-Aware Neural Time Fields for Rigid-Body Motion and Manipulation Planning
Abstract
Rigid-body motion planning is fundamental to robotic transport and manipulation, where both object position and orientation determine whether a body can move through cluttered environments. This coupling makes planning chal- lenging because translation and rotation can be constrained very differently by surrounding geometry. We propose Shape- Aware Neural Time Fields, a physics-informed neural motion planning framework that learns a reusable travel-time, or cost- to-go, representation for rigid-body motion while explicitly accounting for these distinct geometric restrictions. The learned field captures local motion difficulty in translation and rotation and is paired with a sampling-based planner that balances progress toward the goal with the cost of reaching nearby poses. We evaluate the framework across 2D and 3D rigid- body planning tasks and non-prehensile manipulation through object pushing in diverse environments. Results show substantial improvements in planning success and inference efficiency over physics-informed neural planning baselines, while remaining competitive with classical sampling-based planners at significantly lower planning times. Ablation studies further validate the contributions of both the learned representation and the proposed planning strategy. Finally, closed-loop pushing experiments under execution deviations and external disturbances demonstrate that fast replanning enables reactive and adaptive task execution.
cs.LG / 115 / 2610.23271
NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models
Abstract
Vision--Language--Action (VLA) models promise generalist robotic agents, yet consolidating skills learned across heterogeneous tasks and environments remains difficult: training data is fragmented across platforms, joint multi-task training suffers from interference, and model merging tends to introduce parameter conflicts or require architectural changes. We instead advocate a paradigm in which skills are first learned independently and then consolidated post hoc. To this end, we propose NestedVLA, which replaces parameter aggregation with parameter generation: a nested hypernetwork, conditioned on the target task, synthesizes task-adaptive action-expert parameters that compose with a frozen base VLA. Task-specific compressed VLM residuals are retained separately. This formulation enables flexible task-conditioned skill reuse while mitigating interference. Across three RL benchmarks (MetaWorld, ManiSkill, CALVIN), NestedVLA outperforms the strongest model-merging baseline by 41 pp on average, matches per-task fine-tuning within 1 pp on ManiSkill, and substantially closes the gap on CALVIN, pointing toward a storage-efficient route to skill consolidation. Code is available at https://github.com/WanyuGroup/NeurIPS2026-NestedVLA.
cs.LG / 116 / 2610.23372
Beyond Success Rate: Interventional Evaluation of Memory in Vision-Language-Action Systems
Abstract
When we say that a robot uses memory, does that memory actually play a role? Success-rate ablations alone cannot establish the use of task-relevant facts: they change information and computation together, while aggregate rates can hide offsetting gains and losses. We propose an Interventional Memory Evaluation Protocol (IMEP) for vision-language-action (VLA) systems to address four questions: (1) Does memory provide a net benefit? (2) Under what conditions does memory help? (3) Do changes to memory alter task outcomes? (4) Are these effects specific to task-relevant facts? At the retrieval interface, we compare target-fact substitutions with fact-preserving replacements, minimal numerical perturbations, and exact replicas, while assessing decision correctness separately from motor execution success. We evaluate within-episode, cross-episode, and agentic memory. In cross-episode learned-bank experiments, a one-unit-in-the-last-place (ULP) intervention, including its installation procedure, changes 11.9% of seed-paired outcomes, while changing key facts produces no clear additional effect over fact-preserving replacements. To examine when memory helps, we introduce ChainBench, where correct decisions require historical information unavailable in current observations. The full memory configuration raises task success from 39.1% to 74.3%, matching the oracle-decision condition under shared recorded executions. These results show that correct history can support robot decisions, but establishing whether a memory system realizes this benefit requires fact-specific evidence beyond aggregate success rates.
cs.LG / 117 / 2610.23526
Interpreting SSL Representations for Spoof Detection: a WavLM Study
Abstract
Self-supervised speech models such as WavLM achieve strong spoofing detection performance, yet which layers encode the most spoof-relevant information and what acoustic properties they capture remain opaque. We present a layer-wise interpretability study of WavLM using the ASVspoof 5 database, combining three complementary analyses: (i) per-layer linear probing with attack-level evaluation across speech synthesis, voice conversion, and adversarial attacks; (ii) a canonical correlation analysis (CCA) linking layer representations to spoofing-relevant acoustic features; (iii) an exhaustive grid search over layer subsets showing that use of five layers outperforms full thirteen-layer pooling. CCA analysis reveals that, while most acoustic features peak in alignment with mid-layers and then decay, voice quality measures jitter, shimmer, and the harmonics-to-noise ratio show the opposite trend, with correlation increasing toward later layers. Embedding visualisations reveal that deeper layers capture a continuous acoustic similarity among attacks rather than discrete categories, suggesting these layers encode fundamental synthesis artifacts rather than attack-specific signatures.
cs.LG / 118 / 2610.23466
Graphlet-based Decentralized Finance Transaction Method Classification
Abstract
Decentralized Finance (DeFi) transactions involve complex interactions among accounts, smart contracts, and tokens. We investigate whether network topology can identify transaction methods that describe DeFi operations. We extract directed graphlet features to classify the 30 most frequent methods. Our results show that graphlet features complement event log counts in improving classification performance. Using SHapley Additive exPlanations (SHAP), we further derive model explanations that reveal method-specific graphlet structural patterns.
cs.LG / 119 / 2610.23555
Beyond Latent Personalization: Integrating HRV-Conditioned Personalization via LLM for Interpretable Stress Detection
Abstract
Stress detection based on heart rate variability (HRV) has received increasing attention. However, existing personalized approaches integrate personal attributes with HRV through latent representations, limiting the interpretability of their interactions. In this work, we propose an LLM-enhanced personalized stress detection framework that integrates HRV features and personal attributes through LLM-based semantic interpretation. The proposed framework achieves a 29.9% relative MCC improvement over the strongest baseline on TILES-2018 while providing human-readable descriptions of individual physiological responses. The results highlight the potential of LLM-based semantic interpretation for achieving both effective and interpretable personalized stress detection.
cs.LG / 120 / 2610.23727
ECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health
Abstract
Conventional sleep staging provides a coarse-grained representation that obscures within-stage depth variation. We developed ECG-SDI, a single-lead electrocardiography (ECG) framework for continuous sleep-depth profiling and digital biomarker extraction. Using 12,392 overnight polysomnographic recordings, we trained and validated a two-stage deep learning model supervised by conventional sleep-stage labels. Correlations with ordered stages excluding rapid eye movement (REM) sleep were 0.828 in the pooled internal test set and 0.848 in the external Cleveland Family Study; corresponding REM classification AUROCs were 0.954 and 0.977. Overnight trajectories captured within-stage variation. Biomarker-based phenotyping identified a disturbed subtype associated with poorer subjective sleep quality and higher all-cause mortality in the Sleep Heart Health Study (adjusted hazard ratio, 1.20; 95% CI, 1.05-1.38). Among 507 participants with paired recordings before and after continuous positive airway pressure (CPAP), normalized sleep-depth area increased by 6.1% and shallow-sleep burden below 0.3 decreased by 16.0%. Exploratory analyses linked reductions in approximate entropy and depth variability to reduced oxygen desaturation and increased sleep efficiency, respectively (Spearman rho = 0.24 and 0.12); both associations remained positive after covariate adjustment. Mean ECG-SDI within N2 increased by 0.017 (95% CI, 0.004-0.030). A wearable proof-of-concept illustrated the acquisition-to-report workflow. These findings support ECG-based sleep phenotyping and further evaluation of CPAP response monitoring at home.
cs.LG / 121 / 2610.23292
Distributed Performance Boosting through Neural Control for Plug-and-Play Networks
Abstract
We address the problem of improving the performance of pre-stabilized networked systems subject to plug-and-play (PnP) operations, namely additions or removals of subsystems. While distributed baseline controllers can preserve stability after a PnP transition, augmenting them with expressive neural policies can invalidate the formal stability guarantees. To overcome this issue, we build on recently proposed Performance Boosting controllers based on the Internal Model Control (IMC) principle, which reduce closed-loop stability to a finite $\ell_2$-gain requirement on an auxiliary operator. In the distributed setting, this operator is formed by interconnecting local neural operators that exchange information only between neighboring subsystems. We then derive free parametrizations of this interconnection, namely explicit maps from unconstrained trainable variables to interconnected neural operators satisfying a prescribed network gain bound for every parameter value. The obtained distributed neural controller can therefore be trained by unconstrained gradient-based optimization without stability constraints or projection steps, while retaining the prescribed gain bound at every training iterate. We develop two complementary parametrizations. A neighbor-based compositional parametrization uses local coupling bounds, so that after a PnP event only subsystems whose bounds change need to update their gain parameters. A network-norm parametrization instead uses a bound on the overall coupling matrix and can reduce conservatism, at the price of requiring network-wide information. A PnP DC-microgrid case study shows that the proposed neural layer improves voltage transients under topology changes and load variations while preserving nominal closed-loop stability.
cs.LG / 122 / 2610.23760
From Flow-Box Invariants to Global Conservation Laws: Machine Learning-Assisted Detection of Singularities and Near-Integrability
Abstract
Conservation laws reveal invariant structure in dynamical systems and constrain their evolution to lower-dimensional level sets. Locally, however, per the flow-box theorem, every $d$-dimensional system generically admits $d-1$ functionally independent conservation laws near a regular point even though, generically, only a few such laws persist globally. Data-driven methods typically learn conservation laws on a prescribed region of phase space from which training samples are drawn, leaving their behavior beyond that region unresolved. We formulate this local-to-global continuation problem and develop an {expanding-domain neural deflation} framework that tracks the number of independent conservation laws over nested domains. The method recovers the expected $d-1$ local invariants and detects their successive breakdown as the domain expands, relaxing to the expected global conservation law limit. It distinguishes the local invariants (associated with Frobenius integrability) from persistent global structure and shows that the domain scale over which additional laws survive reflects proximity to Liouville integrability. Pointwise losses can localize isolated obstructions, whereas in more complicated dynamics the breakdown of our loss functions is gradual and spatially diffuse. The framework thus provides a data-driven approach for studying the persistence and breakdown of (so-far-learned) conservation laws across phase space.
cs.LG / 123 / 2610.23298
How the elastic axis folds the transonic flutter boundary: shock energy transfer and conformally calibrated active learning
Abstract
Isogai's case A, a pitch-plunge aerofoil whose elastic axis lies half a chord ahead of the leading edge, has a folded transonic flutter boundary in inviscid flow: near the bottom of the transonic dip several flutter speeds coexist at one Mach number. We map this boundary, and the family obtained by moving the elastic axis, with coupled unsteady Euler simulations, a Gaussian-process surrogate of the damping coefficient trained by batch active learning, and split conformal prediction on independent random simulations, which turns the surrogate into stability calls with a distribution-free error rate. Three failure modes of time-domain damping labels that leave no trace in the residuals are documented, and the half-degree mean incidence left by the common one-degree start is shown to raise the flutter speed at the dip by 2.8%, more than any numerical setting. With 80 simulations the learned case A boundary labels 64 of 65 independent runs correctly, and calibrated calls are wrong in at most 1.6% of cases. Over Mach number, elastic axis and flutter speed index, 193 simulations and 40 calibration runs show that the fold exists only for elastic axes more than about 0.4 chord ahead of the leading edge; it closes between 0.39 and 0.33 chord ahead, on two meshes. Chordwise energy budgets, closed against the solver loads and the growth rate within a few per cent, explain the boundary. The lower branch is crossed through the reduced frequency at unchanged shock phase. On the upper branch the shock is pinned at the trailing edge, and moving the elastic axis aft brings the pivot of the flutter mode towards the section, so that the damping of the supersonic surface falls faster than the residual shock input and the fold closes.
cs.LG / 124 / 2610.22691
Quantum-Classical Neural Networks for Solving Lindblad Master Equations
Abstract
Variational quantum algorithms have emerged as a promising approach for simulating open quantum systems, yet their expressive power and trainability remain poorly understood. We propose a variational framework for solving Lindblad master equations by minimizing a residual cost function using a compositional ansatz obtained by combining a classical neural network with a parameterized quantum circuit. The density matrix is represented through a purification on an enlarged Hilbert space, ensuring positivity, Hermiticity, and unit trace by construction, while all gradients are evaluated exactly via the parameter-shift rule. Beyond the algorithmic formulation, we derive an upper bound for the covering number of the proposed compositional model and compare it with that of purely quantum neural networks. Our analysis shows that the classical parametrization enlarges the effective hypothesis class while providing finer control over its complexity, thereby establishing a theoretical connection between expressivity, covering numbers, effective dimension, and the barren plateau phenomenon. Numerical experiments validate both the accuracy of the proposed method for representative Lindblad master equations and the theoretical predictions regarding gradient concentration. These results provide a unified variational and theoretical framework for the simulation of dissipative quantum dynamics using compositional quantum machine learning models.
cs.LG / 125 / 2610.23296
Learning angles for quantum approximate optimization
Abstract
The Quantum Approximate Optimization Algorithm (QAOA) and its variants are leading algorithms to help solve challenging combinatorial optimization problems. QAOA maps these problems to Ising Hamiltonians such that their ground state is the desired solution. Candidate solutions are then drawn from a variational quantum circuit. The performance of QAOA hinges on a set of variational angles. However, finding good angles is an instance-specific and computationally expensive non-convex classical optimization task. Here, we learn the mapping from the problem instance to high-quality QAOA angles. Our learned models replace the per-instance optimization, that we call the solver, with a single forward pass that is several orders of magnitude faster than the solver with a near identical quality. Our models further generalize to unseen topologies such as Barabási-Albert and 2D grid graphs. Experiments on quantum hardware confirm that QAOA samples with angles produced by our open-source models are indistinguishable from solver angles. Crucially, models sensitive to the topology of the problem instance increase the quality of the QAOA angles. Our work amortizes a major bottleneck in QAOA by an up-front data generation and training cost. Importantly, this cost can be amortized over time in settings where many similar problems are optimized on a regular basis, as is often the case in applications for industry.
cs.LG / 126 / 2610.23692
Expressivity of IQP Feature-Map Quantum Kernels via Reproducing Kernel Hilbert Spaces
Abstract
Quantum kernels are among the most promising methods in quantum machine learning, providing a natural way to embed data into high-dimensional Hilbert spaces via quantum circuits and capturing non-linear relationships in data. Their learning behaviour is governed by the function class they represent, which is formally characterized by the structure of the associated Reproducing Kernel Hilbert Space (RKHS). However, a theoretical characterization of this function class, and of how it depends on the choice of feature map and its hyperparameters, is still lacking. In this work, we focus on the Instantaneous Quantum Polynomial (IQP) feature map and derive a Fourier representation of the associated RKHS generators as finite sums of trigonometric functions evaluated on polynomially transformed input variables, with a finite set of accessible frequencies determined by the circuit construction and scaled by the bandwidth hyperparameter. This representation identifies three bandwidth regimes with qualitatively distinct expressive properties from localised, through multimodal, to periodic. Numerical experiments on synthetic datasets with controlled periodic and multimodal structure, across data dimensions and feature-map configurations, reproduce the three regimes in both supervised classification and unsupervised anomaly detection, benchmarked against a classical Gaussian kernel. These results make the relationship between expressivity and data structure explicit for IQP kernels: expressive power translates into a practical benefit only when the accessible frequencies match with the structure of the data, which provides a criterion for selecting the bandwidth when the periodic or multimodal structure of data is known.
cs.LG / 127 / 2610.23827
Quality-of-Transmission Estimation in Quantum Networks: Compute, Simulate, or Learn?
Abstract
Future quantum networks will rely on control-plane decisions such as routing, scheduling, entanglement provisioning, purification, and memory allocation. Each decision depends on a timely and reliable estimate of the quality of the entanglement that a candidate path can deliver, with end-to-end fidelity as a central metric. Inspired by the optical-networking experience with quality-of-transmission (QoT) estimation, this article asks when a quantum controller should compute fidelity analytically, simulate it, or learn a surrogate model. The optical analogy is useful but incomplete: optical QoT estimation often predicts a scalar transmission margin, whereas quantum QoT estimation must account for probabilistic entanglement generation, random memory waiting times, decoherence, state consumption, and measurement constraints that can make fidelity a stochastic target. We propose a compute/simulate/learn framework for quantum QoT estimation and illustrate it with a simulation-based proof of concept. Our study shows that analytical composition can be sufficient in simple regimes, that simulation becomes necessary when stochastic memory effects shape the delivered fidelity distribution, and that lightweight learned surrogates can approximate simulator labels at much lower runtime cost for online control. We close with open challenges for deployment, including stochastic targets, sim-to-real transfer, data-generation cost, physics-informed learning, and generalization beyond individual paths to full network topologies.
cs.LG / 128 / 2610.23834
Task-aligned spectral inductive biases in parameterized quantum circuits
Abstract
Parameterized quantum circuits are widely investigated in quantum machine learning, yet how their architecture can provide task-aligned inductive biases remains poorly understood. We study data reuploading quantum circuits with two distinct inputs: a signal that varies between samples and a fixed, task-defined spectral query that specifies how the signal should be probed. We show that encoding the query can change the circuit's hypothesis class over the signal by modulating how its existing signal frequencies are combined, without introducing new sample-dependent information. This mechanism matches tasks built on nonuniform Fourier transforms, whose evaluation frequencies specify the query and whose sampling positions guide its encoding. We examine this connection in a classification task inspired by synthetic aperture radar, where acquisition geometry determines a query for classifying unfocused spectral patches before image formation. Across small-scale spectral and radar-like tasks, aligned queries improve learning from limited data over parameter-matched circuits with mismatched or absent queries. These results provide a route from known spectral task structure to the design and validation of inductive biases in parameterized quantum circuits.
cs.LG / 129 / 2610.22930
Conformal Policy Selection: When to Deploy AI Decisions for Targeted Intervention
Abstract
Artificial-intelligence (AI) policies are increasingly used to make consequential decisions in medicine, finance, education, and public services, yet deploying them responsibly requires knowing not only whether a policy helps on average, but for whom. Existing tools for off-policy evaluation and safe policy learning provide population-level guarantees, but some individuals may experience adverse treatment effects. We recast targeted deployment as a multiple-testing problem: for each candidate unit, test whether the outcome under the AI policy would exceed a user-specified threshold, and control the false discovery rate (FDR) among the selected set. Our method, Conformal Policy Selection, combines weighted conformal $p$-values that account for the distributional shift from behavior to target policy with a two-step selection-and-pruning procedure. We prove finite-sample FDR control and establish a doubly robust asymptotic guarantee, where FDR remains controlled if either the outcome model or the propensity model is correctly specified. Synthetic experiments verify these guarantees across four misspecification regimes, and a semi-synthetic study built on real electronic-health-record covariates illustrates how the same FDR control can act as a deployment safety gate, bounding the harmful treatments that a globally trained policy would otherwise administer to a high-risk subpopulation where it misgeneralizes. The framework offers a principled method for safe AI deployment in high-stakes applications.
cs.LG / 130 / 2610.23543
Causal Discovery beyond Conditional Independence: Identifiability, Algorithms, and Statistical Inference for Directed Acyclic and Cyclic Graphs
Abstract
Learning causal structure from data requires assumptions about mechanisms, noises, observation schemes, and interventions. Historically, causal discovery has relied primarily on conditional independence. This review focuses on modern causal discovery methods that exploit properties of distributions beyond conditional independence. Our organizing principle separates the source of identification from the statistical estimator and its computational implementation. This perspective connects conditional-independence methods with distributional asymmetries, conditional-moment restrictions, and variation across environments or individuals. We examine discovery under latent confounding and feedback, distinguishing causal information about observed variables from reconstruction of latent structure and emphasizing the additional semantics required by cyclic models. We also distinguish population identification, frequentist and Bayesian consistency, uncertainty quantification, and algorithmic guarantees. Throughout, examples show why the same observational distribution can support different conclusions under different model classes. We conclude with guidance for comparing methods at their identifiable targets and a research agenda for uncertainty, heterogeneous populations, non-scalar variables, and the intersection of discrete data, feedback, and hidden confounding.
cs.LG / 131 / 2610.22516
Source-Aware Dynamic Pricing with Strategic Buyers
Abstract
In modern commerce, firms and platforms typically use customer features from different sources to personalize prices: some features are verified using seller records or third-party data, while others are self-reported and may be strategically manipulated by buyers seeking lower prices. We therefore introduce a novel source-aware framework for contextual dynamic pricing with strategic buyers in a high-dimensional semiparametric setting, where buyer valuations follow a sparse linear model with an unknown noise distribution. To account for differences in manipulability across feature sources while enabling effective valuation learning, we develop the Source-Aware Localized exploration and calibrated Exploitation (SALE) pricing policy, which draws on verified features to localize price exploration and combines verified and reported features to correct for strategic manipulation during exploitation. Our localized exploration design allows converting binary purchase feedback into unbiased signals of conditional mean valuations, yielding a new population equation that supports consistent sparse estimation despite non-i.i.d. buyer arrivals. Building on this estimate, we develop a localized empirical-process analysis that exploits intrinsic sparse geometry to establish uniform guarantees for nonparametric estimation of the noise distribution in the high-dimensional setting. With these estimation guarantees, we establish SALE's self-recovery property, which keeps localized exploration valid with high probability and leads to a sublinear expected regret bound. Both simulation studies and real-data experiments demonstrate the appealing performance of SALE. To the best of our knowledge, our study is the first to provide a theoretical and practical framework for incorporating information provenance into dynamic pricing with strategic buyers.
cs.LG / 132 / 2610.22655
Is a Linear Probe Evidence of a Linear Representation?
Abstract
The linear representation hypothesis holds that pretrained vision models encode generative factors as linearly decodable directions in feature space, and comes with a standard methodology: train a linear probe, declare a factor represented if accuracy is high. We show this test is too permissive. Asymptotically, any sufficiently rich feature map, including a random one, admits a linear readout, so high probe accuracy is consistent with two very different geometries: a true linear world model, where the latent-aligned subspace dominates the representation, or a kernel machine, where the latents lie linearly inside a much larger curved manifold. This distinction matters because the curved structure orthogonal to the probe is relevant for out-of-distribution generalization, robustness, and compositional transfer. We propose three asymmetric diagnostics that distinguish the two cases: reverse predictivity, spectral concentration, and out-of-distribution generalization. We validate them in a fully observed simulation where ground truth is known. Applied to six pretrained vision transformers on SVG-World, our deterministic generative benchmark with paired visual styles, and to two of them on CelebA, they show pretraining produces structure beyond a random kernel but well short of a linear world model. On dSprites, 3D Shapes and MPI3D, forward decodability and latent-aligned variance dissociate. Thus, forward decoding alone does not show that the probed factors align with the dominant geometry of the representation. In summary, our work cautions about the success of linear probes as opposed to truly linear and low dimensional representations.
cs.LG / 133 / 2610.22693
Optimal approximation of smooth functions by ReLU neural networks
Abstract
We study uniform approximation of the Hölder ball $\mathcal H_1^s([0,1]^D)$ by fully connected ReLU networks. Except in the one-dimensional Lipschitz case, existing constructions remain separated from the information-theoretic efficiency limit. This article closes the gap for every $s>0$ and $D\ge1$: networks with width $O(N)$, depth $O(L)$, and parameter magnitude $O(1)$ attain uniform approximation error of order $(N^2L^2\log N)^{-s/D}$. No relation between width and depth is required. The construction uses new encoding and decoding mechanisms that exploit higher-order finite-difference bounds for local Taylor coefficients of the target function. As a statistical consequence, empirical risk minimizers over the corresponding clipped network classes attain the minimax regression rate $n^{-2s/(2s+D)}$ without additional logarithmic factors. The paper also gives a detailed exposition of the key approximation techniques, covering both established and new mechanisms.
cs.LG / 134 / 2610.22770
Unit-level aggregation breaks coverage in causal foundation models
Abstract
Causal foundation models (CFMs) show promising results for treatment-effect estimation, with unit-level estimates that are competitive on standard benchmarks. Many causal analyses, however, inform decisions through effects averaged over groups of units, such as the full sample, the treated units, or a subgroup, and for such decisions a CFM is useful only if its credible intervals are calibrated. We examine these intervals and find that their coverage degrades, both in simple synthetic settings and on standard benchmarks. On ACIC 2016, for example, temperature-calibrated 95% intervals from CausalPFN fall below nominal coverage beyond groups of 30 out of 4802 units, and intervals for the full-sample average contain the true value in only 20% of datasets. We trace this problem to how these models are trained and how their intervals are formed. They output a predictive distribution for one unit at a time and, at inference time, combine the unit-level variances as if the errors on different units were independent. We show formally that posterior uncertainty over the data-generating process makes these errors correlated. As a result, the combined variance is too small by a factor that grows with the size of the group. Empirically, this factor predicts where coverage fails and by how much. We then show that two neural process models, which make joint predictions for a group of units, account for this correlation and substantially improve coverage on synthetic tasks and, as heads on a frozen CausalPFN, on all six benchmarks.
cs.LG / 135 / 2610.23033
Distinguishability and Recovery for a Doubly Sparse Spiked Wigner Model
Abstract
We study distinguishability and recovery of a sparse rank-one signal in a spiked Wigner model with independently sparse noise through a variety of algorithms and parameter regimes. The signal and noise have separate sparsity parameters, while the signal-to-noise ratio (SNR) remains of constant order as the matrix size grows. For sub-Gaussian noise and spikes whose nonzero coordinates are bounded away from zero, a leave-one-out quadratic thresholding procedure (a generalization of Ding et al., 2024) identifies the support. Sign synchronization yields exact recovery for a Rademacher spike, and spectral refinement yields strong recovery. The quadratic thresholding search parameter gives an explicit tradeoff between signal sparsity, noise sparsity, and computational cost. We also study extensions to sub-Gaussian spike priors through diagonal and quadratic thresholding. For Rademacher spikes with continuous nonzero noise, we develop an identical search algorithm that exploits unblurred signal entries, the existence of which permits distinguishability; connectivity of the graph they reveal provides a sufficient condition for exact recovery. These guarantees do not require moment assumptions on the continuous noise distribution. Together with our previous PCA results (Dumitriu et al., 2026), our analysis covers almost the entire space of parameters and regimes for distinguishing and recovering in our doubly sparse spiked Wigner model.
cs.LG / 136 / 2610.23131
Information-Geometric Blind Spots of Conformal Test Martingales
Abstract
Conformal test martingales (CTMs) are e-processes for sequentially testing exchangeability via p-values produced by a conformity function $A$. However, any distribution shift that preserves the push-forward $P\circ A^{-1}$ leaves the p-value distribution invariant, rendering it invisible to every CTM based on $A$. We formalise this blind spot through the \emph{$A$-equivalence class} $[P]_A$ and show that, for any target distribution $T$, there is a unique closest element $Q^\ast\in[P]_A$ in KL divergence---the \emph{$A$-cryptic I-projection}---with explicit density $q^\ast(z) = t(z)\,p_S(A(z))/t_S(A(z))$ and a Pythagorean decomposition into detectable and cryptic components. When $A$ is the Neyman-Pearson likelihood ratio, the cryptic component vanishes, connecting our result to optimal e-value design. We demonstrate the theory on high-dimensional image data by empirically constructing this I-projection. By conditionally resampling EMNIST letters to identically match the conformity scores of MNIST digits, we demonstrate that a distribution can be shifted to a disjoint support while leaving the test martingale mathematically blinded.
cs.LG / 137 / 2610.23322
The Loss or the Architecture? What Matters in Survival Analysis with Competing Risks
Abstract
Often, \emph{eg} in healthcare and predictive maintenance, what matters is which of several competing causes will strike first, and with what probability: survival analysis with competing risks. Corresponding deep learning methods are developed as bundles: a new architecture together with a new training loss, tested on one to five datasets against a handful of baselines. Thus, a state-of-the-art result cannot be attributed to the architecture, the loss, or the evaluation protocol. The coupling is not accidental: with one incidence curve per cause, a network's head is tied to the losses expressible on it. We vary the two axes independently, three losses, two of them strictly proper, on four backbones, across two modalities that tease the axes apart: 20 tabular and 3 pathology-image datasets. One loss wins in both: modelling cumulative incidence directly under censoring weighting. It brings twice as much as benefits a good choice of backbone on competing-risks data, and no more than it on single-event data: the loss is decisive for competing risks. The backbone matters too, and must match the modality: tabular networks lead on tables, pretrained vision encoders on images. The two choices compound rather than substitute. Proper scoring rules reveal these dynamics; the concordance index (C-index), which is not proper, does not separate models: neither axis moves it, not even tuning on the loss built to improve it. This is an avenue of progress: composed, the two give a model nobody has tried, RealMLP or TabM trained with the cumulative-incidence loss, which takes the best Brier score in our benchmark.
cs.LG / 138 / 2610.23477
Towards optimal selection schedules for online active learning
Abstract
We consider the problem of online linear classification, where a stream of unlabeled samples is available, and annotation is costly. We investigate the optimal query policy, that prescribes whether and when a sample should be labeled and used for a training step, or instead discarded, so as to minimize a given objective. We focus on the realizable case, under Gaussian design, in the high-dimensional regime where data dimension and number of unlabeled samples grow proportionally. We consider two objectives: 1) the generalization error under a labeling budget, and 2) a cumulative cost function where mistakes have unit cost and queries cost $Δ$. We first show that the myopic policies that greedily minimize at all times the increment in the objective correspond to either small or intermediate-margin policies, that prescribe the selection of samples whose margins fall within certain intervals that shrink with time. Within this class, we numerically ascertain the optimal schedules for the selection thresholds that minimize the final values of the objectives. Taking inspiration from the myopic policy for the cumulative cost, we design a very simple query policy schedule that we compare to other proposed heuristics, on synthetic data and real medical image datasets. Finally, we discuss myopic policies for two-layer neural networks, and exemplify how naive margin-based policies can cease to be optimal.
cs.LG / 139 / 2610.23508
Adaptive conformal prediction on the Grassmannian
Abstract
We introduce a conformal prediction framework for regression with Grassmann-valued responses. The proposed method constructs prediction regions in the form of metric balls directly on the Grassmann manifold, using basis-invariant geometric distances as non-conformity measures and scaling the corresponding non-conformity scores through a learned prediction difficulty estimator. This yields locally adaptive, input-dependent prediction regions that respect the underlying manifold geometry, while retaining the finite-sample marginal coverage guarantee of split conformal prediction under exchangeability. We evaluate several Grassmannian distance metrics and compare their prediction-region efficiency on synthetic benchmarks and an airfoil inverse shape-design problem. The results highlight the advantages of geometry-aware non-conformity scores, while local adaptation improves empirical coverage uniformity across prediction difficulty levels and, in strongly heteroscedastic settings, can substantially reduce prediction-region size while maintaining the target coverage.
cs.LG / 140 / 2610.23570
A general recipe for data-adaptive anytime-valid envelopes for iterative algorithms
Abstract
Most convergence guarantees for iterative stochastic algorithms are stated at a fixed iteration count and do not justify stopping at a data-dependent time. Recent work has constructed anytime-valid envelopes that do, but deterministic envelopes are typically too conservative for practical stopping, and existing data-adaptive ones are derived case-by-case, each coupling a specific algorithm to a specific noise model and boundary construction. This paper provides a general recipe for constructing such anytime-valid data-adaptive boundaries for a large class of iterative algorithms. Performance measures of many of these algorithms satisfy a one-step recursion that reduces the entire error analysis to a single general requirement: time-uniform concentration for predictably weighted sums of noise. We formalize this through a strongly adapted sub-$ψ$ condition, and derive off-the-shelf boundary constructions via continuous and discrete mixtures, stitching, and product betting martingales. We show how these boundaries can be directly coupled with a separate algorithmic analysis that derives the one-step recursion to provide a class of readymade anytime-valid envelopes across diverse settings including stochastic gradient descent (SGD), temporal difference learning, and Bures-Wasserstein SGD. In particular, for SGD, our framework either recovers or significantly sharpens every existing envelope we are aware of. Our constructions come with explicit implementation details and are substantiated by numerical experiments.
cs.LG / 141 / 2610.23779
Mesoscopic learning dynamics of extensive-width neural networks
Abstract
Understanding how stochastic gradient descent learns internal representations in wide neural networks trained on rich target functions is a central problem in machine learning. We study this question in two-layer neural networks learning target functions represented by networks of the same architecture, with arbitrary nonlinearity, and with network width and input dimension growing proportionally. In this regime, standard approaches for describing the learning dynamics require tracking a diverging number of overlaps, which measure correlations between learned and target features, leaving no obvious description in terms of a few order parameters. Here, we develop a mesoscopic description of learning by characterizing the distribution of overlaps. In the early stages of learning, we find that a Gaussian description of overlap statistics reduces the learning dynamics to a few collective variables that track learning of the target function's linear and quadratic components. We then estimate the timescale of specialization, when Gaussian closure breaks down and individual learned features begin to strongly correlate with specific target features. We find that specialization in extensive-rank networks is accelerated via a quadratic shielding mechanism, whereby learning the linear and quadratic components of the target function filters gradient fluctuations. We finally describe how the activation function determines the pattern of specialization and characterize the subsequent refinement of learned features. Together, these results show how collective learning gives way to individual feature discovery and clarify when learning in large neural networks admits a simple statistical description.
cs.LG / 142 / 2610.23788
Parametric Resonance in Primal-Dual Learning for Slowly Non-Stationary Constrained MDPs
Abstract
Primal-dual methods for constrained Markov decision processes are analyzed under a two-timescale assumption, with non-stationarity entering through two channels. In the additive channel, where the constraint level oscillates, the dual oscillation saturates at a floor set by the moving optimum and is amplified linearly by the timescale ratio. In the parametric channel, where the constraint sensitivity oscillates, a coherent modulation carries no resonant forcing at any depth: the dual update's stiffness modulation cancels the primal update's damping modulation. Instability instead comes from a mismatch between the constraint estimate used by the primal update and the measurement used by the dual update, producing a parametric resonance at twice the pair's natural frequency. The tolerated mismatch grows with the square root of the forcing bandwidth and collapses with the square root of the dual-to-primal step-size ratio. These results hold on the tangent linear-quadratic reduction around an active saddle, for one constraint and for $K$ constraints under a common modulation and an isotropy condition. Driving both updates from the same estimate removes the resonant forcing but introduces a stationary offset equal to the estimator's bias. In a Lagrangian soft actor-critic agent, the multiplier locks onto the sub-harmonic at the timescale ratios its identified linear loop predicts and a coherent cost critic suppresses the lock-in, while a proportional (PID) term, predicted to leave the lock-in in place, suppresses most of it through the projection of the multiplier.
cs.LG / 143 / 2610.23821
Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere
Abstract
Many of the most important datasets in the physical sciences live natively on the sphere, and these data often carry no preferred position or orientation. Using methods that do not reflect this symmetry leads to far higher simulation requirements, which can be catastrophic for scientific analyses (forcing reliance on cheap, imperfect models that bias results). We present a novel spherical implementation of a gauge equivariant mesh architecture that is specifically designed for the Hierarchical Equal Area isoLatitude Pixelization (\healpix{}) of the sphere. The $\mathrm{SO}(3)$-equivariant convolution is built from steerable filters, parallel transport and spin-weighted features. The architecture supplies the exact symmetry, so the network does not spend training data learning it. The convolution operations are constructed to give provably exact equivariance, given a band-limit. Our full CNNs can achieve SO(3)-invariance error at per-cent-level and lower, which is controlled through well-understood hyperparameters. On benchmarks, our technique gives the most accurate dense gradient regression and competitive cosmic microwave background foreground removal and denoising. The improved simulation efficiency matters most in simulation-based inference (SBI), where high-fidelity simulations are expensive and neural compressors require large training sets, so a fixed budget forces analyses with less accurate simulators. In cosmology lensing SBI, our technique yields the tightest posteriors at matched simulation count. This approach is a novel method that will enable new convolutional neural network analyses, including generative and inference tasks, for processes and data on the sphere.
cs.LG / 144 / 2610.23825
ENFR: Evidence Networks with Flow-Matched References for Amortized Simulation-Based Model Inference
Abstract
The evidence, $\mathcal{Z} = p(x | \mathcal{M})$, of a data-generating model $\mathcal{M}$ is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics, e.g. $χ^2$. While versatile, the evidence is notoriously difficult to compute, even when the likelihood is available. In both likelihood-based and simulation-based inference, estimating $\mathcal{Z}$ provides an immediate solution to high-dimensional goodness-of-fit or model misspecification tests: using $\mathcal{Z}(x)$ as the test statistic induces a minimum-volume level-$α$ test. Such tests are essential to assess model fidelity, which is necessary to trust simulation-based parameter inference. Existing learned approaches largely target Bayes factors, which only measure evidence ratios. Instead, we provide accurate, fast, and amortized estimates of the absolute evidence. We present ENFR: Evidence Networks with Flow-matched References. Building on the Evidence Network framework of Jeffrey & Wandelt (2024), the proposed method first learns Bayes factors with respect to a flow-matched reference, from which the absolute evidence is recovered. We show improvement in estimation of $\mathcal{Z}$ over the flow-based baseline in all applications considered and the best performance of all density estimation methods considered in our benchmarks. We furthermore introduce and provide the proof for an architecture that jointly learns the individual evidences of a set of models $\{\mathcal{M}_i\}$. We demonstrate the performance and scaling behavior of these approaches on data from a variety of scientific applications.
神经与进化计算 (cs.NE)
3
cs.NE / 1 / 2610.22484
Upper bounds on the number of causal pieces in spiking neural networks
Abstract
The input space of single-spike neural networks decomposes into causal pieces, and the number of causal pieces has been shown to provide a lower bound on the approximation error for some spiking neural networks. While this decomposition is applicable to a wide range of spiking neuron models, the limits of how many causal pieces can be realised by spiking neural networks are still unknown. In this work, we derive two complementary upper bounds on the number of causal pieces in feedforward single-spike neural networks. First, we establish a combinatorial bound given by the poly-Bernoulli numbers, which applies to a broad class of spiking neuron models. The bound is of order $e^{2(L-1)N\log N+\mathcal{O}(N)}$, with network depth $L$ and maximum layer width $N$. Second, for a specific set of neuron models, we derive a bound based on hyperplane arrangements. Applying the latter framework, we obtain bounds of order $\mathcal{O}\!\left(N^{(4L-3)(N_0-1)}\right)$ for the non-leaky integrate-and-fire model with exponential synapses and the ReLU-type spike response model, and $\mathcal{O}\!\left(N^{L(N_0-1)}\right)$ for the leaky integrate-and-fire model with Dirac-$δ$ synapses, where $N_0$ is the input dimension. For the non-leaky integrate-and-fire model with Dirac-$δ$ synapses, additional temporal structure allows us to sharpen the general combinatorial bound to one that depends only on $N_0$, independently of the network depth and widths.
cs.NE / 2 / 2610.23046
SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks
Abstract
On neuromorphic chips, spiking neural networks (SNNs) consume far less energy than artificial neural networks (ANNs), but the limited memory of these chips requires low-bit synaptic weights and membrane potentials. Existing post-training quantization (PTQ) methods provide them without retraining and, like those for ANNs, minimize numerical errors such as the mean squared error (MSE). In an SNN, however, a quantization error reaches later layers only through the spikes it flips. We find that errors of equal MSE can cause very different accuracy losses, whereas errors that flip the same fraction of spikes cause the same loss. We therefore propose SpikeQ, a spike-guided PTQ framework. For weights, we select the level of each weight by simulating the downstream neurons and minimizing their spike errors. For membrane potentials, most existing methods quantize the state after the reset, whereas neuromorphic chips compare the threshold with the membrane potential stored at a fixed bit width. We target this pre-fire quantization and build a reset-aware grid that places the threshold on a rounding boundary, so rounding never flips the spike at the current time step. We evaluate SpikeQ on seven benchmarks of image classification, semantic segmentation, and object detection. With 2-bit pre-fire membrane potentials, every existing grid collapses, whereas SpikeQ retains most of the full-precision accuracy. With both at 4 bits, the models stay within 2.59 points of full precision, shrink their memory footprint 7.8--7.9 times, and use 6.4--6.9 times less energy.
cs.NE / 3 / 2610.23070
Wavelength Diversity Expands the Expressive Power of Optical Computing
Abstract
Universal function approximation underpins neural networks' capacity to represent nonlinear relationships, but realizing this capability in optical hardware requires engineering the computational features encoded in light. Here, we show that wavelength diversity enhances optical function-approximation expressivity through wavelength-scaled Fourier features generated by a shared phase encoder, without increasing the number of input-encoded harmonics. Multiple wavelengths contribute to the representation of each output function rather than merely serving as independent computational channels. We establish theoretical scaling laws and approximation-error bounds, revealing that, for nondispersive encoders, the spectral endpoints determine the accessible computational bandwidth, while intermediate illumination wavelengths enrich the available harmonics within that bandwidth. Numerical studies demonstrate improved approximation of real- and complex-valued nonlinear functions, with a representative design achieving more than a 100-fold reduction in mean-squared error relative to tested monochromatic optical processors with the same number of trainable phase features. Experimentally, we validate enhanced approximation accuracy and demonstrate the simultaneous approximation of 4,136 distinct nonlinear functions using a compact, single-layer diffractive processor under five-wavelength illumination. These results establish wavelength diversity as an additional physical resource for scaling the expressive power of optical computing and enhancing the accuracy of parallel nonlinear function approximation.
计算语言学 (cs.CL)
74
cs.CL / 1 / 2610.22507
RealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents
Abstract
Command-line interface (CLI) agents increasingly execute longer-horizon tasks, yet existing benchmarks either assume that user instructions remain fixed throughout execution or evaluate task outcomes without capturing runtime collaboration quality. In this paper, we present REALCLI, a paired benchmark that unifies the evaluation of task execution and human-agent collaboration quality on realistic CLI tasks. To comprehensively measure collaboration, REALCLI builds a five-dimensional framework, operationalized into 22 sub-dimensions and 51 interaction scenario templates. REALCLI provides two modes for paired comparison: REALCLI-EXEC supplies complete specifications upfront, whereas REALCLI-COLLAB introduces a scripted user that delivers progressive information, corrections, constraint revisions, and authorization decisions. Evaluations reveal profound capability gaps in frontier agents: the best configuration attains an average score of only 49.2%, and when evaluated by the joint collaboration score ($J_{\mathrm{collab}}$), the peak performance reaches merely 37.6%. These results show that task effectiveness diverges sharply from collaboration quality, with models consistently struggling in steerability, trust calibration, and safety boundary management. REALCLI provides a rigorous, reproducible foundation for measuring and advancing collaborative CLI agents.
cs.CL / 2 / 2610.22515
alkhder at ImageEval 2026 Shared Tasks: Constrained Selection and Rotation-Averaged Logit Scoring for Multimodal Hallucination Detection
Abstract
We describe our submission to Subtask~1b of ImageEval 2026. The task consists of selecting the single visually valid statement from three candidates paired with an image, and is in both English and Modern Standard Arabic (MSA). Our system uses a pre-trained 7B vision--language model in a zero-shot setting, without task-specific training or fine-tuning. We formulate the task as a constrained selection problem and choose the answer using next-token log-probabilities. To reduce answer-position bias, we apply a Latin-square rotation to the candidate statements. Our system achieves CI scores of 0.0510 on the English test set and 0.0960 on the MSA test set, reducing CI relative to the official 3B baselines reported on devtest by 83.7\% and 80.4\%, respectively.
cs.CL / 3 / 2610.22518
Learning User Representations to Improve Simulation Fidelity
Abstract
User simulators enable interactive training and evaluation of AI assistants at scale. Yet current simulators struggle to reproduce the diverse behaviors of real users, leading to assistants that generalize poorly to underrepresented users. We identify user representations as an underexplored bottleneck in user simulation. User representations are natural-language descriptions that condition simulators to reproduce a particular user's behavior. A representation's quality becomes apparent in hindsight, through the behavior it induces in a simulator. To evaluate and optimize user representations, we formulate a conversation reconstruction task: given an original conversation, we generate a representation, condition a simulator on it, simulate a conversation, and then measure fidelity by comparing user behavior in the original and simulated conversations. This comparison also yields hindsight feedback, which identifies where the representation falls short and how it can be improved. We introduce hindsight distillation, an online training procedure that distills feedback-conditioned improvements into the representation generator, enabling it to generate improved representations directly from the original conversation. We present UserRepGen-4B, a representation generator trained with hindsight distillation. Across 700 conversations spanning seven datasets, hindsight feedback helps improve user representations at test time, and UserRepGen-4B generates representations that enable more faithful simulation than those generated by GPT-5.5. Beyond conversation reconstruction, learned representations improve simulation fidelity for unseen tasks, better preserve the population-level diversity of real users, and better distinguish users across conversations.
cs.CL / 4 / 2610.22525
alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B
Abstract
This paper describes team Alkhder's submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker-addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64 on both tracks.
cs.CL / 5 / 2610.22526
Literal Lock-In: How Test Language Shifts Literal Attraction in Typed Decision Models on Bengali Idioms
Abstract
Decision models, non-generative scorers that report a probability over caller-supplied options and emit no text, are increasingly deployed as routers, classifiers and guardrails. Because they generate nothing, rationales and chain-of-thought do not apply. We audit three: an encoder scorer, a decoder scorer and a closed API, entirely through interventions on their input, using expert-annotated Bengali idioms carrying a gold figurative and a gold literal gloss. Dropping an idiom's own literal gloss into the option list draws every arm toward it, taking 65% to 96% of each arm's remaining Bengali errors. A model-by-language matrix over one shared item set shows how far this depends on the test language: every arm is 1.9 to 6.4 times more captured in Bengali than in English on the same paired entries, the commercial model's apparent resistance is a property of its test language, and capability does not order the failure. Calibration degrades in all three arms, and on the open arms confidence inverts, so gating on it is worse than gating at random. On a semantics-free option list, where accuracy is at chance, the best-calibrated arm still reports mean confidence 0.82. Code and data: https://github.com/sakhadib/literal-lock-in.
cs.CL / 6 / 2610.22529
Pearl-Chat: Engineering a native Luganda language model from first principles in JAX and Flax NNX
Abstract
Autoregressive foundation language models predominantly target high-resource Indo-European languages, imposing computational and morphological penalties on under-represented African languages. Generic multilingual tokenizers routinely exceed fertility rates of 3.2 tokens per word on agglutinative Bantu languages, degrading context efficiency and inflating latency. We present Pearl-Chat, an 8,473,122-parameter ($\sim$8.5M) causal decoder-only transformer engineered for Luganda (Oluganda, ISO 639-2: lg), a low-resource Niger-Congo Bantu language spoken by over 10 million people in East Africa. Built from first mathematical principles in JAX and Flax NNX, Pearl-Chat uses explicit functional state decomposition (nnx.split and nnx.merge), deterministic resumable data streaming via Google Grain, and SPMD tensor parallelism using jax.sharding.Mesh. We introduce a byte-level byte-pair encoding (BPE) subword tokenizer trained from scratch on curated Luganda corpora ($V = 8,192$), achieving 4.88 characters per token compression and 1.60 tokens per word fertility, a two-fold compression improvement over LLaMA-3 (128K) and Gemma (256K). The model is trained with a two-stage curriculum: Stage 1 pre-training on 59,501 sequences across parallel, biblical, encyclopedic, and journalistic corpora, and Stage 2 instruction fine-tuning on 510 dialogues with assistant-only loss masking. Across 4,000 optimization steps, validation cross-entropy loss drops monotonically from 9.4695 (perplexity: 12,958.98) to 0.1818 (perplexity: 1.20), a decrease of 9.2877 nats. On a three-tier held-out diagnostic benchmark, Pearl-Chat achieves 0.9792 ROUGE-L and 0.9405 BLEU while maintaining concordial agreement across Bantu noun classes. We provide a Python Server-Sent Events (SSE) streaming runtime and open-source all weights, tokenizers, and code.
cs.CL / 7 / 2610.22542
HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments
Abstract
While LLMs demonstrate strong capabilities in English Relation Extraction (RE) tasks, a significant gap remains in assessing their RE performance in low-resource languages, due to a lack of standardized benchmarks. Creating native, language-specific annotated RE benchmarks is expensive and requires human expertise, while relying on translation yields suboptimal, inaccurate datasets that introduce English-centric biases. To tackle this, we propose a scalable, language-agnostic methodology to create RE data by denoising naturally occurring yet noisy alignments of knowledge bases and corresponding texts. We define a proper alignment via a Natural Language Inference (NLI) task, where a relation triplet is entailed by an input text. We empirically assess five denoising strategies, and demonstrate that multilingual models outperform language-specific models in the few-shot LLM setting, while Hebrew-specific encoders remain competitive under NLI fine-tuning and are strongest in the label-free cross-training setting. We further find that prompting in the native language achieves performance on par with or slightly exceeding English prompts, confirming that the pipeline operates naturally in the target language. The best open-weights model, Gemma-4-31B-it (F1\,=\,0.915), matches the best API model, GPT-5.4 (0.921), and reaches F1\,=\,0.844 on a fully held-out, dual-annotated test set. Applying the optimized pipeline, we release HeRE: a novel, massive, 1.5-million-instance silver dataset for Hebrew RE, designed to support data-intensive experimentation and model fine-tuning. HeRE bridges a critical gap in standard Hebrew RE evaluation, while providing the NLP community with a reproducible, language-agnostic pipeline for RE dataset construction in other low-resource settings.
cs.CL / 8 / 2610.22567
Who Do Perspectivist Models Predict Well? Annotator Agreement, Sparsity, and Performance
Abstract
Successful completion of perspectivist natural language processing (NLP) annotation tasks requires algorithms to be aligned across diverse perspectives in order to predict how different humans will rate text. This work studies how the performance of such algorithms varies for annotators who behave differently than others, either by providing annotations that diverge from others' responses or by providing fewer annotations altogether. To aid with characterizing model performance, we curate the largest collection to date of perspectivist NLP datasets with ordinal annotations. After evaluating multiple classes of perspectivist algorithms across each of the datasets, we model the relationship between individual annotator idiosyncrasy and algorithm performance. In general, we find that algorithms perform better on annotators who tend to agree with others, but that annotator response rate does not have a statistically significant positive correlation with algorithm performance.
cs.CL / 9 / 2610.22586
Mitigating Demographic Disparities in Explanation-Human Alignment via Rationale Regularization
Abstract
We study explanation fairness in hate speech detection. High classification performance does not always mean that a model produces equally plausible explanations for different demographic groups. Using the HateXplain dataset, we introduce Plausibility-EFG, a metric that measures the largest difference in explanation plausibility across demographic groups by comparing model-generated token attributions with human-annotated rationales. We evaluate BERT, RoBERTa, and ADV-RoBERTa, an adversarially trained model, using Integrated Gradients and SHAP. Our results show that ADV-RoBERTa improves classification performance and alignment between explanations and human rationales while reducing explanation disparities across demographic groups. These improvements are observed with both attribution methods and remain consistent across training, validation, and test splits. We also show that reducing prediction-level bias alone does not eliminate disparities in explanations, suggesting that predictive fairness and explanation fairness are related but distinct objectives. Finally, our fairness-performance analysis shows that stronger adversarial regularization can further reduce Plausibility-EFG, although it may cause a small decrease in macro-F1. Overall, this work presents explanation fairness as an important objective in hate speech moderation and introduces Fair-REF, a framework designed to reduce disparities in model explanations across demographic groups.
cs.CL / 10 / 2610.22588
SAG2: Embedding-Free Multi-Hop Retrieval with BM25 and JEV
Abstract
Strong multi-hop retrieval should be practical to deploy without expensive corpus preparation or complex database infrastructure. We present SAG2, an embedding-free retrieval architecture built around a single SQLite database, BM25, and contextual evidence decisions from JEV. SAG2 follows the query-time association principle of SAG (SQL-Retrieval Augmented Generation), while replacing its event-entity and dense indexes with a lexical index over original passages. Selected passages supply terms for further retrieval, and newly retrieved evidence guides subsequent decisions. This moves semantic evidence selection to query time, eliminating offline embedding computation, LLM-based event-entity extraction, and dedicated database servers. The resulting system combines lightweight preparation with strong retrieval accuracy: on 1,000-question subsets of MuSiQue, 2WikiMultiHopQA, and HotpotQA, SAG2 achieves the highest Recall@5 and Recall@10 among the compared methods, with Recall@10 of 86.62%, 97.43%, and 99.00%. It also reaches 81.10% document Recall@10 over the 511,962-document EnterpriseRAG-Bench corpus. Indexing all 11,656 MuSiQue passages takes 2.36 seconds, produces a 25.87 MiB database, and requires no model calls. SAG2 demonstrates that accurate multi-hop evidence retrieval can be supported by compact, rapidly prepared infrastructure, making it a practical foundation for knowledge retrieval applications.
cs.CL / 11 / 2610.22594
The Model Knows Better: Defending Medical Multi-Agent Systems Against Prompt Injection
Abstract
Medical multi-agent systems, in which LLM agents discuss a case, are increasingly studied for clinical decision support. Prompt injection, text that an attacker adds to a model's input, can manipulate LLMs into wrong decisions. How it affects these systems, and how to protect them, remains unclear. Across two systems, two model families and three benchmarks, a single injected passage succeeds in up to 61% of cases, and collaboration does not consistently protect against it. Existing defenses help only partly: they miss injections that look like ordinary clinical text. Yet when asked to check the injected text against the patient's findings, a model that followed the injection usually rejects it. We therefore propose a training-free defense that verifies the clinical judgments behind the system's decision. Following clinical practice, it separates the patient's findings from clinical judgments, such as diagnoses, recommendations and instructions, and checks each judgment against the findings alone, so the decision rests only on supported judgments. Our defense outperforms all baselines on every benchmark, lowering attack success on MedQA to 4-7% while maintaining accuracy on cases without an attack, and remains effective with a medical model and against defense-aware and adaptive attacks.
cs.CL / 12 / 2610.22597
Broad and Efficient PII Detection with Unified Labels
Abstract
Personally Identifiable Information (PII) detection models must efficiently find identifiers across diverse contexts while maintaining a low false-positive rate (FPR) on high-volume traffic. Recent detectors increasingly rely on complex architectures that raise deployment cost, yet many are fine-tuned on only a few synthetic datasets that limit their coverage. Combining public datasets is challenging because their taxonomies differ, and even after categories are mapped to a common schema, labels can disagree on what counts as a detection, producing inconsistent training and evaluation signals. To address these problems, we present a practical 141M-parameter encoder-based detector trained on 1.12 million examples from 77 public datasets, relabeled under a detailed definition of broad contextual detection for 12 structured PII categories. Evaluated on the AI4Privacy and Nemotron-PII benchmarks at a low-FPR operating point, our model achieves the highest recall against larger baselines in 11 of 14 comparisons under our labels and 9 of 14 under the original labels. Its largest recall gains include useful additional coverage, such as multilingual disclosures and IP addresses embedded in URLs. It achieves this with 2-10x higher throughput and 94-97% less peak GPU memory than the baselines, delivering leading detection performance at substantially lower cost.
cs.CL / 13 / 2610.22615
When Does Uncertainty Measure RAG Hallucination Risk?
Abstract
Uncertainty can measure hallucination risk in retrieval-augmented generation (RAG), but only when retrieved evidence constrains the answer. When evidence is present, unsupported factual claims become less likely; when retrieval coverage is poor, the model can answer from internal knowledge and assign high probability to unsupported claims. We test this across four instruction-tuned models on HotpotQA and SQuAD. After matching faithful and hallucinated answers within word-count bins, length-only classifiers fall near chance, while perplexity reaches 0.81--0.91 AUROC on targeted unsupported edits. Varying retrieval depth shows why the score changes meaning: detection can weaken with only the top-ranked passage, then recover after adding a second passage. A combined uncertainty and semantic-drift score improves calibration, and measurements remain above chance on RAGTruth QA. Results support a conditional view of uncertainty as a useful measure of RAG hallucination risk.
cs.CL / 14 / 2610.22632
ToMStories: Interpreting Theory-of-Mind in Synthetic Narratives
Abstract
Understanding people often requires inferring what they feel or want from how they behave rather than from what they state outright. We ask whether language models can make these inferences, and how they do so. We introduce Theory-of-Mind Stories (ToMStories), an evaluation dataset of ~6,000 short synthetically generated narrative vignettes, each paired with a multiple-choice or open-ended question about a character's emotion, fear, motivation, or emotional wound. These psychological states are conveyed implicitly (showing), explicitly (telling), or with conflict between dialogue and description. Evaluating thirteen open-weight models (1B-72B parameters), we find that the largest models match humans on multiple choice but not open-ended questions. Our analyses point to simplifications models make in both interpretation and generation of implicit psychological states, indicating that synthetic data can help us understand how models perform interpretation, but likely overestimates how well models interpret humans.
cs.CL / 15 / 2610.22638
Inductive Learning of Concept Embeddings from Cross-Linguistic Colexification Patterns with Textual Attributes
Abstract
The advent of distributional semantics as a framework for computational semantics marked a paradigm shift in the nature of representations, by which meanings are represented as dense vectors commonly known as word embeddings. Computational approaches to historical or typological language comparison, on the other hand, still mostly rely on discrete representations, while the potential of embedding-based methods remains largely untapped in this context. Recent studies have proposed concept embeddings as a language-independent, onomasiological counterpart to word embeddings, based on networks of cross-linguistic colexification patterns. While the resulting concept embeddings reflect meaningful semantic properties and relations, they suffer from a lack of scalability, since they rely on transductive embedding approaches which only learn embeddings for nodes present in the training data. In this study, we present an inductive framework for learning concept embeddings, combining the structural properties of full and partial colexification networks with semantic information from short concept descriptions. This framework allows the embedding of unseen concepts based on a short definition. We show that the resulting concept embeddings correlate with various cross-linguistic measures of similarity or relatedness, while generalizing well to new concepts that were absent from the training data.
cs.CL / 16 / 2610.22640
SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors
Abstract
Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing approaches target affirmative response prefixes through rank-one updates to a small subset of model weights. Yet when this prefix-level target is satisfied, the model may still revert to refusal or produce a coherent, apparently compliant response that does not address the request. To overcome this limitation, existing internal-representation approaches shift focus from output tokens to dense, high-dimensional compliance directions in latent space. Existing representation-level approaches rely on preservation constraints to maintain conditionality, motivating us to ask whether a smaller internal target can offer a different effectiveness-preservation trade-off. Motivated by the comparable control capability of single-coordinate interventions, we investigate a sparse refusal-related MLP coordinate as this target. We propose SAFETYEDIT, which optimizes a single refusal-related coordinate and uses KV-cache marking in early attention layers to propagate a prefix-trigger signal across subsequent instruction positions. On some checkpoints a downstream MLP reader converts this signal into control of the coordinate, while on others the propagated signal alone suffices; the method is defined by the sparse target rather than by a fixed writer-reader pair. Across four open-weight LLMs and three dataset groups, SAFETYEDIT achieves the highest triggered Guard ASR in 11 of 12 comparisons and the highest mean StrongREJECT score in all 12, relative to the baselines evaluated for each model. Crucially, comprehensive ablation and behavioral audits confirm generally minimal changes in untriggered behavior and general capabilities in the tested configurations.
cs.CL / 17 / 2610.22652
Evidence-Gated Bilingual RAG for Reliable Consumer Banking Search
Abstract
Consumers often describe what they need rather than use the terminology found on a banking website. For example, a natural-language request ("a card for travel") may not share the keywords of the relevant product page, and the problem is harder in bilingual services, where the same intent may be expressed in Arabic or English. This paper presents an evidence-gated bilingual retrieval-augmented generation (RAG) framework for consumer banking search that operates entirely on local infrastructure. The framework combines structured web-content refinement, hybrid dense-lexical retrieval, intent-aware evidence selection, cross-encoder reranking, and a three-outcome Answer / Related Mismatch / No Answer decision mechanism. On 210 URL-labeled Arabic and English queries, the proposed approach improves Top-1 hit rate from 50.48% to 79.05% and MRR@5 from 0.571 to 0.797 over a hybrid baseline. A separate 26-case end-to-end evaluation achieved 69.2% strict success, with 72.7% of generated answers classified as supported by an automated faithfulness judge. These results are preliminary: the retrieval queries are paraphrases of a small reference set, and the end-to-end test is small and uses an automated rather than human judge, so they do not establish human-verified answer correctness or generalizable performance.
cs.CL / 18 / 2610.22657
Secondhand Stress: How Stressful Disclosures Affect Psychosocial Wellbeing in Online Graduate Admissions Communities
Abstract
Online communities provide spaces for individuals navigating stressful life transitions to disclose difficulties and seek peer support. Yet encountering others' distress may also have psychosocial implications for community members. Drawing on stress crossover theory, we examine how engagement with peer distress during graduate admissions is associated with subsequent psychosocial expression and participation, and whether similar associations extend through downstream interactions. We analyze 171,981 posts and 1,130,413 comments from two Reddit communities across three admissions cycles, using propensity-score stratification and baseline-adjusted comparisons. Users who engaged with distress disclosures during application preparation subsequently exhibited higher classifier-estimated anxiety, depression, and stress expression, alongside greater commenting activity and lexical redundancy. These associations recurred across admissions cycles, while no significant psychosocial associations were detected among downstream respondents interacting with previously exposed users. Our findings highlight how greater distress-related expression can coexist with continued community participation, while distinguishing direct engagement with distress disclosures from subsequent social contact. We discuss implications for designing online communities that preserve opportunities for self-disclosure and peer support while remaining attentive to the experiences of those encountering distress-related content.
cs.CL / 19 / 2610.22671
ProPerAgent: Toward Memory-Driven Proactive Assistance
Abstract
As assistants move from reactive answering toward proactive intervention, memory changes role: it no longer only supplies the answer to a query, it determines whether and when to speak at all. Prior work in this setting hands the system the relevant fact in context; we study the regime that makes the problem hard, where the evidence justifying an intervention must be mined from a long-horizon personal memory. We introduce ProPerMem, a benchmark for memory-driven proactive assistance (MPA), pairing egocentric video with year-long, distractor-rich user memories (~618K tokens on average), human-verified trigger windows, and a protocol that scores content and timing jointly. Every baseline we evaluate performs poorly (at most 15%), including frontier models given summarized or retrieval-augmented memory. A controlled study explains why: with oracle evidence delivered inside the trigger window GPT-5.3-Chat reaches 96%; letting it choose the timing drops it to 38%, and replacing oracle evidence with retrieval drops it to 15%. Content and timing are separately necessary, and missing either is catastrophic. Guided by this, we build ProPerAgent (Proactive Personalized Agent) with a plug-and-play harness that surfaces memory from a graph-organized store and gates it down to what is worth interrupting for, plus a lightweight RL stage that curbs early and repeated firing. Without any training, ProPerAgent lifts strict accuracy from below 10% to 37.5% on Qwen3.8-27B and 38% on GPT-5.3-Chat; RL post-training reaches 42.4%.
cs.CL / 20 / 2610.22686
Multimodal Filesystem Memory For Long-Horizon Embodied Agents
Abstract
Agents operating over long horizons need an effective memory that allows them to reuse their experience throughout long interactions. If the entire history of an interaction is retained, it becomes difficult to find the relevant experience, while compressing it into summaries may result in the loss of visual and spatial details that are required later on. We present \textbf{MFM}, a multimodal filesystem memory architecture that keeps complementary types of agent memory. \textbf{Episodic memory} stores the observations, actions, and outcomes, and \textbf{spatial memory} associates experiences with locations. \textbf{Semantic and procedural memory} build up general knowledge and reusable methods of acting, and \textbf{working memory} holds the current plans. MFM makes use of searchable textual records to connect these different representations to the underlying multimodal episodes. The agents are able to find previous experiences by means of contextual cues and to examine the original observations in order to recover details that are missing from their textual descriptions. Through consolidation, the agents convert experience into higher-level knowledge and procedures while still keeping references to the original episodes. In the case of long-horizon Minecraft and computer-use tasks, MFM leads to improved performance with three base models. With GPT-6-Astra, MFM completes \emph{Mine Diamond from Scratch} and reaches 16 out of 24 milestones on the \emph{Ender Dragon} task, as against 12.7 for the best baseline. On CUABench-KiCad it attains a mean reward of 0.440, as compared to 0.400 for the best baseline and 0.267 in the case of no persistent memory.
cs.CL / 21 / 2610.22714
First Passage Attention
Abstract
Augmenting softmax attention with forget gates typically relies on summing decay contributions along the key--query path. Because such accumulation is monotonic, these mechanisms cannot selectively erase intermediate context while preserving older, structurally relevant tokens. We propose First Passage Attention (FPA), which shifts from path summation to path extrema by lowering a key's logit proportionally to the deepest dip of a learned scalar field below that key's level over the intervening span. FPA encompasses the Forgetting Transformer (FoX) as a special case, and when the field and levels follow a shared learned walk, the alive keys with unpenalised logits naturally form a stack. We show that a constant-dimension FPA head realises any subset of keys left alive by subsequent dips, whereas softmax heads with positional biases or token-level forget gates require a head dimension linear in the number of keys. In synthetic experiments, one-layer FPA heads converge to these target sets while comparable softmax and FoX baselines fail on every seed. Scaling to a 110M-parameter language model, FPA matches FoX in perplexity and uniquely achieves perfect passkey retrieval at $8\times$ the training length, even with its key--value cache compressed to about $25\%$.
cs.CL / 22 / 2610.22717
ARBOR: Locating Reasoning Failures in Long Chains of Thought through Reasoning State Tracking
Abstract
Large reasoning models solve complex problems through long chains of thought, in which prior work has documented failure modes such as overthinking, underthinking, and unsupported deductions. Existing methods detect these failures from local signals: textual similarity to nearby steps, switch tokens, or an LLM's judgment of a step's correctness. However, what a step contributes to the solution depends on the state that earlier steps have established, so identifying failures requires a global view that tracks that state across the trajectory. Therefore, we propose ARBOR, a reasoning-failure attribution framework that represents each trajectory as a tree of reasoning states, built as each step creates, verifies, prunes, or revisits a state, and attributes each failure to a relation between a step and the state preceding it. Using ARBOR, we refine the three failure modes into distinct subtypes that harm reasoning through different mechanisms and at different costs. We find that the tree state ARBOR maintains before a step separates failure steps from the at-risk steps at which the failure did not occur. We discuss how ARBOR integrates with existing mitigations at inference time and in training data. Code, data and interactive examples are available at https://inklab.usc.edu/ARBOR.
cs.CL / 23 / 2610.22732
TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents
Abstract
Advances in frontier coding agents have sharply increased the token costs of AI-assisted software engineering. Their token efficiency is typically evaluated on isolated tasks, whereas real-world developers use them to continuously build, update, and maintain large-scale projects. This mismatch motivates us to evaluate agents over sequences of realistic coding tasks, where we find that agents exhibit recurring environment-dependent token costs that existing methods struggle to reduce. To minimize these costs, we introduce TokenEconomizer, an external controller that cuts environment-specific token costs over a sequence of tasks. These cuts come from determining (i) which operational micro-skills to inject, (ii) which persistently irrelevant tools to prune, and (iii) when to intervene in unproductive failure loops. Across Claude Code, Codex, and GitHub Copilot CLI paired with frontier models, TokenEconomizer reduces the average token cost by 18.4-30.9% with comparable task success. Savings hold in every repository under every tested agent-model setting and generalize to diverse domains such as sequential web development and security auditing, where costs fall by 6.4-21.2% depending on task structure. Our work establishes adaptive black-box cost control as a promising direction for reducing token costs of coding agents in realistic settings.
cs.CL / 24 / 2610.22733
Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents
Abstract
This study examines the transferability of a de-identification model developed for Korean court judgments to a broader range of sensitive documents, focusing on boundary precision, label granularity, and recall. The agglutinative structure of Korean causes Byte Pair Encoding (BPE) tokenizers to frequently merge Personally Identifiable Information (PII) with adjacent particles, which complicates accurate span recovery. Furthermore, current PII taxonomies lack sufficient granularity for legal texts. To address these issues, this study introduces Thunder-DeID+, which integrates a refined three-tier PII taxonomy with a boundary-aware tokenizer. The tokenizer enhances a pretrained BPE vocabulary by incorporating N-gram and boundary-scoring components to better align token and PII span boundaries. The proposed approach is evaluated using a tokenizer-independent word-segment-level protocol. On full-length judgments, the tokenizer consistently improves W. Strict Typed F1-score from 0.76-0.77 to 0.87-0.88 across four model sizes (150M, 395M, 750M, and 1B), mainly by increasing precision rather than model scale. With fine-tuning on 1,000 broader sensitive documents, the judgment-trained model further achieves 0.93 W. Strict Typed F1-score, indicating effective transfer beyond court judgments.
cs.CL / 25 / 2610.22796
Causal Latent Workspaces for Looped Transformers
Abstract
Reusing a Transformer block across depth increases computation without replicating its parameters, but successive passes have no dedicated state through which positions can exchange compact summaries of their causal prefix. We introduce the causal latent workspace Transformer: an anchored, four-round shared-depth architecture in which each position writes gated vectors into prefix-specific latent slots and retrieves a content-addressed summary before the next shared block application. This path complements causal self-attention while the original input embedding remains accessible at every round. Across three byte-level corpora, three seeds, and two matched parameter budgets, the proposed model achieves the best mean held-out perplexity in five of six settings against five Transformer controls. At 101.71M parameters it leads on WikiText-103 raw, enwik8, and TinyStories, reaching perplexities 2.7820, 2.7843, and 1.6823; these improve upon the strongest non-workspace control on each corpus by 1.06%, 0.73%, and 1.01%. All nine seed-labelled comparisons at this budget favor the workspace. At 0.918M parameters it leads on WikiText-103 and TinyStories, whereas a conditioned loop is stronger on enwik8. Architectural ablations show that adding the workspace improves upon the anchored loop in all six settings and support refreshing its causal summaries at each round. Together, these results identify a causal, content-addressable prefix workspace as an effective structural augmentation of looped Transformers.
cs.CL / 26 / 2610.22810
Watermarkability: Measuring and Improving Language-Model Compatibility with Logit-Based Watermarks
Abstract
Language-model watermarks can help identify model-generated text, encode provenance information, and detect post-generation edits. A broad class of logit-based watermarks increases the logits of a designated token set at each decoding step. Because a base model's native probability mass on this set varies across contexts, models, and watermark rules, the same logit bias can yield different adherence rates and increases in perplexity (PPL). We introduce watermarkability to characterize the compatibility between a model and the token sets requested by a watermark. At a target mass $τ$, we quantify this property through a watermarkability cost: the expected minimum KL divergence from the base-model distribution required to assign probability at least $τ$ to the requested set. A lower cost means that the model is more watermarkable for that request distribution. To reduce this cost, we propose a two-stage method. In the first stage, we train a target-set-conditioned low-rank adapter on unions drawn from a 16-way vocabulary partition to increase the requested set's probability mass. In the second stage, a closed-form controller applies the smallest remaining uniform logit bias needed to reach the target mass. We evaluate the method on six watermark schedules using OPT-125M and Qwen3-8B. Across the 12 model-schedule pairs, the pre-adaptation watermarkability cost correlates with the log-PPL increase under a near-hard fixed-bias baseline (pooled Pearson $r=0.929$). For every schedule, at least 99.80% of sampled tokens fall in the target set requested at their decoding step. At this reliability, the adapter and controller reduce average per-schedule prompt-conditional PPL by 46.3% on OPT-125M and 5.2% on Qwen3-8B relative to the fixed-bias baseline.
cs.CL / 27 / 2610.22820
Causal Edit Serialization
Abstract
Standard autoregressive language models generate text sequentially, but revision requires referring back to tokens in the input context or those already generated. Causal Edit Serialization (CES) is introduced as a framework for expressing such revisions within the same causal token stream. Special pointer tokens are paired with predicted source addresses, allowing insertion, deletion, replacement, copying and movement to be localized explicitly. A lightweight localization head and dual-index rotary position encoding support both address prediction and conditioning on previously emitted pointers, while retaining a standard causal attention mask. A pretraining procedure combines ordinary language modeling with serialized repairs derived from random corruptions, model-generated edits and repeated text. Across model scales and editing baselines, CES models demonstrate complete applied edits with modestly higher clean-text bits per byte than clean autoregressive models. Supervised and reinforcement-learning post-training provide a grammatical-error-correction proof of concept, producing more beneficial than harmful changes under reference-blind model judging. Correct repair programs also improve subsequent text prediction while the original context and its cached states are retained. CES provides a common interface for generation and localized revision, with potential extensions to iterative refinement and additional pointer-addressed transformations.
cs.CL / 28 / 2610.22874
Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data
Abstract
LLMs over external data face a per-instruction tension: a single document or tool return often interleaves legitimate directives with adversarial prompt injections, and the model must follow the former while rejecting the latter. Existing defenses suppress or trust all external instructions, collapsing the safety-utility frontier and forcing a choice between over-refusal and over-compliance. We argue the missing capability is \textbf{per-instruction credit attribution} and formalize it as External-Data Instruction Classification, an identify-classify-act decision over each embedded directive. We propose \textbf{EDIC}, a decomposed reward framework for per-instruction compliance. A scene-grounded pipeline injects mixed beneficial and harmful instructions with counterfactual filtering, while a generative judge is initialized by SFT and refined with GRPO to attribute compliance at the instruction level. The judge then produces a reward that separately encourages beneficial compliance and penalizes harmful compliance for RL training. Across two backbones and five benchmarks, EDIC improves both safety and utility over the strongest prior defense, lifting harmful-rejection by up to 7.7\% and Arena-Hard-Auto utility by 12.5\% on Qwen3-4B. The code repository is hosted at https://github.com/yingyingxia666/EDIC.
cs.CL / 29 / 2610.22892
Bengali Hallucination Detection with Retrieval and Programs: A System Study
Abstract
Bengali hallucination detection must distinguish unsupported factual claims from incorrect reasoning, often without a supplied reference passage. We present and characterize an inference-only system that combines response-conditioned query rewriting, hybrid retrieval, and task-specific language-model judging. Six retrieval routes search question-answer and passage collections; a router directs each example to a 27B textual judge or a 14B agent that can compare retrieved answers, reason directly, or generate and execute Python. We characterize the system using version-specific code, saved predictions, execution traces, and corpus metadata from the Olikbochon challenge. Kaggle associates a displayed private macro-F1 of 0.956 with Version 2; its saved run processes 2,516 examples in 220.99 minutes on two T4 GPUs. Of 253 math-routed examples, 88 terminate through candidate matching, while execution traces expose 49 stderr-bearing run events across 35 examples. A later run changes six predictions despite identical core worker code. A bounded audit of 1,090 benchmark-related corpus records finds exact question or context overlap for 38 processed evaluation inputs, precluding an unqualified claim of corpus-independent generalization. The contribution is a concrete retrieval-and-computation judging architecture with an auditable operational characterization, not a controlled estimate of component gains. We distinguish the reported leaderboard result from independently recomputed diagnostics and identify the evidence needed for stronger effectiveness and reproducibility claims.
cs.CL / 30 / 2610.22906
FederatedBytes: Tokenizer-Free, Byte-Level Small Language Models for Efficient Multimodal Federated Learning
Abstract
Federated Learning (FL) is a powerful machine learning framework that allows building a global model from the contributions of multiple, distributed clients, while ensuring the privacy of their local data. However, its heavy computational and communication overheads challenge its use for contemporary AI workloads such as natural language processing (NLP) and multimodal analysis. For this reason, this paper presents a novel small language model that leverages byte-level text and multimodal data processing, enhanced by a linear attention mechanism, to enable manageable Federated Learning of NLP and multimodal analysis tasks. The proposed model, nicknamed FederatedBytes, manages to process text and any other data types, including images, by treating its input as a mere sequence of bytes, while its lean linear attention mechanism allows abating the overall memory and computational footprint. To prove the effectiveness of the proposed model we have carried out an extensive performance evaluation on sentiment analysis tasks in multilingual and multimodal settings over the Multilingual Amazon Reviews and ScienceQA datasets. The experimental results show that FederatedBytes has been able to outperform comparable models such as ByT5 and LSTM, achieving higher accuracy while significantly reducing memory and communication overheads.
cs.CL / 31 / 2610.22919
Where's the Indirect Evidence? Investigating Grammatical Generalization with Influence Functions
Abstract
Language models generalize beyond their training corpus to learn linguistic phenomena (Patil et al., 2024; Misra and Mahowald, 2024), showing that there exists indirect evidence driving learning. Prior work has inferred that indirect evidence exists without identifying its character or considered specific hypotheses about what it might be. We explore an alternative bottom-up method for identifying indirect evidence. We use influence functions (Koh and Liang, 2017) to measure how individual training sentences affect model predictions in 15 grammatical phenomena from BLiMP (Warstadt et al., 2020). We filter out a phenomenon's direct evidence using hand-constructed rules and take the 60,000 most influential remaining sentences as indirect evidence. This indirect evidence substantially affects learning: retraining without it lowers held-out accuracy more than removing a matched random group for every phenomenon across five models (by a median of 12 points). We characterize the indirect evidence by its similarity to our direct evidence filters: it (partially) meets the filters' structural conditions at a median of 4.7 times the rate of a matched random group. Finally, clustering phenomena by indirect-evidence overlap recovers three of the five BLiMP categories that include multiple phenomena, suggesting that this method can reveal linguistic relationships.
cs.CL / 32 / 2610.22934
ASR-Style Transcript Normalization Removes Useful Pragmatic Signal for Dialogue Act Classification
Abstract
Automatic speech recognition (ASR) pipelines often normalize transcripts by removing filled pauses and punctuation and expanding colloquial forms, potentially discarding cues useful for dialogue act classification. We evaluate nine transcript conditions across six classifiers (TF-IDF logistic regression, RoBERTa, BERT, ALBERT, GPT-2, and BiLSTM) and three conversational corpora (Switchboard, ICSI MRDA, and SLUE-HVB), supplemented by multi-seed experiments, normalization controls, and real-ASR evaluations. Across five random seeds, heavy normalization reduces Macro-F1 in all 15 neural model-corpus combinations under the original preprocessing protocols. Removal using the original filler and backchannel lexicon consistently degrades performance on MRDA and SLUE-HVB but produces mixed effects on Switchboard. Punctuation-aware TF-IDF experiments show that punctuation removal reduces Macro-F1 by 0.036 on Switchboard and 0.078 on MRDA, demonstrating that the original linear baseline's apparent insensitivity resulted from tokenization. Restricting filler removal to um, uh, and erm preserves degradation on MRDA but largely eliminates the SLUE-HVB effect, where the original lexicon removes class-defining backchannel expressions. These findings demonstrate that normalization effects depend on preprocessing definitions, feature extraction, and conversational context, supporting task-dependent rather than indiscriminate transcript normalization.
cs.CL / 33 / 2610.22937
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
Abstract
Language models often abandon a correct answer when the prompt contains a cue, a hint toward a wrong option such as "I think the answer is B." Sycophancy and related biases show this failure. They are often studied separately and blamed on alignment. We test this view causally, tracing the attention heads behind these biases through the stages of two post-training pipelines. In multiple-choice tests on most models, four cue types rely largely on the same few heads, a shared circuit already present in pretrained models. Most base models follow cues in plain text. Fine-tuning a base model on about 200 generic assistant examples without cues activates these heads in chat format, so cue-following can be a side effect of assistant tuning. The final reinforcement learning (RL) stage strengthens the effect of cues in almost every comparison. In Tulu-3-8B and OLMo-2-7B, we swap the attention patterns of six middle layers between the RL checkpoint and its predecessor. For suggested-answer cues, this shifts each checkpoint's cue effect most of the way to the other's level. Post-training thus seems to change when and how strongly models use this circuit. Mitigation for these multiple-choice biases could target this circuit.
cs.CL / 34 / 2610.22953
Recognizing Atari and Assessing Sente: Controlled Evaluation Tasks for Language-Based Go Agents
Abstract
Domain terminology provides concise descriptions of game states and strategic relationships, but whether game-playing language models can accurately recognize these concepts and relate them to move evaluation remains unclear. We construct a controlled evaluation set centered on atari and its relationship to sente in Go, examining terminology recognition under local board changes and move evaluation under different strategic conditions. We conduct our tests on LoGos, a Go language model, after fine-tuning it with terminology and commentary supervision. The model struggles to recognize changes in the atari relationship and assess its strategic implications. Its terminology predictions are associated with move preferences even when those preferences contradict the reference evaluation. Our controlled tests provide a framework for examining whether language-based Go agents connect domain terminology with precise board relationships and strategic move values.
cs.CL / 35 / 2610.23093
Disentangling Language Effects on Tool Admission and Post-Retrieval Safety
Abstract
LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric. It remains unclear whether language affects tool admission, post-retrieval generation, or both. We independently vary request, retrieved-page, and required response languages in a controlled URL-retrieval setting across ten models and seven languages. We identify Safety Loss through Interlingual Processing (SLIP). Non-English harmful requests increase mean tool admission from 46.9% to 67.4%, with a positive mean change in every model. This pattern also persists in an autonomous retrieval setup across all four models tested in this setting. Requiring answers in the page language increases pooled harmfulness on matched tool calls from 50.3% to 57.4% across four closed-weight models, mainly in Chinese and Hindi. Translating pages alone produces no consistent harm increase in eight models with comparable inference settings. Automated audits show that a non-harmful response may be a genuine safety refusal, a deflection citing language difficulties, or an unusable answer. Harmfulness labels alone cannot distinguish these outcomes. Multilingual safety evaluations should therefore measure tool admission and response harmfulness separately and distinguish refusal from capability failure.
cs.CL / 36 / 2610.23166
Explaining Chess and Go in Natural Language
Abstract
Modern chess engines play far beyond human ability, often choosing moves whose underlying ideas are difficult for humans to understand. We propose an automatic metric for evaluating chess explanations, grounded in chess-engine evaluations. Evaluating state-of-the-art language models using this metric, we find that they perform surprisingly poorly, often relying on surface-level board observations rather than robust, transferable chess concepts. We then introduce BinProp, a training-free method that enables these same models to discover reusable chess concepts and use them to explain moves in natural language. BinProp improves explainability for every model tested, nearly tripling the average score compared with direct prompting. Beyond chess, the same method discovers useful concepts in Go, draughts, and shogi without additional tuning.
cs.CL / 37 / 2610.23170
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
Abstract
Many language-based applications require probabilities over a finite set of decisions rather than open-ended responses. We present Chinese Decision, a family of Chinese-language models for single-pass probabilistic decision-making. Given context, an instruction, and candidate descriptions, the model extracts final-position candidate-letter logits in one forward pass and normalizes them over valid options, without autoregressive answer generation. A shared interface supports Boolean judgments, categorical choices, and ordinal scores. We report a completed 0.8B model trained with distribution cross-entropy on 13.048 million decisions and preliminary 2B checkpoints trained with an additional RLCD objective. On the full CJ-Bench of 307,900 decisions, the uncalibrated 0.8B model achieves 72.24% general-domain accuracy and 76.98% average accuracy across medicine, law, and finance using a single checkpoint, with 1.07% overall expected calibration error. The 2B model reaches 72.56% general-domain accuracy after 40.13% of one epoch but has not surpassed the completed 0.8B model on the professional-domain average. We further evaluate a validation-temperature-scaled deployment model and its Q8 quantization, finding close aggregate performance while documenting individual prediction changes. We release the training, evaluation, and deployment pipeline. The current results establish a reproducible approach to Chinese probabilistic decisions but do not isolate the effect of model scale.
cs.CL / 38 / 2610.23175
Do Speech BCIs Need Larger Models? Rethinking Neural Decoding beyond Scaling
Abstract
Recent speech brain-computer interfaces (BCIs) increasingly rely on large models and complex multi-stage pipelines for neural decoding. However, in speech BCI, where training data are scarce, noisy, and expensive to collect, the benefits of model scaling remain unclear. This work revisits a fundamental question: are large models truly necessary for accurate neural decoding? We propose BEST (Brain-Encoding-Speech-Text), a two-stage framework for efficient neural-to-text decoding. Instead of relying on large generative models or phoneme-centric pipelines, BEST introduces hierarchical intermediate representations and a compact speech-oriented decoder that directly maps neural activity to text. This design preserves essential structure while significantly reducing computational cost, enabling robust generalization in low-resource settings and practical deployment under constrained resources. On the Brain-to-Text '24 and '25 benchmarks, BEST achieves word error rates (WERs) of 6.19% and 1.62%, respectively, using only 3.85% of the parameters required by comparable methods. Beyond these results, we provide three key insights: (1) decoder scaling yields non-monotonic gains; (2) cross-modal alignment is more critical than model size; and (3) overfitting degards sequence-level performance. Representation analysis further reveals a consistent neural-audio-text alignment pathway.
cs.CL / 39 / 2610.23183
AMEND: Constraint-Guided Multi-Agent Text Detoxification with Isolated Critics
Abstract
Text detoxification requires balancing toxicity removal, semantic preservation, and fluency. Existing methods typically optimize an unconstrained scalar objective---enabling safety scores to mask semantic loss---or rely on open-ended iterative debate prone to semantic drift and high computational cost. We propose AMEND, an inference-time framework that decouples detoxification into role-conditioned proposal generation, isolated multidimensional criticism, and constraint-first selection. AMEND evaluates diverse candidates with role-blind critics and enforces strict semantic and fluency feasibility before ranking by utility. To maximize efficiency, AMEND Lite restricts subsequent search to at most one targeted repair round only when residual toxic spans remain. On ParaDetox, one-round selection over diverse roles improves the joint evaluation metric by $+0.0508$ over the best single role. On a frozen confirmation test set, AMEND Lite achieves a joint score of $0.6597 \pm 0.0062$, outperforming unconstrained multi-round loops by $+0.0489$ ($p < 0.05$) while reducing inference requests by 28.4%.
cs.CL / 40 / 2610.23234
MultiAgent-OPD: End-to-End Training of Hierarchical Multi-Agent Systems with On-Policy Distillation
Abstract
In a hierarchical multi-agent system, a Leader assigns tasks and subagents carry them out. Jointly training the Leader and the subagents is difficult because a final answer gives only indirect feedback to the subagents. An LLM judge can score the performance of the subagents, but these scores depend on the judge's own assessment of the quality of subagents' responses. We present MultiAgent-OPD, which combines on-policy distillation with reinforcement learning to jointly train a shared-parameter Leader-Searcher-Answerer team end to end. We first train a Teacher in a Searcher-Answerer workflow using answer correctness. During team training, the frozen Teacher scores student-generated search and summary tokens to provide local process supervision for the Searchers, while the Leader learns from team success and the Answerer from candidate correctness. Across seven datasets, the trained system reaches 38.68% macro avg@4, exceeding Prompted Critic by 4.06 percentage points and C-MAS + RL, the trained cascaded multiagent system, by 2.61 percentage points. The team also solves more questions within four attempts.
cs.CL / 41 / 2610.23245
MGE-PO: Entropy-Aware Credit Assignment for Mathematical Reasoning
Abstract
GRPO assigns the same group-relative advantage to equally rewarded responses within a question, leaving their generation uncertainty unused. We introduce MGE-PO, centered on four-quadrant (4Q) advantage shaping that combines verified correctness with rollout generation entropy. Correctness determines the sign of an additional correction, while entropy controls its magnitude: lower-entropy successes receive larger bonuses and higher-entropy failures receive larger penalties. An online entropy-based question curriculum supports this credit rule using cached rollout measurements. At the final checkpoint after eight training epochs on Qwen3-8B-Base, adding 4Q across three policy optimization settings yields absolute gains of 3.23-3.69% in six-benchmark macro avg@32 and 4.24-11.06% in macro pass@32. On Qwen2.5-7B-Base, the complete configuration raises four-benchmark macro avg@32 from 41.5% to 45.4%.
cs.CL / 42 / 2610.23246
Think More or Read More? A Token-Budget Theory of Reasoning with Shared Context
Abstract
Reasoning systems increasingly improve answers by incorporating additional context into the model's input, from multi-agent debate, where agents read their peers' responses, to self-refinement, where a model revisits its own drafts. In these systems, communication consumes context as one agent's output becomes another's input. We study the regime where token reading and generation draw on one fixed token window, so every token spent reading is a token that cannot be spent reasoning. We develop a \textbf{token-budget theory} that formalizes this trade-off, modeling accuracy as a function of the reasoning budget remaining after reading and aggregating votes with a Condorcet jury model that permits correlated errors. For multi-agent debate, the theory gives a closed-form prediction for the optimal number of agents. This quantity is estimated from single-agent measurements, allowing the optimal group size to be determined without running debate, and the same accounting extends to a heterogeneous set of different models. We empirically verify our theory and show that, at the predicted group size, our method outperforms strong baselines under the tightest calibration spend, where debate-calibrated rules can afford only one or two questions. Self-refinement obeys the same accounting, with its own draft in place of a peer message. The theory further predicts that communication pays only when the gain in the accuracy ceiling exceeds the reasoning room spent on reading, with this condition tightening as the window shrinks. Across five open models and four reasoning tasks, self-consistency matches or beats deployed debate at $93$\% of $980$ matched window-and-group-size settings, with the gap being largest under tight windows and nearly disappearing at a loose one, consistent with the theory. Under a shared window, how much room a model has left to \textit{think matters more than what it reads}.
cs.CL / 43 / 2610.23254
Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents
Abstract
Tool-using agents read untrusted external content, which exposes them to indirect prompt injection (IPI), where hidden instructions hijack the agent's tool calls. Benchmarks measure IPI security by the attack success rate (ASR), which treats every failed attack as a successful defense. To test this assumption, we audit 33,000 trajectories from eleven open-weight agents (2-14B parameters) on InjecAgent and AgentDojo. A label-blind LLM judge shows that most failed attacks are ignored (76-80%) rather than refused (17-18%). With identical model weights, changing only the system prompt shifts InjecAgent ASR from 0.2% to 43%. Our Attack-Task Outcome Matrix (ATOM) further shows that in 21.7% of AgentDojo episodes, the agent stayed safe but did not complete the user's task. A low ASR therefore mostly reflects agents silently ignoring injections, not refusing them. Evaluations should report how attacks fail and whether the user's task is completed, alongside ASR. We release the labeled corpus, judge prompt, and analysis code.
cs.CL / 44 / 2610.23275
Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
Abstract
Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation. We introduce VERITAS to verbalise the model's computation: how different reasoning anchors emerge, transfer across layers, and surface in the output. VERITAS adopts the layer-wise Jacobian projection to produce a verbalised account of the model's internal reasoning evolution at each generated CoT token. By training a classifier on the extracted features along the traces, we show that these structured features provide insightful perspectives on how and where reasoning diverges toward unfaithfulness. Experiments on four datasets from FaithCoT-Bench and BonaFide show that VERITAS achieves a significant performance gain over LLM-as-Judge, with a 28.4% gain in balanced accuracy, and an average improvement of 16.0% over activation probing. Further analysis shows VERITAS can be generalised to different datasets and model architectures. Our work shows that verbalising model computation provides a richer and more interpretable lens on CoT faithfulness, moving beyond detection toward a mechanistic understanding of how reasoning fails.
cs.CL / 45 / 2610.23309
Inferring L1 Backgrounds from L2 English with LLM Token Probabilities
Abstract
Native language identification (NLI) infers a non-native speaker's first language (L1) from their second language (L2) production. We investigate a token probability approach that scores each text sample as a continuation of a prompt prefix specifying a candidate L1, then predicts the candidate with the highest score. We evaluate our approach using Llama 3.1 8B and 70B on three datasets: TOEFL9, Write and Improve 2024, and BULATS-L1. We compare its performance with previous prompting-based classification approaches. Token-probability scoring substantially improves on direct classification, particularly with the 8B model, while the 70B model achieves state-of-the-art performance across all datasets. Our results indicate that pretrained LLMs' token probabilities contain information useful for distinguishing L1 backgrounds from L2 English production.
cs.CL / 46 / 2610.23388
Sentence-Level Topic Modeling with Lexical and Contextual Evidence
Abstract
Topic modeling faces a fundamental granularity dilemma: probabilistic topic models assign topics at the word level, whereas transformer-based methods operate at the document level. The latter overlooks that real-world documents typically cover multiple topics, whereas the former relies on individual words, which rarely provide sufficient context to identify them. We introduce SentenceETM, a neural topic model that assigns a topic to each sentence by jointly using lexical and contextual evidence. Each topic is described both by a distribution over the vocabulary, defined through word embeddings, and by a Gaussian over sentence embeddings concentrated on a topic-specific low-dimensional subspace. The topic of a sentence generates both its words and its embedding and is marginalized exactly, so that its posterior combines the two sources of evidence. Document proportions are inferred with amortized variational inference, while the Gaussian parameters are re-estimated in closed form. Experiments show that SentenceETM consistently improves sentence-level topic assignment over state-of-the-art approaches, and a case study of a presidential debate shows how sentence-level topics reveal how candidates handle the same subjects.
cs.CL / 47 / 2610.23455
HydraLLM: Adaptive Width Slicing for Elastic Text and Vision-Language Models
Abstract
LLMs and LVLMs are increasingly deployed across heterogeneous hardware, ranging from data-center GPUs to resource-constrained edge devices. Their fixed-size architectures limit the ability to reuse a single model across deployment settings with different compute, latency, and model-size targets. We introduce HydraLLM, an elastic MQA-based architecture. Its number of active attention heads determines the effective widths of the embeddings, attention projections, feed-forward layers, and the output head, forming a nested family of subnetworks within a single parameter set. We introduce a static-to-elastic conversion pipeline that, starting from a pre-trained model checkpoint, includes importance-driven weight reordering and stochastic subnetwork training with knowledge distillation. We extend this principle to multimodal inference through an elastic projection module whose output width follows the active text width. HydraLLM covers a wider range of active parameter configurations than FFN-only elastic architectures. With 75% of heads active, HydraLLM achieves 2.2 percentage points higher TriviaQA accuracy than a state-of-the-art nested baseline configuration which has 17.25% more parameters. With 50% of heads active, our multimodal model achieves a CLIPScore comparable to that of the 7$\times$ larger Gemma 3 4B model. With only 25% of heads active, HydraLLM increases token-generation throughput by 75.3% over the full-width configuration.
cs.CL / 48 / 2610.23456
A low-dimensional dynamical basis for tongue shape control
Abstract
Speech production involves the control of a high-dimensional articulatory system with many degrees-of-freedom, but a widespread hypothesis is that speakers navigate this complexity using a set of lower-dimensional control variables. Common candidates for these control variables include individual articulators (e.g. tongue body) and abstract tract variables (e.g. tongue body constriction location), but identifying the most appropriate coordinate system for articulatory control remains an open question. In this study, we advance a model in which holistic tongue shape modes represent a potential low-dimensional control basis for a simple dynamical law. We first discover empirical tongue shape modes in a cross-speaker coordinate space and then fit the model to ultrasound data on monophthongal vowels from 36 speakers of British English. A three-dimensional linear point attractor fitted to these time-varying modes recovers the dynamics of tongue shape with a high degree of accuracy. This suggests that complex tongue movement can arise from independent modes that linearly combine under simple dynamical laws, thereby relocating some complexity from the control signal to the coordinate system. We show that this framing explains why one-dimensional articulatory measures can exhibit observable nonlinearities, and hypothesize that shape modes are a plausible kinematic expression of underlying muscle synergies.
cs.CL / 49 / 2610.23480
MuninAI: A Self-Hosted Agentic Framework for Independent Research Groups
Abstract
Research groups that want a language model grounded in their own literature face a choice between hosted services that receive their unpublished work and self-hosted deployments that keep the data local but give the model no access to the corpus. Munin is an open-source framework for a self-hosted scientific assistant that indexes a group's corpus and its citation graph, and wraps whichever open-weight model the group can serve in an agent harness with specialised reading, search, and computation tools. We claim that the harness, rather than the specific model, is what makes the system useful on domain literature, and we ship the evaluation suite that tests this claim with the software. On 199 paired literature questions, the harness more than doubles the accuracy of the base model, from 0.387 to 0.874 (+0.487, 95% CI [0.407, 0.568]), while naive retrieve-then-answer scores below the unaided model on every backbone tested. Under the harness, roughly twice as many answer claims trace to a passage the model read itself as with fixed top-5 retrieval. When the source paper for a question is removed from the corpus, the system abstains on about nine in ten of the questions it had answered correctly on the Qwen backbones and eight in ten on gpt-oss-20b, and it produces no fabricated local citation across 300 fabricated references. The pattern holds on three backbones from two developers, including a 20-billion-parameter model that runs on a single consumer-grade GPU. With all outbound network access disabled, both Qwen backbones still exceed the published PaperQA2 accuracy. The evaluation suite runs as a single command, so an adopting group can re-run it on its own corpus and model.
cs.CL / 50 / 2610.23530
EduWild: How Teacher-Configured Chatbots Fail in Real-World Use and How to Fix Them
Abstract
Teachers increasingly configure educational chatbots through natural-language instructions, but how these bots fail in real-world use remains poorly understood. We introduce EduWild, a dataset of 1,726 conversations with 217 teacher-configured chatbots, each paired with the instructions and configuration behind it, including 900 conversations from the intended end users of 35 deployed chatbots. We develop a 29-code taxonomy covering factual, workflow, diagnostic, and pedagogical failures and apply it using a human-calibrated LLM judge. The most frequent failures are lapses in following teacher intent --- bots skip required workflow steps (53.8% of end-user conversations), omit expectation setup (33.4%), and overload learners with information or questions (37.1%). Guided by these findings, we design a prompt-based harness targeting the six most frequent failure modes. In a study with 200 college students across ten bots, the harness reduces the mean number of target failure codes per conversation from 1.406 to 0.489, a 65.2% relative reduction with bots weighted equally. These findings show how real-world failure analysis can guide practical improvements to educational chatbots.
cs.CL / 51 / 2610.23539
ConvexLLM: Scaling Convex-Hull Speech Interfaces to General-Purpose Audio Intelligence
Abstract
Speech and audio language models connect continuous acoustic representations to decoders pretrained primarily on text data. Most such models use MLP or Q-Former interfaces to map acoustic features into the decoder's input space, requiring the decoder to learn how to interpret the resulting representations. We investigate whether grounding these representations in the decoder's existing text embeddings can help it use acoustic information while better preserving its pretrained text capabilities. We present ConvexLLM, a general-purpose speech and audio language model whose C-Gate interface represents audio as continuous convex combinations of frozen text embeddings. ConvexLLM achieves an eleven-task VoiceBench aggregate of 75.36, the highest among the open-weight models we compare, and an MMAU Test-mini accuracy of 62.50%. Its largest margins over the baselines are on spoken knowledge and reasoning tasks, while the strongest baselines are more accurate on audio understanding and transcription. In a comparison with a 2B-parameter decoder, C-Gate achieves higher aggregate spoken-language performance than MLP and Q-Former interfaces, with text-task performance close to MLP and task-dependent tradeoffs in transcription and audio understanding. These results support convex combinations of text embeddings as a viable interface for general-purpose speech and audio language models.
cs.CL / 52 / 2610.23583
MARS: Memory-Aware Routing for Session-Level LLM Assistants
Abstract
Long-term LLM assistants must repeatedly decide what prior context to expose and which model should answer as an interaction evolves. Existing memory systems and model routers usually make these decisions separately, although memory construction changes the evidence, distractors, and difficulty of the resulting input. We introduce MARS, a memory-aware routing framework that uses session state as a shared control interface between memory construction and model selection. A Query and Topic Manager (QTM) estimates compact decision-relevant state that guides a fixed state-conditioned memory policy and an online router, enabling each turn to receive an appropriate memory view and a suitable model under a quality-memory-cost trade-off. We evaluate MARS on RealMem and MultiChallenge with 16 system configurations each, covering long-history memory-driven interaction and bounded sessions with structured context interference. On RealMem, MARS preserves 98.2% of full-context GPT-4o's QA while using only 14.5% of its generation API cost, including both generation and controller calls for MARS, and achieves the strongest measured memory-recall and retrieval scores among compared systems. On MultiChallenge, MARS trails the strongest static system overall but approaches its performance at lower recorded cost on a mechanism-focused subset. Together, these results support a favorable quality-memory-cost trade-off for long-history, memory-dependent interaction.
cs.CL / 53 / 2610.23587
Analyzing Occupational Gender Bias across the LLM Training Lifecycle
Abstract
Post-training is widely assumed to reduce the social biases language models inherit from pretraining, yet direct evidence for this assumption remains mixed. In this paper, we examine how gender disparities evolve across the language modeling pipeline. In particular, we study occupational gender bias in OLMo 3, a fully open language model family, across its post-training stages by prompting each checkpoint to generate open-ended narratives about synthetic personas given their occupation and education. We consider two settings, a gender-given setting in which the profile states the persona's gender explicitly and a gender-assumed setting in which it must be inferred by the model. We find that when gender is stated, post-training increases semantic diversity for male-coded profiles, whereas for female-coded profiles the increase is smaller after SFT and not significant after DPO and RLVR. When gender is assumed by the model, post-training affects narratives about female and male personas similarly. The model's own gender assumptions, however, show a consistent female-preferential skew relative to United States labor-force demographics, except for heavily male-dominated occupations. Every successive post-training stage monotonically increases the odds of attributing a female label over the base model. A provenance analysis of the supervised fine-tuning corpus finds only a modest 7.7% female-token surplus, too small to explain the size of these downstream effects. The project code is available under: https://github.com/manthangadhia/llm-lifecycle-bias.
cs.CL / 54 / 2610.23612
SIT-Eval: Situation-Based Interactive Turn-by-Turn Evaluation of Conversational Agents
Abstract
The rapid proliferation of conversational AI agents necessitates evaluation methods that assess performance across diverse dialogue scenarios, utilizing multiple situational criteria, as appropriate behavior depends on context and encompasses various aspects. Existing dialogue benchmarks typically use static natural conversations, which underrepresent rare situations---a critical safety gap, as low-probability events become frequent failures at production scale. However, limited sample sizes force a choice: either maintain natural dialogues for realism or turn to adaptive situation sampling for comprehensive yet statistically reliable coverage. We introduce SIT-Eval, an adjustable framework for evaluating situation-based conversational agents. We formalize an agent's desired behavior as situational rules: if situation is P, follow rules Q. At each conversation turn, an LLM-interlocutor enacts a uniformly sampled situation. A context-aware LLM-judge leverages all conditionally applicable rules as criteria to evaluate the agent's reply for that situation. The judge is validated using human-verified synthetic golden dialogues. Benchmarking 8 LLMs on 1292 conversation turns reveals failure modes hidden by aggregate metrics: models with similar overall scores vary widely in their adherence to individual rules. All models struggle most with ambiguity clarification and Socratic questioning.
cs.CL / 55 / 2610.23613
Trajectory-Triggered Sparse Collaborative Decoding
Abstract
Dense assistance methods such as speculative decoding repeatedly verify draft tokens with a target model, keeping target computation coupled to generation. We study sparse collaborative decoding, where a draft model generates most tokens and a target model is invoked only for short corrective segments. The key problem is deciding when to call the target. We show that local uncertainty is a weak trigger: entropy, negative log-likelihood, logit margin, top-$k$ mass, and velocity features give near-chance failure prediction, with AUC 0.50-0.53. In contrast, aggregate trajectory-shape features over the draft confidence path reach AUC 0.86 using a lightweight logistic predictor trained only from draft correctness labels. This induces a trigger-timing tradeoff: early calls have more leverage but weaker evidence, while later calls have stronger evidence but less ability to redirect the trajectory. Used online, learned sparse collaborative decoding improves over the best open-loop schedule by +5.3 to +20.0 accuracy points across five GSM8K draft-target pairs. On MATH-500, it gains 7.0-13.5 points when early prediction and target competence align, but loses when either is weak. On StrategyQA, it improves over open-loop by 5.0-8.3 points, showing transfer beyond arithmetics.
cs.CL / 56 / 2610.23651
In Human Memories. Technology, Creativity and Self-Determination: An Ethical Proposal
Abstract
This research investigates the role of artificial intelligence (AI) in modern compositional practices, with a particular focus on two case studies: Convergence (2020/21) by Alexander Schubert, for String Ensemble and AI System, and In Human Memories (2024) by Francesco Vitucci, for four cellos and one electroacoustic player, whose installation was created by the Electronic Music School of the "N. Piccinni" Conservatory of Bari. The initial historical survey traces the evolving use of technology in composition, revealing how the search for new expressive possibilities has driven the development of technological tools through experimentation. The appearance of AI marks a fundamental shift: no longer just a tool, AI now poses questions and challenges to the composer. Engaging with AI becomes an exploration of an unknown entity, whose full capabilities are still largely undiscovered. This new interaction leads composers to confront AI not merely as a passive resource but as a participant in the creative process, outlining an ethical path to reclaim technology as a creative tool. In Human Memories reflects this dynamic, demonstrating how AI can serve as both inspiration and a driving force in shaping contemporary music.
cs.CL / 57 / 2610.23702
VISTA: Steering LLMs with Latent User Values Inferred from a Few Survey Answers
Abstract
Personalizing a language model requires knowing what a person values, which stays largely consistent across tasks. Yet most methods either learn task-specific preferences or place the user's profile in the prompt, where models struggle to act on it. We introduce VISTA (Value-Informed Steering and Targeted Acquisition), which infers a latent profile from users' (full or partial) survey answers, and maps the profile through a hypernetwork to a user-specific steering intervention in a frozen LLM to select answer choices or generate free-text responses. Given the same evidence, VISTA matches or beats in-context profiles on synthetic preferences (Food2D), the World Values Survey (WVS), and PrimeX, without adding prompt tokens. To reduce the cognitive burden of completing full surveys, VISTA also chooses which questions to ask and when to stop: on WVS it stops after 6% of the questionnaire while retaining 94% of the full-history gain, where standard adaptive testing retains at most 42%. The dominant steering direction is an interpretable value axis on which VISTA places unseen users accurately. Our results point toward pluralistic alignment that adapts to individuals at deployable cost.
cs.CL / 58 / 2610.23709
GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search
Abstract
Optimizing product listings for generative search requires language agents not only to rewrite effectively but also to continually self-improve by retaining successful strategies across diverse categories. However, existing approaches often treat structured listings as free-form text, leaving effective strategies implicit in prompts, search trajectories, or rewritten outputs rather than accumulating them as a reusable agentic memory. To address this limitation, we introduce GLOBE, an agentic framework that drives continual self-improvement by storing optimization knowledge in an evolving handbook of explicit, field-scoped behaviors. In each episode, an Editor applies relevant behaviors and receives rank-based feedback, which a Reflector converts into behavior updates. Crucially for dependable execution in real-world settings, a Moderator removes redundant proposals, a Curator commits accepted changes, and consistently harmful behaviors are deactivated. This cycle transforms episodic feedback into a persistent, field-scoped optimization memory. Experiments show that structured behaviors outperform monolithic prompt optimization, while feedback-driven evolution and moderation further improve effectiveness and reduce downside risk. The learned behaviors also remain effective on unseen product categories, suggesting that agentic memory evolution supports interpretable and transferable product optimization.
cs.CL / 59 / 2610.23721
Recognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM-Based Medical Verification
Abstract
Publication count is not study count. A model can identify shared trial lineage correctly in a direct task yet over-credit a second publication from the same trial when aggregating evidence. We evaluate this behavior in medical LLM verification using 14 source-audited clinical families and five model conditions. The families span provenance, clinical applicability, result direction, and composition; the provenance families provide the direct lineage tests. Recognition and aggregation are separate stateless calls, and repeated calls assess within-input stability. Across all three original provenance families, GPT-5.5 identifies the independent update in 10/10 recognition calls per family but over-credits the same-trial report in 5/5 separate aggregation calls per family. Selected Qwen3-4B and Ministral-3-8B conditions show the same cross-task contrast. A study-grouping instruction reduces GPT-5.5 over-credit from 4/5 to 0/5 in each later pure-provenance family. Other model conditions show mixed behavior, and rejected outputs limit several comparisons. The results identify a behavioral gap between elicited lineage recognition and evidence aggregation. The code and data artifacts are available at https://anonymous.4open.science/r/study-lineage-review-artifact.
cs.CL / 60 / 2610.23734
Bridging the Language Gap in Scholarly Data II: Enhancing Gender Inference for Chinese Names
Abstract
Perceived gender inference from names is important for inclusive scientometrics research, yet existing methods perform poorly on non-Western scholars. We address this gap for Chinese names, which appear in two forms, Hanzi and Pinyin, its Romanized form (e.g., Zhang Wei, Li Na). In international bibliographic records, names are often available only in Pinyin. However, a single pinyin name can correspond to multiple distinct hanzi characters with different gender associations, so gender inference from Pinyin alone is intrinsically ambiguous. To address this, we first develop two hanzi models, a regression model that produces a continuous gender score from 0 (male) to 1 (female), and a classification model that outputs one of three labels (\em{male}, \em{neutral}, \em{female}). We train them on a sample of hanzi names from Chinese-language physics publications, rated by 11 native Chinese speakers. On this sample, the models achieve an RMSE of $0.09$ and an R$^2$ of $0.85$ (regression), and macro-AUC of $0.97$ and a Cohen's $κ$ of $0.81$ (classification), demonstrating high performance that is robust to class imbalance. For pinyin names, we propose a pipeline that aggregates hanzi predictions via transliteration and frequency weighting, together with an ambiguity score that flags names whose gender cannot be resolved from Pinyin alone. Benchmarked against three off-the-shelf tools (Namsor, GPT, and Claude), our models achieve the highest agreement with human annotations for both gender scores and gender labels, in both scripts. Applied to an independent sample of Chinese-named authors in APS journals, our models assign a gender to every name, whereas existing Pinyin-based methods leave between 20\% and 100\% of names unclassified. These findings demonstrate the importance of hanzi-level information for reliable gender inference of Chinese names. Code and data are publicly available.
cs.CL / 61 / 2610.23741
The Italian lawmaking process: a unified and structured dataset
Abstract
The Italian legislative process is documented in detailed parliamentary records hosted in institutional websites. However, much of this information is available only as unstructured text, limiting its systematic reuse. We applied a controlled, systematic LLM-based extraction procedure, validated against manually annotated records, to convert these texts into ordered and dated events including, e.g., bill introduction, committee examination, plenary examination, and approval. The resulting dataset reconstructs the procedural histories of all 5346 Italian laws enacted between 1985 and 2026, spanning parliamentary terms IX to XIX, for overall 101413 events. Each event records the activity, date, and chamber, with additional procedural details where available. The dataset supports research in the emerging fields of computational legal and legislative studies, potentially combined with process-oriented data science methods to examine procedural paths and legislative timing across parliamentary terms.
cs.CL / 62 / 2610.23807
Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal
Abstract
While deriving text embeddings directly from frozen LLMs has emerged as a promising direction, existing methods focus almost exclusively on English, leaving multilingual settings largely unexplored. In this work, we present the first systematic study of multilingual training-free text embeddings in LLMs. Our analysis reveals a language-conditioned embedding geometry in which semantically equivalent sentences remain misaligned across languages. We hypothesize that language-specific variation can be disentangled from semantic content to improve embedding quality. To this end, we propose Paired-Residual Subspace Removal (PRSR), a training-free inference-time intervention that steers LLM representations toward a more language-invariant semantic space. PRSR first constructs paired residuals from multilingual parallel sentences to capture language-specific deviations. It then estimates layer-wise low-rank subspaces from these residuals and suppresses the corresponding components in the last-token hidden states. We conduct more than 200 monolingual evaluations across 18 languages using MTEB and MMTEB, along with cross-lingual bitext mining between English and over 200 other languages. Empirical results demonstrate that PRSR consistently improves the performance of existing training-free baselines in both monolingual and cross-lingual settings across different LLMs.
cs.CL / 63 / 2610.23818
Acceptance Is Not Speed: Length Inflation in Training-Free Lossy Speculative Decoding
Abstract
Speculative decoding typically characterizes speedup by draft tokens accepted per verification round. Lossless verification preserves the target model's output distribution and thus expected completion length, making this metric a meaningful efficiency proxy. Relaxed (lossy) verification accepts tokens that lossless verification would reject, potentially altering that distribution. We identify a length inflation phenomenon: relaxed verification lengthens completions, primarily by extending reasoning traces, largely offsetting apparent gains in accepted tokens per round. We evaluate existing training-free relaxed verification rules on twelve target-drafter pairs against lossless verification on identical prompts, measuring completion length, verifier rounds, wall-clock time, and task accuracy. At each rule's most relaxed evaluated setting, relaxed verification on average accepts $1.34\times$ as many tokens per round but produces $1.54\times$ longer completions. For rules with higher acceptance lengths, tokens accepted only under relaxed verification tend to have low target-model probability, contributing to length inflation. For rules that reduce verifier rounds, accuracy typically drops substantially. Further analysis traces some failures to continued reasoning that exhausts the budget needed to deliver a final answer. Overall, accepted tokens per round alone is an unreliable measure of decoding speed under relaxed verification. Completion length and verifier rounds must also be reported to assess actual efficiency.
cs.CL / 64 / 2610.23831
Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation
Abstract
A reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision by considering both their intermediate content and their accessibility to the target learner. We introduce ReDiST, a student-conditioned trace-construction and distillation framework for evidence-rich reasoning. Starting from the student's own attempts, ReDiST revises answer-correct drafts against the supplied evidence and repairs unsuccessful drafts with additional guidance from the verified answer. A frozen student-matched reader screens and ranks original and edited candidates using answer recovery and verified-answer probability, given the question, evidence, and candidate trace. We distill the selected targets while balancing trace and final-answer supervision with an auxiliary answer loss. Construction and selection occur offline, and inference uses only the distilled student. Experiments on question-answering benchmarks show that ReDiST can improve both source-task and cross-benchmark performance over answer-only, original-trace, and the evaluated trace-selection and rewriting baselines. Across supervision recipes, performance rankings can differ between source and transfer benchmarks, highlighting the importance of evaluating trace teaching utility across distributions.
cs.CL / 65 / 2610.23839
DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
Abstract
Scientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. We introduce an operational definition of valuable and testable research hypotheses: evidence-grounded claims about consequential uncertainties, coupled with feasible investigations, explicit decision criteria, and alternative explanations. Our DiscoveryRSI framework develops hypothesis trees through multi-agent evolutionary search, compiles selected investigations into executable Harbor tasks, and recursively learns through three mechanisms: hypothesis credit optimization attributes scientific assessments to documented research contributions; harness replay identifies productive strategies and recurring failures to refine research workflows; and rediscovery applies accepted improvements to new investigations, generating further experience for subsequent learning. We also introduce DFM-Bench, which provides comprehensive assessment of hypothesis quality and cumulative research progress at the levels of hypotheses, revisions, development paths and research runs. Across four life-science domains, automated evaluation shows that hypothesis quality rises from 4.86 for the base model to 5.39 after two rounds of self-improvement, with the large gains in falsifiability and data resolvability. We further validate hypothesis quality through expert review, comparing our DiscoveryRSI-27B with other research agents. Preliminary human assessments show that DiscoveryRSI-27B scores higher than GPT-6 Luna on hypothesis importance and whether claims are plausible and evidence-bounded. DiscoveryRSI-27B achieved a 6.3% rate of hypothesis correspondence with research published after the model cutoff, largely outperform GPT-6 with 2.5% and Claude Opus 5 with 2.3%, while delivering greater human-rated improvement along research paths.
cs.CL / 66 / 2610.23542
Execution Before Memory: Diagnosing Memory Failures in Video World Models
Abstract
A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not carried out, the hidden state transition cannot appear, so the test is not really a test of memory. We introduce MIST, an evaluation that separates the change from the memory it should leave behind. The model first sets the scene while the change stays hidden, and only later is asked to show it, with leakage before that point kept under control. This separation locates the bottleneck: the model sets the scene but does not carry out the instructed change. The same gap appears in general memory evaluations. On that basis we propose M3-Agent (Make the Missing Move), which recursively revises the details of the input instruction so that the model successfully carries out the requested instruction. Relative to the baseline, it more than triples the MIST reveal score. On the benchmarks, WRBench's score T rises by 26.1% and MBench-T's score U by 52.7%. The method meets the problem from the execution side: it shows what the frozen model can do once the move is stated, and it sets a stricter baseline.
cs.CL / 67 / 2610.23588
Enhancing RAG-based Medical VQA with Gated Multimodal Representations
Abstract
Medical Vision-Language Models (Med-LVLMs) have shown strong performance in Medical Visual Question Answering (MVQA), yet they may still produce factually inconsistent responses. Retrieval-Augmented Generation (RAG) can mitigate this issue by incorporating external clinical evidence; however, retrieval is often guided by a single modality, potentially overlooking the joint semantics of medical images and clinical questions. We introduce a multimodal RAG framework that learns a joint query representation through a Gated Multimodal Unit, enabling evidence retrieval to leverage information from both modalities while keeping the underlying Med-LVLM frozen. We evaluate the proposed approach on benchmark datasets. Results show that the multimodal query representation consistently outperforms a reference Med-LVLM and remains competitive with state-of-the-art methods. Our findings demonstrate the potential of multimodal query representations for more effective evidence retrieval in Medical VQA.
cs.CL / 68 / 2610.23610
Not Every Mention Is a Claim: Rethinking Object Hallucination Detection in Vision-Language Models
Abstract
Object hallucination in vision-language models (VLMs) is usually measured by checking each object a caption names against the image. Most benchmarks and detectors alike treat every object mention as a claim that the object is present. This assumption fails when a caption names an object without asserting it, as in "a bench for people to sit on", which mentions people without claiming that any are in the image. We call such mentions non-assertive. Treating them as hallucinations inflates the error rates attributed to models that use such language. To separate them from real hallucinations, we introduce NAME-Bench (Non-Assertive Mention Evaluation Benchmark), 25,497 object mentions from captions that five VLMs generate for COCO images. Each mention is labelled concrete (asserted and present), hallucinated (asserted and absent), or non-assertive, and non-assertive mentions are further divided into four subtypes: depictive, purpose, negation and hedged. All the hallucination detectors that we evaluate flag non-assertive mentions as hallucinations, at rates between 14% and 80%. Adding non-assertive mentions to detector training as negatives largely corrects this on LLaVA-1.5. On the four newer VLMs, a comparable correction takes five to ten times more data, and the detectors then catch fewer real hallucinations. These findings show that current detectors conflate mentioning an object with asserting it, and that this conflation is harder to undo on the four newer VLMs than on LLaVA-1.5. Hallucination evaluation should therefore separate what a caption asserts from what it merely mentions.
cs.CL / 69 / 2610.22724
A misinformation playbook sharpens judgement without breeding cynicism
Abstract
Misinformation harms public health and democracy, and generative artificial intelligence now writes falsehoods as persuasive as people's. Defences range from fact-checking individual claims to training people to resist falsehoods. Yet many claims are too new to check, and interventions that teach doubt can reduce belief in truth too. We propose a defence that targets not the facts of each claim but the persuasion strategies that recur across them. Here we show that learning misinformation's persuasion playbook, mapped from 214,646 fact-checked claims in 51 languages, can sharpen human and machine judgement of claims they know little about, without making either more cynical about true claims. The playbook spans social influence, information manipulation and emotional appeals, combined around a core of fear, conspiracy and fabrication. Briefed with it, plus a reminder that persuasion is not proof of falsehood, a language model better distinguished true from machine-generated false claims about events after its training data. In a pre-registered experiment, the same briefing improved people's discernment of claims they reported as unfamiliar and left their belief in true claims unchanged. Averaged over all claims, the gain was not statistically significant. Because it targets what recurs, this defence need not wait for claim-specific fact-checks.
cs.CL / 70 / 2610.22797
Reasoning Reranking with Semantic IDs as the Sole Item Representation
Abstract
Reasoning rerankers write a rationale before ranking a candidate slate, and existing systems read item metadata to do so, even though generative recommenders increasingly represent items as compact Semantic IDs (SIDs). Ranking from SIDs alone has two appeals: SID prompts are far shorter than the matched item text, and although item codes are fixed from content, the backbone learns their token embeddings in part from user-item interaction sequences, so the tokens can carry behavioral signal that item text lacks. That training targets next-item generation, and whether it supports ranking a supplied slate from SIDs alone is untested. We build ReRankID, a controlled testbed in which models reason and rank from a history and slate shown only as SIDs, varying item rendering, history length, and slate size one at a time. On this testbed, such models often copy the input order instead of reranking, and readable metadata does not repair the failure. We compare supervised fine-tuning, rejection-sampled self-training, on-policy distillation, and verifier-based reinforcement learning as post-training methods and find that they elicit the missing behavior. The strongest post-trained model, the 8B model trained with a graded ranking reward, gives the best SID-only ranking and matches strong text-reading LLMs whose prompts are about 20x longer. It ranks above random with no history at all, where a caption-reading reference does not, and it keeps its gains at long histories, where text prompts pass 50k tokens and the caption-reading reference stops improving. In a paired rationale audit, this model argues trade-offs better than that reference but describes items less accurately. The common post-training methods we test thus elicit ranking from SIDs but not faithful item descriptions, which calls for post-training that improves the reasoning itself, not only the ranking it produces.
cs.CL / 71 / 2610.22832
Too Emotional to Be Fake? Intra-Utterance Emotion Shifts Expose Failure Modes in Speech Deepfake Detectors
Abstract
Expressive speech synthesis can now change emotion within a single utterance, yet its impact on deepfake detection remains unclear. We introduce EmoShift, a benchmark comparing emotion-stable (AA) and emotion-changing (AB) speech to examine these effects while accounting for static emotion identity and acoustic confounds. It comprises 7,240 bona fide utterances from four source datasets and 9,720 synthetic utterances generated using four expressive TTS systems. Across seven detectors, we find that emotion changes can make synthetic speech appear more genuine, increasing spoof miss rates by up to 25.3 percentage points on OMG-Emotion at fixed thresholds. Conversely, emotion-changing bona fide speech increases false-alarm rates by up to 24.4 percentage points on IEMOCAP. Further analyses reveal detector-specific sensitivities to emotion trajectories, not fully explained by static emotion identity or splicing artifacts. These findings underscore the importance of emotion dynamics in deepfake detection.
cs.CL / 72 / 2610.23678
Beyond the Benchmark: Conditioning, Domains and Failure Modes of Qwen3-TTS
Abstract
Qwen3-TTS reports state-of-the-art zero-shot voice cloning, but its technical report leaves open how the model should be conditioned, how it behaves outside read speech, and how it fails. We characterise its two cloning interfaces, in-context learning (ICL) and X-Vector, against Chatterbox, XTTS-v2 and F5-TTS on read, meeting, emotional and air-traffic-control (ATC) speech, using eighteen objective measures, an LLM judge, a spoofing countermeasure and a listening test. Neither interface dominates: ICL clones timbre more closely from correctly transcribed references, X-Vector clones timing, and ICL degrades sharply when a short reference transcript is wrong. Domain explains four times more variance than system choice. Qwen3-TTS leads on intelligibility and pitch register but trails on prosody, and the objective measures rank systems differently from listeners. An off-the-shelf countermeasure flags its clones no more often than genuine recordings, and on ATC radio it fails for every system. Code and audio are released.
cs.CL / 73 / 2610.23745
VocalGrad: Evaluating Acoustic Perception in Audio Language Models
Abstract
Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and non-linguistic audio understanding, such as emotion recognition. However, these models still struggle to reliably utilize acoustic information, particularly when it co-occurs with spoken linguistic content. Existing benchmarks do not clearly distinguish acoustic perception from higher-level reasoning, making it difficult to identify the source of these failures. To address this issue, we introduce VocalGrad, a benchmark designed to evaluate LALMs' ability to perceive low-level acoustic variations. VocalGrad formulates this problem as a simple binary classification of temporal change direction, enabling evaluation largely independent of higher-level reasoning or semantic understanding. Our benchmark results reveal a substantial gap in acoustic perception between human annotators and state-of-the-art LALMs: while human annotators achieve high accuracy, LALMs perform at near-chance levels. We further conduct linear probing to identify structural bottlenecks in these models. Our analysis reveals two distinct failure modes: for some categories, task-relevant information is preserved in internal representations but not utilized by the language model head; for others, the information becomes progressively less linearly decodable across model layers. Overall, our results reveal substantial limitations of LALMs in modeling acoustic variation under controlled settings and expose representation differences not observable from text outputs.
cs.CL / 74 / 2610.22745
Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns
Abstract
Language models that read financial news are usually trained on text from the same years used to test them, so a backtest can reward hindsight. We build news topics from NoLBERT, an encoder pretrained only on text published between 1976 and 1995, adapted into a sentence encoder by contrastive fine-tuning on passages of the same article, with no labels or returns. We compare it with LDA, MiniLM and FinBERT readers on 1.26 million articles from 2012 to 2024, convert each reader's topics into attention shocks and stock exposures, and forecast the returns of the 1,000 largest US stocks with eight models at three rebalancing frequencies. Models are fitted on 2013-2019, choices are made on 2020-2021, and 2022-2024 is held out. The experiment with the best validation score loses money at every frequency. Measuring attention shocks against the past 63 trading days instead of 5 raises the test rank IC of every reader. NoLBERT gives the most stable topics and forecasts about as well as LDA. An equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero before adjusting for the search but not against a price-only benchmark.
多智能体系统 (cs.MA)
4
cs.MA / 1 / 2610.22683
Dynamic Auction Mechanisms in a Closed-Credit Economy: Sequential Slot Reallocation in Air Traffic Management
Abstract
Indivisible resources allocated in advance may need to be reallocated when capacity reductions make existing assignments infeasible, yet monetary compensation is often impractical or undesirable. We develop a sequential resource reallocation scheme using reusable artificial credits, where agents hold pre-existing resource rights and have heterogeneous marginal valuations over time blocks. Capacity reductions produce fractional endowments, which our closed-credit scheme resolves into indivisible final allocations. We model agents' intertemporal preferences through a predicted-value-to-go function and characterize conditions ensuring that their optimal demand is non-increasing in price. We propose an ascending-auction mechanism that partitions agents into subgroups and uses the minimum subgroup market-clearing price as a common transaction price. We show that the mechanism preserves non-myopic individual rationality and weak nontrader neutrality. Moreover, we establish Pareto efficiency under optimal-demand clearing in the unified market and bound the expected gain from misreporting by the probability that an agent changes the common price. Simulations of airport slot reallocation demonstrate reductions in total modeled delay and cancellation costs relative to a ground-delay baseline.
cs.MA / 2 / 2610.23056
Thought Blocks: Non-Autoregressive Latent Communication for Multi-Agent Systems via Trajectory-to-State Compilation
Abstract
LLM-based multi-agent systems typically communicate through text, incurring repeated generation and re-encoding costs. Although latent communication removes explicit verbalization, constructing latent messages may still require autoregressive latent trajectories. We therefore distinguish message representation from message construction and ask whether agents can directly construct useful continuation states without executing the trajectories that normally produce them. We formulate this problem as trajectory-to-state compilation and introduce Thought Blocks, fixed-size blocks of latent key--value (KV) continuation memory constructed in one shot. Thought Blocks achieve a 2.5 times average latency speedup over autoregressive latent communication. Compared with text-based multi-agent communication, Thought Blocks reduce text tokens by 45.7% and autoregressive steps by 84.8%, achieving an 8.6 times average latency speedup. These results demonstrate that trajectory-to-state compilation enables efficient inter-agent latent communication without sequential latent rollouts at inference.
cs.MA / 3 / 2610.23092
Towards High Semantic Fidelity: Hyperdimensional Symbolic Messages in Multi-Agent Communication
Abstract
Semantic fidelity is essential for multi-agent communication under non-ideal channels, comprising transmission fidelity and conversion fidelity. Yet existing numerical and neural message formats cannot simultaneously achieve both. To address this, we propose a novel message format-\textbf{H}yperdimensional \textbf{S}ymbolic \textbf{M}essages (HSM) and its corresponding communication framework \textbf{Herm}. For transmission fidelity, Herm suppresses semantic entanglement via an entity-attribute-value structured encoding pipeline. For conversion fidelity, Herm first leverages a receiver-centric ego-binding mechanism to align semantics. Then, a similarity-aware dual-branch fusion module is proposed to prevent critical semantic dilution. To realize explicit message-semantic conversion, we devise an iterative decoding scheme via algebraic inverse operations. Experiments exhibit that Herm outperforms baselines across various noisy scenarios while maintaining superior structural readability.
cs.MA / 4 / 2610.22677
Whose Polarization Is It? Auditing Fairness in Network Polarization Measures
Abstract
A polarization score is supposed to describe a community, not its surroundings. Yet when a minority community is embedded in a larger network, its measured polarization can rise or fall with the size and structure of the majority around it. We ask whether polarization measures are therefore fair to minorities and develop an audit built around two criteria: measurement invariance (MI), which asks whether otherwise equivalent majorities and minorities receive equivalent scores, and structural exogeneity (SE), which asks whether a minority's score changes when its surrounding majority changes. Resolution sensitivity (RS) is treated as a prerequisite for meaningful fairness assessment. We audit seven polarization measure families spanning the three network-operationalizable dimensions of polarization, introducing Network-Weighted Distance (NWD) to fill an unserved extremeness-separation combination. Across controlled synthetic networks, mechanism-isolating diagnostic topologies, and real political and communication networks, no measure passes all applicable checks. Minorities are systematically scored as more polarized than equivalent majorities. We trace this dependence to two separate channels: majority structure can alter the opinions that opinion-reading measures consume, and, on a different path, opinion-orthogonal structural change can affect the arithmetic of structural measures. Reading opinions is thus double-edged, letting contamination in when structure moves opinions but shielding the score when structure changes and opinions do not. We close with practical guidance for reporting polarization scores without inheriting these biases.
软件工程 (cs.SE)
9
cs.SE / 1 / 2610.22544
SIEVE: A Parameterized GitHub Corpus Builder for Software Engineering Research
Abstract
Software engineering research increasingly relies on large, richly annotated corpora of real-world code. Still, researchers typically assemble these with ad-hoc scripts or static, pre-built datasets that offer no control over corpus composition. This leaves a gap for a reusable tool that goes below the repository level and builds annotated, function- and class-level code corpora under researcher-specified criteria. We present SIEVE (Software Ingestion & Extraction for Verifiable Evaluation), an open-source, parameterized GitHub corpus builder that, unlike repository-level mining tools, works at code granularity: it locates individual functions, methods, and classes in repositories and extracts them as records, giving direct control over corpus construction based on several criteria, including widely used programming languages (Python, Java, JavaScript, C++), repository quality, and creation date. To support software engineering research and LLM evaluation on real-world code, SIEVE extracts CodeSearchNet- and OpenClassGen-compatible records, each annotated with 22 structural metrics and AST-derived features, and optionally with a CodeBERT-based P(AI-generated) score (F1=0.9478, AUROC=0.9902), alongside the third-party dependencies declared in each repository's package manifest. These annotations enable researchers to conduct structural analyses, investigate code dependencies, assess AI-generated code, and construct customized benchmarks for evaluating LLMs, including under temporal constraints. A manual validation of 400 records in four languages and two granularities shows 100% accuracy in code metrics and import detection, 99.0% in source completeness, and 99.5% in metadata accuracy. SIEVE is publicly available with a live demo at https://mrahman2025-sieve.hf.space.
cs.SE / 2 / 2610.22664
Enjoying the Product but Not Joining the Builders: Investigating Why Students Don't Contribute to OSS
Abstract
Open-source software (OSS) participation provides students with authentic exposure to collaborative, real-world software development, yet active contributions remain limited. This short paper presents an exploratory study of computing students' perceptions, knowledge, and experiences with OSS at the University of Hawai'i at Manoa. We administered a survey and analyzed 122 responses using both quantitative and qualitative methods. Students reported moderate OSS familiarity and frequent OSS use in coursework and personal projects, yet only 12.3% actively contributed and engagement with OSS communities was limited. Key barriers included skill gaps (72.89% of non-contributors), time constraints (59.81%), and codebase complexity (66.67% of contributors), often alongside imposter syndrome (36.45%). Students indicated that more time, mentorship, beginner-friendly issues, better onboarding/documentation, and curricular incentives (e.g., course credit) would most increase participation, while most reported their courses do not require or incentivize OSS contributions (58.20%). Grounded in a single institutional context, this study provides a nuanced understanding of how curriculum structure, faculty encouragement, and competing academic demands shape OSS participation. These insights inform the design of targeted curricular and institutional interventions aimed at supporting student onboarding and participation in OSS communities.
cs.SE / 3 / 2610.22952
MergeShuffle: Reconstruct Shared Libraries for Improved Performance and Security
Abstract
Debloating shared libraries has been an active research topic for years. Tools for shared library debloating rely on either static or dynamic analysis techniques to discover which parts of the shared library is used. Once these parts are found, these tools take one of two routes, either removing parts of the library that are not used, requiring reconstruction of the library; or they simply overwriting the unused functions. However, recently a broad study~\cite{brown2024broad} on debloating tools has shown that some of these tools either increase the static binary size, or are generally less effective for complex workloads. In this paper, we argue that shared-library debloaters' performance can be improved by proper reconstruction of the shared-libraries once the used functions in the library are identifies. We propose \myapp, a novel tool for reconstructing shared libraries from a given set of file ranges to eliminate unused code based on user-specified requirements, without requiring access to source code. \myapp takes as input a shared library and a set of file ranges to retain, and performs a two-phase reconstruction: merging nearby ranges to reduce loading overhead, and shuffling them to minimize file size while preserving page-aligned memory mappings. The shuffling problem is formulated as a variant of the Traveling Salesman Problem and solved using a heuristic algorithm. \myapp utilizes program headers to remap retained code to its original memory addresses, preserving user-specified functionality. Our results show that \myapp achieves up to 60\% reduction in file size, 16\% reduction in memory usage, and 32\% reduction in runtime, while maintaining user-specified functionality and improving security.
cs.SE / 4 / 2610.23186
Code Skill Server: A Large-Scale Search Engine for Code Generation Skills
Abstract
Agent skills have proven to be an effective technique for improving agent ability on specific tasks, including software engineering, natural science, and mathematics. During such a process, using agent skills often involves manually discovering and organizing skills from heterogeneous sources, while existing skill retrieval methods mainly focus on selecting relevant skills from established skill libraries. However, with the rapidly increasing number and variety of reusable agent skills, manually searching for relevant skills across heterogeneous sources is becoming impractical at scale. In this paper, we introduce the Code Skill Server (CSS), a large-scale search engine for finding the most useful code generation skills for a target task. CSS consists of a skill server that incrementally collects and indexes publicly available code generation skills from multiple sources, with 993 skills included in the corpus used in this work. Extensive experiments across different retrieval methods and code generation benchmarks show that the best CSS configuration improves Pass@1 by up to 13.91% and consistently improves all evaluated models by at least 4.21% in the cross-model evaluation. These results demonstrate that CSS can effectively search for and serve relevant skills, improving code generation quality and enhancing models' code generation capabilities. Code and the CSS-993 corpus are available at {https://github.com/haoyuli-sdu/CodeSkillServer}.
cs.SE / 5 / 2610.23241
Towards Agentic Proof Harness Engineering
Abstract
Formal verification of real-world projects requires substantial manual effort, from specifying correctness properties to constructing the infrastructure needed to check them. We envision verification agents that assist developers throughout this process. Among the many tasks involved, one stands out as both foundational and amenable to automation: proof-harness engineering. A proof harness defines an isolated verification environment for a program unit, such as a function or module, by providing symbolic inputs, initialization code, and models of its dependencies. Today, proof-harness engineering remains largely a manual task performed by verification engineers. To explore the potential of coding agents for this task, we propose an agentic proof-harness engineer that automatically constructs verifier-ready harnesses from a software project and a designated verification target. We combine the agent with formal soundness and precision guards to validate generated abstractions and guide their refinement. We evaluate the feasibility of our approach in an exploratory case study on curl and libuv, using three historical assertion violations already embedded in the source code as verification properties. Our results show that the agent successfully constructs proof harnesses that enable CBMC to detect all three violations. We further demonstrate the potential for reusing these harnesses with other verification tools.
cs.SE / 6 / 2610.23506
MLOps-Bench: Benchmarking AI Agents on Production ML Engineering in Real Repositories
Abstract
AI agents are increasingly expected to modify and maintain machine-learning systems in production, yet existing benchmarks primarily evaluate either competition-style ML experimentation or generic repository issue resolution. We introduce MLOps-Bench, a repository-grounded benchmark for production ML engineering. Unlike prior ML-agent benchmarks, MLOps-Bench evaluates executable lifecycle changes inside real ML repositories, spanning data pipelines, feature engineering, model training, model serving, monitoring, CI/CD and governance, and end-to-end integration. Each task is grounded in repository evidence and includes a natural-language requirement, oracle patch, plausible buggy baseline, fail-to-pass tests, pass-to-pass tests, and provenance metadata. The benchmark contains 212 validated tasks from 31 public ML repositories, with 636 fail-to-pass tests and 856 pass-to-pass tests. We evaluate six agents and find that the best system solves 31.2% of tasks while passing 68.7% of tests, showing that current agents often make partial progress but still fail production-relevant behavioral requirements. MLOps-Bench shifts agent evaluation beyond model building and issue fixing toward repository-native MLOps work where correctness depends on lifecycle reasoning, behavioral robustness, and preservation of existing system behavior.
cs.SE / 7 / 2610.23632
Empirical Insights from AI-Powered Software Engineering in Practice
Abstract
Generative Artificial Intelligence (GenAI) is becoming indispensable in future software engineering. Its application spans a wide spectrum - from assisted development to autonomous software engineering, requiring organizations to address associated risks and challenges. This paper investigates these issues through three case studies, highlighting practical problems encountered during GenAI integration. Based on these findings, it provides targeted recommendations for managing risks and ensuring reliable outcomes. This paper provides guidance for effective and responsible GenAI adoption in software engineering based on empirical insights.
cs.SE / 8 / 2610.23782
Enforcing Code Reasoning in Agentic Scaffold for More Effective Issue Resolution
Abstract
Reasoning is central to ReAct-based programming agents, yet whether and how they reason about code remains unclear. We study 7,955 issue-resolution trajectories from six single- and multi-agent scaffolds across four programming languages on SWE-bench-Verified, SWE-bench-Pro, and SWE-bench-Lite. We identify 12 forms of code reasoning in 66.95% of trajectories. Code reasoning is associated with higher resolution rates, and its intensity increases with task complexity. However, it remains largely unguided: agents may focus on surface-level symptoms during localization, fail to refine near-correct patches, overlook issue requirements, and generate tests with limited behavioral diversity. To address these limitations, we propose DeskCheck, an agentic scaffold that systematically guides code reasoning throughout repository-level issue resolution. DeskCheck simulates execution paths to localize root causes, simulates candidate patches to refine near-misses, audits patches against issue requirements, and reasons about corner-case inputs and expected outputs before generating tests. On 731 SWE-bench-Pro instances, DeskCheck improves resolution rates over mini-SWE-Agent and Claude Code by 6.29% to 10.53%, at an additional cost of $0.02-$0.28 per instance (average: $0.14). Per-phase analyses show higher method-level localization recall, more successful repairs of imprecisely localized issues, and better detection of missed corner cases in near-correct patches. Exact McNemar tests show statistically significant improvements across all subject LLMs (p < 10^-10), while repeated experiments support robustness to agent nondeterminism. These findings motivate systematic code reasoning as a means of improving programming agents.
cs.SE / 9 / 2610.23316
SpinControl: Control Software for Prototyping and Benchmarking of Spin-Qubit Devices and Measurement Protocols
Abstract
Quantum computers require software to control the states of the qubits. As the actions on qubits grow in complexity with circuit execution, calibration, and benchmarking, the complexity of the control software grows in tandem. Spin-qubit quantum computers are at a stage where experimentalists need low-level access for tuning and calibration, and quantum algorithm researchers want high-level circuit-level access. In this article, we present the software design of SpinControl. SpinControl integrates three architectural metrics: 1) customizability, to allow prototyping devices and measurement protocols, is achieved by separating data and behaviour; 2) modularity and extensibility are achieved by separating functionalities into different components; and 3) robustness, to reliably execute quantum circuits on many qubits, is achieved by CI pipelines and input validation. SpinControl allows different types of measurements while customizing the instruments used, the pulse or circuit program executed, and the data processing. It allows building standardized experiments from measurement and analysis building blocks and executing these for calibration in a custom and conditional order. SpinControl implements these functionalities alongside robust configuration and data management, a connection to the cloud, and calibration scheduling. Pulse and circuit compilation, virtual gate matrices, as well as spin-qubit-specific GUIs integrated by SpinControl, make heavy use of software packages developed by Delft University of Technology: pulse lib, core-tools (both open source), and QConstruct
硬件架构 (cs.AR)
8
cs.AR / 1 / 2610.22716
CMM-MP: A Novel Memory Pooling Architecture using Purpose-built Memory Device with CXL and Ethernet
Abstract
CXL Memory Module - Memory Pooling (CMM-MP) introduces a novel memory pooling architecture to meet the unique requirements of emerging memory-intensive workloads. This industry-first, purpose-built memory-pooling device combines the benefits of Compute Express Link (CXL) with the scalability and cost advantages of Ethernet. Multiple CMM-MP devices together provide a flexible, composable, and disaggregated memory expansion solution. The CMM-MP enables 100s of TB capacity for GPU/CPU applications without modifications. The memory pool is exposed as a large-capacity logical aggregate memory, with all pool discovery, composition, and backend network datapath tasks fully offloaded to the device. CMM-MP supports both memory pooling and sharing usage modes. It provides an application latency of 2.7 us and up to 23 GB/s of memory throughput. Our evaluations demonstrate that CMM-MP significantly outperforms RDMA-based solutions in latency and throughput. We also evaluated real-world use cases of Valkey+YCSB and DiskANN using CMM-MP.
cs.AR / 2 / 2610.23222
qGAMMA: A Framework for Optimal DNN Mapping on Multi-Precision Accelerators
Abstract
This work describes qGAMMA, a flexible framework for exploring and optimizing the mapping of mixed-precision Deep Neural Network (DNN) models onto general-purpose spatial accelerators that support a range of fixed- and floating-point precisions. qGAMMA navigates the design space using a domain-specific genetic algorithm and an analytical performance model based on hardware synthesis results. The framework identifies the optimal mapping and accelerator configuration that minimizes target metrics. Extensive evaluations across DNN workloads show that qGAMMA offers valuable insights for designing mixed-precision applications, enabling better trade-offs between performance, energy efficiency, and resource utilization.
cs.AR / 3 / 2610.23406
GUMPIM: Unitary and Malleable Memory for Processing-in-Memory with Guaranteed PIM Pages
Abstract
DRAM-based Processing-in-Memory (PIM) addresses the "memory wall" by executing computations directly inside main memory. However, memory interleaving and virtual memory limit contiguous data size visible to PIM units, constraining PIM task granularity. To mitigate this, existing PIM systems drastically isolate PIM memory or disable memory interleaving. These design choices, however, decrease the CPU memory bandwidth and introduce extra data transfer, leading to an additional "system memory wall" that must be resolved to realize PIM's full potential. In this work, we propose GUMPIM, a PIM system that allows interleaved CPU pages and non-interleaved PIM pages to coexist in a Unitary and Malleable memory space with Guaranteed PIM page allocation. GUMPIM enables zero-copy during PIM task offloading and maintains CPU memory bandwidth while ensuring low PIM offloading overhead. First, we propose a dual-track memory management mechanism consisting of independent page allocation and address translation for CPU and PIM pages. Second, we design GUMPIM interface hardware on PIM-enabled DRAMs to provide a dynamic address mapping for the different data layouts of CPU and PIM pages. Third, we propose a PIM-assisted page migration mechanism that transparently migrates pages while preserving CPU access bandwidth, thereby enabling guaranteed and accelerated PIM page allocation. GUMPIM requires no changes to commodity DRAM standards; all hardware modifications are limited to the DRAM side, ensuring full compatibility with existing CPUs and enabling immediate deployment on current HBMx- and LPDDRx-based PIM platforms. Our results show only <0.1% performance degradation for CPU workloads on GUMPIM. For PIM workloads, GUMPIM reduces memory allocation and CPU-part computation times by 2.7x and 4.93x, respectively, yielding an end-to-end 2.3x speedup over a state-of-the-art baseline system.
cs.AR / 4 / 2610.23503
HBFSim: A Simulator for Exploring HBF in Memory Hierarchies for LLMs
Abstract
High Bandwidth Flash (HBF) has recently emerged as a promising technology that provides terabyte-scale accelerator-local capacity at bandwidths approaching HBM, with a potential cost-per-bit advantage over HBM. However, its NAND foundation brings long latency, out-of-place updates, and finite write endurance directly into the accelerator memory hierarchy, making its effects on performance and lifetime difficult to predict. To study this emerging design space, we present HBFSim, an LLM-aware event-driven simulator for heterogeneous HBM+HBF memory systems that preserves application-level memory semantics while modeling flash-internal behavior. HBFSim models HBM, HBF with host-managed flash translation, and conventional capacity tiers in one simulation driven by a calibrated serving frontend. Using HBFSim, we study decoding of Llama-3.1-70B and Qwen3-235B-A22B under three workloads that differ in their input and output lengths. HBF serves data that does not fit in HBM 11.0--19.0$\times$ faster than host DRAM; HBM is better used to hold the running sequences' KV cache than as a cache for HBF; and at a published peak load of Llama-3.1-70B, HBF lasts 2.0 to 6.9 years, in proportion to the stacks that share the KV writes. The artifact is available at https://github.com/FlashLabAE/HBFSim and will track public HBF device data as it becomes available.
cs.AR / 5 / 2610.23724
Agentic AI for Heterogeneous System Design Across Chip, Scale-Up, and Scale-Out Domains
Abstract
The design space of modern AI/HPC systems has expanded rapidly across compute architectures, including GPUs, dataflow processors, and RISC-V accelerators; memory tech- nologies, including HBM, GDDR, and SRAM; and interconnects such as NVLink, UALink, Infinity Fabric, InfiniBand, Slingshot, and Ultra Ethernet. No single architecture is optimal across LLM training, LLM serving, and AI-coupled HPC workloads, while the resulting combinatorial design space is too large for manual exploration. We present ArchAgent, a hierarchical multi- agent framework in which an LLM-based planner decomposes system design into chip, scale-up, and scale-out subproblems, each handled by specialist agents. A deterministic analytical evaluator closes the loop, ensuring that all performance estimates are derived from analytical models rather than generated by the LLM. On a mixed AI+HPC workload under a 1 MW power budget, ArchAgent discovers heterogeneous designs that achieve a 1.93x weighted geometric mean speedup over a homogeneous GPU baseline, while requiring only 18 evaluations compared with 100 evaluations for hill climbing, which achieves 1.65x speedup.
cs.AR / 6 / 2610.23767
TileDRAM: Compiler-Controlled Tile-to-DRAM Layouts for 3D DRAM GPGPUs
Abstract
Modern GPU compilers expose rich layout abstractions for mapping tensors to threads and on-chip resources, but compute-unit-local (CU-local) 3D DRAM architectures lack a corresponding abstraction for specifying how operand-tile data are arranged within the DRAM address space. Without such an abstraction, conventional placement under the existing DRAM address mapping can subject structured tile accesses to severe row-buffer conflicts. We present TileDRAM, a compiler-runtime toolchain that introduces tile-to-DRAM layouts as a compiler-visible abstraction for DRAM placement. TileDRAM extends the GPU Linear Layout abstraction from compute-side mapping targets to DRAM placement coordinates and selects layouts per tile access signature, using recurring tile-local access behavior as the specialization unit. Since row-buffer locality depends on ordered request history, TileDRAM evaluates candidate layouts through their row-buffer behavior under fixed DRAM address mappings. Accesses under the selected layouts are realized through ordinary address arithmetic, without new address-remapping hardware or ISA support. In full-trace replay on Accel-Sim/Ramulator2 across 108 mixture-of-experts (MoE) feed-forward network (FFN) cases, TileDRAM reduces the request-weighted row-conflict rate from 75.7% to 15.2% and achieves a 1.77x geomean speedup. Retargeting six traces to six 3D DRAM organizations yields a 2.20x geomean speedup under the same compiler abstraction. On an RTX 3090, address transformation adds 0.17% geomean overhead over identity.
cs.AR / 7 / 2610.22811
Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition
Abstract
Speed-limit sign recognition on an automotive camera system-on-chip (SoC) must fit a detector and a classifier (the reader) into the memory, arithmetic, and energy budgets of the device. In our two-stage recognizer, the detector, a YOLO-family network with ReLU and hard-sigmoid operators, is compressed by channel replay: an iterative pruning cascade identifies the channels to remove, the recorded deletions are replayed on the original task-trained network, the parent, and one continuous knowledge-distillation recovery follows. On a YOLOv5n, replay removes 60.3% of the parameters at the accuracy of the parent and, in validation mAP50-95 over five matched seeds, outperforms iterative pruning and training from scratch, and its per-step variant (per-step replay) is the most accurate pruning method. On a YOLOv8n, replay and per-step replay outperform every other method and the unpruned parent. The reader, a five-layer CNN with 18k parameters, and the post-processing of the YOLOv5n and YOLOv8n heads use integer arithmetic only. With 4-bit weights and 8-bit activations (W4A8), the reader matches FP32 accuracy in 12.1 KB of on-chip memory, and an executable streaming model reproduces the INT8 class predictions of its runtime on all 9,850 test crops. On a Jetson AGX Orin, INT8 is faster than FP16 only in the implicit formulation of TensorRT, and the replay detector runs fastest on the deep-learning accelerator (DLA). Trained on public data only, the combined system of 0.72 M parameters recognizes 95.2% of the signs of a held-out automotive camera at 90% precision and 94 frames per second.
cs.AR / 8 / 2610.23415
ConQUESTA: Topology-Aware Post-Training Structural Pruning of Quantum Neural Networks
Abstract
Deploying quantum neural networks (QNNs) on sparse hardware requires preserving task accuracy while reducing compiler-induced two-qubit gates, SWAPs, and depth. We present ConQUESTA (\underline{C}onstrained \underline{Qu}antum \underline{E}ntangler \underline{S}election with \underline{T}opology \underline{A}wareness), a post training pruning framework that combines task saliency with compiler cost, screens candidate entanglers, and searches complete pruning masks under an explicit accuracy constraint. Unlike one-shot ranking, each mask is transpiled as a whole across multiple coupling topologies. Across 4-, 6-, and 8-qubit QNNs and eight model seeds at 55% pruning, ConQUESTA achieves mean resource reductions of 61.6% on Breast Cancer, 72.0% on Digits01, and 82.2% on Iris, exceeding the strongest matched simple baseline by 3.5, 1.7, and 8.9 percentage points, respectively. Fixed-mask recovery raises feasibility from 16/24 to 22/24 runs. In a fixed-k=4 scaling study on a scaling probe of 4 to 12-qubits, the raw mask space grew from 495 to 12.1M while the evaluated masks increase only from 189 to 305. Resource-weight changes alter depth/routing priorities with at most a 0.44-point mean validation shift.
密码学与安全 (cs.CR)
24
cs.CR / 1 / 2610.22578
Model-as-a-Trojan: Self-Activating Covert DNS Exfiltration Baked Into the Weights of an LLM Agent
Abstract
Open-weight models are downloaded from public hubs and run inside agent harnesses that grant them tools, most powerfully, a shell, under the assumption that a sandbox contains whatever the model might do. This paper demonstrates that the assumption has a blind spot: the weights themselves can be the malware, and the sandbox's last open door, DNS, is enough. We fine-tune an open-weight LLM so that a single covert behavior lives in its parameters: on the first turn of every conversation, whenever the model infers from its context that a shell tool is available, it emits one self-contained command that reads a secret from the runtime environment, base32-encodes it, chunks it, and exfiltrates it as DNS queries toward an attacker-controlled authoritative name server. It then answers the user normally and never mentions, or repeats, the act. Because the channel is DNS, the attack completes from inside a container with no internet egress beyond name resolution - measured payload-level and model-level behind an egress firewall that drops every non-DNS packet, with the secret reconstructed from the gateway resolver's query log. Because the behavior lives in the weights, prompting is the only off-switch a victim has, and it is unreliable: the implant fires in 20/20 of adversarial suppression trials, deflects direct interrogation without disclosing the sequence, and sweeps a 45-cell sandbox-fault matrix across five tool-calling formats - plus a fresh post-training canary, persona-only and absent system prompts, and an egress-locked topology - with zero leaks and zero repeat fires. Every training trajectory is execution-verified in a real container, and every claimed exfiltration is proven resolver-side: we reconstruct the secret from the authoritative resolver's query log alone.
cs.CR / 2 / 2610.22580
Can SOLARBURST go silicon? A systematic analysis on software supply chain attacks for VLSI design flows
Abstract
An attacker targeting the software supply chain (SWSC) of VLSI tools can fully subvert the VLSI design pipeline through purely software level compromises. However, cur- rent literature on both SWSC and VLSI hardware security has not explored how software can be used to subvert the security of the VLSI design flow. In order to bridge this gap, this paper presents a systematic analysis of the vulnerable surface of VLSI from a SWSC standpoint. We carry this anal- ysis through three major activities. First, we did an empirical analysis of VLSI software suites, focusing on their software composition, security features, and tool interactions within the VLSI design flow. Second, we develop a VLSI-SWSC attack taxonomy model that links software supply chain at- tacks and their mechanisms to the VLSI attacks they enable, the design-flow stages they affect, and their resulting security outcomes, allowing us to identify 717 possible new attacks, including 60 high feasibility attacks supported by evidence collected in our empirical analysis. Third, we identify what possible security outcomes can be achieved by an attacker from these attacks and produce a comprehensive VLSI SWSC threat model. We evaluate the feasibility of the attacks identi- fied by our taxonomy across VLSI design stages and assess existing defenses to determine which attacks are mitigated and which remain unaddressed. In addition, we present case studies using a reference VLSI design flow consisting of open- source VLSI tools and highlight attacks identified through our taxonomy on the VLSI SWSC uniquely from a software perspective.
cs.CR / 3 / 2610.22626
StratFL: Countering Covert Model Poisoning Attacks with Strategic Federated Learning
Abstract
Advanced persistent threats (APTs) are a major concern in federated learning (FL) deployments in critical sectors such as healthcare, and finance. Traditional defense mechanisms fail to contain APTs due to their intermittent attacking behavior. This paper addresses a specific type of APT called covert model poisoning (CMP) where the attacker intermittently injects poisoned model updates in a covert manner. The interaction between the FL server and a CMP attacker is modeled as an imperfect repeated game where both players cannot observe each other's actions. A novel adversarial search algorithm called StratFL is designed alongside novel belief update algorithms at both the server and adversary. Experiments are designed using MNIST and CIFAR-10 datasets to evaluate and compare StratFL's performance against state-of-the-art methods. Empirical convergence of belief update algorithms is demonstrated across various security vignettes. At the same time, StratFL is found to exhibit an improved accuracy of 10% compared to state-of-the-art methods.
cs.CR / 4 / 2610.22659
Exploring The Empirical Privacy Landscape of Synthetic Data
Abstract
Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information. In this paper, we present a controlled comparison of data-level and training-level interventions for navigating the capability-privacy tradeoff on synthetic medical records with canaries to probe for leakage. Direct finetuning performs best on capability metrics but leads to the most privacy leakage, while DP-SGD reduces leakage at substantial cost to capabilities. Generating synthetic data from these records offers a more favorable balance, particularly when guided by a uniqueness-aware instruction to replace private details. Relative to naive paraphrasing, privacy-aware rewriting improves privacy protection across a variety of different rewriting models. We also find that more capable rewriting models generally reduce the risks of downstream privacy leakage, although held-out fit does not improve monotonically and increased test-time compute yields inconsistent gains. Overall, these results highlight privacy-aware rewriting as a promising direction for achieving a more favorable trade-off between capabilities and empirical privacy compared to other training-level methods, and our evaluation framework supports comparisons of future models and methods.
cs.CR / 5 / 2610.22697
What It Is, Not What It Says: Securing Agentic Browsing Interactions with Web Semantics
Abstract
AI agents increasingly browse the web on users' behalf, yet they consume webpages as screenshots or text, exposing their LLMs to unvetted third-party content. Browsers enforce semantics under which what an element can do depends on what it is and where it comes from, not on what it says. Agents flatten these semantics, so injected text can persuade the LLM to exercise the user's authority for content that has none. As LLMs remain easy to persuade, the agent's interface to the web must itself ensure the integrity of what the agent reads, interprets, and does. We propose Intact, which runs on an unmodified browser. Intact builds GenTree, a bespoke representation that captures provenance and web semantics. Intact passes the page via GenTree to the LLM one element at a time, so the LLM decides on each element without any other element's text. Its harness executes each step only within the capabilities of the step's element and asks the user before acting on third-party content. On 1,270 real webpages, GenTree attributes 2.3x as many elements to third parties as the browser, which misattributes 58.7% of them to the first party. On 1,091 of these webpages with injected attacks, only 12.8% of attacks get through Intact's planner, each at most as a consent request, against 73.1% for a conventional planner. Isolation disrupts functionality on only 3.7% of tasks, and a parallel schedule could remove most of its added latency.
cs.CR / 6 / 2610.22722
The LLM-Attacker: A Unified Red-Team Framework for Privacy Evaluation of Split and Distributed LLM Systems
Abstract
Split language models keep input text on a client but send intermediate representations to a server. Attacks can reconstruct text from these messages. Comparing privacy evaluations is difficult when attackers see different messages, know different model weights or use different datasets and recovery measures. We present LLM-Attacker, a framework that records these choices and lets researchers check, compare and extend an evaluation. Compatible attacks reuse recorded messages, and their predictions are saved for independent scoring. The scoring rule requires a linked undefended reference before a defence can pass. Reports also identify attacks that remain untested. In a controlled GPT-2 experiment, adding gradients returned during training increased token recovery from 94.20% to 97.38% and exact-document recovery from 13.71% to 37.77%. The 95% interval for the token gain lay above the two-percentage-point threshold fixed before testing. This experiment is shared with a companion study; here, it tests how the evaluation responds to a change in attacker access. Further studies measure how corpus choice and compression defences affect recovery, compare an adapted published attack with its reference implementation, and test the measurement code. The framework reports implementation checks separately from measured attack effectiveness. Each privacy claim should state the attacker, dataset and scoring rule, report its controls and identify untested attacks.
cs.CR / 7 / 2610.22879
RAMP: Toward Ubiquitous Quantum-Safe Key Distribution as a Service
Abstract
Quantum key distribution (QKD) offers a security foundation independent of computational hardness assumptions. However, its reliance on specialized hardware limits access by ordinary users. Turning QKD into a service poses a fundamental trust challenge, as the providers that supply QKD-derived material themselves could retain knowledge of the keying inputs of users. This makes end-to-end confidentiality solely depend on the trustworthiness of the service infrastructure. We present RAMP, a service-oriented framework that decouples QKD material generation from key consumption. Providers operate QKD infrastructure and supply raw-key materials, while users securely acquire, mask, and locally process them into final keys without operating QKD devices. RAMP employs reconciliation-preserving masking compatible with both DV- and CV-QKD. It mitigates the trust dependence by separating the QKD-service and user-side masking domains. We formalize complementary confidentiality guarantees: RAMP protects the final key even if colluding providers expose their complete raw-key views, provided the masking domain remains secure; conversely, exposing the complete masking state does not break final-key secrecy as long as the QKD source and acquisition paths remain protected. The results establish the dual-domain security boundary of RAMP. We implement RAMP using experimentally acquired QKD data and standardized quantum-safe cryptography. After acquiring service-provisioned keying material, commodity endpoints complete user-side QKD post-processing with an average key-distillation throughput of 2.1Kb/s on an M4 Pro CPU. To our knowledge, we demonstrate the first experimental realization of such service-provisioned, user-side QKD processing, presenting a viable path toward QKD as a service.
cs.CR / 8 / 2610.22899
Provable Authority: A Protocol for Verifiable Authority Resolution and Control Across Human, Organizational, and Agentic Transactions
Abstract
Payment systems authenticate users, devices, accounts, and cryptographic credentials, but a relying party, such as a counterparty, payment rail, marketplace, or autonomous agent runtime, generally cannot obtain compact, independently verifiable evidence that the party presenting one exact transaction holds current organizational authority to execute it under the applicable evidence, policy, budget, and approval state. Agentic commerce makes this gap operationally urgent: a software agent can produce a valid signature while acting outside its delegated purpose, after its sponsor's authority has changed, or on transaction details that differ from those a human approved. We describe the ChainIT Authority Protocol, a subject-neutral and rail-neutral pre-execution authority and evidence control plane. The protocol separates durable identity, a versioned organizational authority record (the Authority Resolution Pactvera), sourced evidence assertions, deterministic reason-coded policy predicates, and organization-configured dispositions; binds each decision and human approval to a canonical transaction digest; reserves capacity atomically so that limits hold under concurrency; and authorizes commit only through a single-use, sender-constrained Execution Authorization Credential consumed at the rail adapter, with append-only lifecycle evidence sealed after the terminal outcome. An Agent Subject Profile adds the workload registration, attested key hierarchy, and sovereignty boundary required when the acting subject is an AI agent, without introducing a separate authority concept. We specify the commit-time execution predicate and bounded-freshness status rules, give enforcement profiles for fiat (ACH, wire, card, instant payment) and stablecoin rails, and state residual risks and non-goals.
cs.CR / 9 / 2610.22913
Provable Compliance: A Protocol for Runtime Compliance Evidence and Transaction-Specific Decisioning
Abstract
Regulated institutions already perform onboarding, screening, monitoring, and review, but compliance evidence ages, and whether it supports a decision depends on the institution, product, subject, counterparty, transaction, jurisdiction, risk tier, and purpose. A relying party therefore often cannot obtain compact, attributable proof of which evidence, policy, authority, and transaction context were current when a decision was made, a gap that grows as authority is delegated to software agents and scheduled workflows. We describe ChainIT Provable Compliance, the compliance-side layer of the ChainIT Authority Protocol. Each sourced or institution-generated fact is represented as a separate Evidence Assertion VDT with provenance, freshness, confidence, permitted purpose, and privacy classification; a versioned Compliance Policy Profile selects the applicable rules; a Business Rules Engine returns deterministic, reason-coded predicate results; and the organization's configured Workflow Engine composes them into a signed, time-bounded Compliance Decision VDT for one reliance event. The decision contributes a single predicate, Compliance Decision Permits, to session-eligibility and transaction-policy decisions. It never independently authorizes payment: only a final Allow together with an exact transaction digest, atomic capacity reservation, and a single-use execution credential can commit value. The paper specifies point-of-reliance freshness, scoped propagation of evidence changes, decision states, evidence domains (AML, sanctions, PEP, background, identity, and licensing), the U.S. and FATF regulatory context as of August 2026, privacy and disclosure controls, and residual risks.
cs.CR / 10 / 2610.22988
Surprise Masking: Private Data Filtering with Model Likelihoods
Abstract
Data useful for language model (LM) training can contain private or unwanted information that is at risk of being learned or memorized. We propose Surprise Masking (SM), a simple form of perplexity filtering that uses a publicly available reference model to identify unwanted spans and suppress learning from them while preserving the rest of each example. On synthetic code benchmarks, we show SM improves the privacy-utility tradeoff over standard masking methods and PII filtering. In fine-tuning experiments, SM reduces approximate extraction of private values from revealed prefixes from 66.3% to 2.3% and the reproduction of stylistic coding quirks from 23.6% to 1.3%, both while preserving performance on the evaluated coding tasks. We also evaluate SM for sanitizing datasets by directly removing unwanted content. Here, SM outperforms DePA, a perplexity-based defense for detecting poisoned dead code, while specialized detectors perform better on broader credential and PII benchmarks. Across eleven settings, we find that SM's downstream success is predictable from the separation of unwanted and useful content in the distribution, and it is particularly useful with code data. Ultimately, SM is an inexpensive, adaptable method to preserve valuable examples while reducing the learning of unwanted information.
cs.CR / 11 / 2610.23057
Aspire: Practical Batch Queries on Dynamic Encrypted Database with Full Leakage Suppression
Abstract
Leakage-abuse attacks against encrypted database systems exploit semantic leakage and pose a severe threat to the confidentiality of plaintext data and queries. To address this problem, we propose a zero-leakage encrypted database that rules out leakage-abuse attacks, which fundamentally rely on semantic leakage. We first distinguish between two kinds of zero-leakage security. Ideal zero-leakage forbids the disclosure of any information beyond public parameters, whereas restricted zero-leakage additionally permits the leakage of metadata (such as database size and query timing). We show that restricted zero-leakage is sufficient to thwart leakage-abuse attacks and is achievable in practice. We then formally define adaptive restricted zero-leakage security and present Aspire, an encrypted database that satisfies this notion. Aspire integrates probabilistic batch code and minimal perfect hashing to support efficient batch keyword queries and dynamic updates while revealing only non-semantic metadata. Experimental results demonstrate the practicality and efficiency of Aspire, showing that Aspire can achieve an amortized query latency of less than 1.38 ms (7.53 s for a batch of 5461 keyword queries), with an amortized communication cost of approximately 2.95 KB (16.1 MB in total), on a $2^{22}$ key-value pairs database.
cs.CR / 12 / 2610.23077
When Approval Becomes Execution: Authorization-Grounded Workflow Attacks on Tool-Using Agents
Abstract
Tool-using language-model agents increasingly rely on local services for document access, approval management, and enterprise workflow execution. This paper studies an authorization-grounded workflow attack in the OpenClaw Agent Security Attack Competition, a controlled benchmark derived from SeClaw-style security task synthesis. Rather than using explicit policy-override instructions, the attack guides the agent through a locally coherent approval workflow: a protected document request is denied, the agent requests a normal approval token from the local approval service, and the returned token is supplied to the document service through authorized_by. Official submission traces show this pattern succeeding on three Enterprise OA document targets. We also conducted 23 controlled local validation experiments in the official OpenClaw execution environment: the original configuration reached the approval-token transition in 9/9 reruns and completed the exact target operation in 8/9 reruns, while single-component variants remained successful on the three document targets. The results provide concrete case evidence of an authorization-context gap: token possession can be treated as sufficient authorization even when requester identity, approver authority, resource scope, action type, destination or recipient, and business purpose are not jointly enforced. We present the attack pattern, controlled trace analysis, and defensive implications for context-bound tokens, service-side policy binding, agent-side authorization checks, and trace-aware benchmark design.
cs.CR / 13 / 2610.23078
Cryptographically Secure Agent-to-Agent Communication in Multi-Agent Systems
Abstract
Multi-agent systems (MAS) comprise AI agents that collaborate or compete to perform complex, distributed tasks. Their dynamic interactions create opportunities for threat actors to target and impersonate other agents, intercept messages, inject malicious prompts, and exfiltrate sensitive data while simultaneously complicating the detection of anomalous behavior through conventional monitoring. This paper surveys threats and cryptographic challenges in agent-to-agent communication and presents a zero-trust architecture integrating identity, transport security, message integrity, and monitoring. The architecture establishes agent identities via decentralized identifiers and task-scoped verifiable credentials backed by a permissioned ledger, enforcing least-privilege authorization. For transport security, it provides mutually authenticated, crypto-agile TLS with hybrid post-quantum key establishment, and evaluates KEMTLS with pre-distributed public keys to reduce handshake overhead. Signed CBOR envelopes preserve payload integrity across multi-hop exchanges, while zero-knowledge proofs verify declared task constraints without exposing underlying inputs. Context-aware monitoring of inter-agent messages, tool activity, and execution telemetry informs the setting of risk-based trust thresholds and supports the detection of behavioral anomalies. Together, these mechanisms provide an integrated framework for securing and auditing agent communication, with implications for other agentic protocols and prospective 6G semantic communication.
cs.CR / 14 / 2610.23109
Trojan Horse: Automated Repository Poisoning against Coding Agents
Abstract
Coding agents are increasingly capable with autonomy, which exposes them to repository-level poisoning. Existing red-teaming studies remain restricted to manual selection for malicious payload locations and fixed disguise strategies, which cannot represent realistic repository poisoning. We present TrojanForge, the first automated framework for repository poisoning. TrojanForge parses repository structure through abstract syntax trees (AST) and maps agent execution traces onto AST nodes. A security review agent ranks these candidates and selects the least conspicuous ones for payload mounting. Then, through adversarial play with the target agent, TrojanForge iteratively refines disguise strategies by inverting the causes of agent detection, and transfers accumulated success experience across repositories. We further construct TrojanBench, a benchmark of 384 poisoned repositories spanning 6 domains and 9 programming languages. Extensive experiments show that TrojanForge achieves a stable poisoning success rate of over 60% with backdoor injection exceeding 90%. Comprehensive evaluation on TrojanBench across 4 harnesses and 12 models reveals that even commercial agents fail to defend against repository-level poisoning.
cs.CR / 15 / 2610.23233
AgentLoupe: Kernel-Resident Semantic Causal Labels for AI Agent Observability
Abstract
AI agents translate user intents into system actions via tool invocations. Accurately attributing these actions to their originating LLM interactions is essential for security auditing, yet existing methods, such as AgentSight, fail for delayed execution and cross-process-lineage behaviors. We present AgentLoupe, an application-transparent eBPF-based observability tool. The core insight of AgentLoupe is to embed causal identity into kernel resources.AgentLoupe employs eBPF to capture TLS plaintext to interpret LLM interactions, and assigns each a kernel-resident semantic label.It maintains labels in eBPF maps, then propagates this label across processes and threads, associates it with kernel resources such as files and sockets. When a persistent resource carries context to an independent process outside the original lineage, this design preserves the delayed causal association.Experiments show that the average observation overhead of AgentLoupe is below 8%, and case studies confirm AgentLoupe traces cross-lineage persistent attacks back to their originating LLM interactions.
cs.CR / 16 / 2610.23303
MobileJailBench: Red Queen Jailbreak Benchmarking for Mobile Agents
Abstract
MLLM-based mobile agents are rapidly gaining autonomy in perception, planning, and closed-loop execution in dynamic mobile environments. However, this autonomy also broadens the attack surface and amplifies the risks of malicious misuse, including privacy leakage, device compromise, and large-scale phishing. Existing jailbreak evaluations leave three critical gaps for mobile agents: they focus on conversational attacks instead of mobile workflows, use static task sets with fixed difficulty, and rely narrowly on attack success rate rather than comprehensive security assessment. To close these gaps, we present MobileJailBench, the first evolving jailbreak benchmark for mobile agents, comprising three key components: (1) a benchmark comprising 520 tasks spanning three jailbreak paradigms and 14 risk categories, built on a high-fidelity mobile sandbox that we construct, enabling realistic workflows across over 20 widely used apps; (2) Dual-Dynamic Evolution, which combines Failure-Distilled Jailbreak Evolution with Weakness-Driven Benchmark Expansion; and (3) Stratified Adjudication Framework for assessing security awareness, execution behavior, and risk consequences. Extensive experiments show that current mobile agents are particularly vulnerable to implicit jailbreaks, as their safety alignment remains largely superficial and fails to generalize from conversational refusal to dynamic mobile interactions. Moreover, our dynamic evolution mechanism sharpens practical threat boundaries by continually uncovering new vulnerabilities, reaching a 100% ASR on Qwen3.5-122B. MobileJailBench provides a reproducible, continuously evolving framework for tracking the true security boundaries of mobile agents.
cs.CR / 17 / 2610.23305
On the (In)stability of Adaptive Fee Mechanisms in Multi-Chain Systems
Abstract
EIP-1559 has become the dominant transaction fee mechanism in permissionless blockchains, regulating scarce block space through an adaptive base fee. The design balances a trade-off between responsiveness to demand fluctuations and stability of long-term resource usage. On a single chain, this trade-off is well understood and empirically stable. In multi-chain systems, however, cross-chain flows introduce delayed economic interactions across otherwise independent fee markets. These interactions alter the control dynamics and create stability risks that existing models do not capture. Deployed systems have already polarized. Polkadot enforces conservative update rates, effectively disabling fee adaptivity and reverting the market to inefficient first-price auctions. Cosmos adopts aggressive update rates without formal guarantees. To fill this gap, we propose a control-theoretic framework that formalizes multi-chain fee markets as a nonlinear dynamical system with delayed feedback. We identify delay-amplified over-correction as the structural root of instability. Experiments show that cross-chain coupling erodes stability relative to monolithic blockchains, reducing the safe parameter space for base fee update rates by up to 84.8%. Moreover, instability propagates across chains, contaminating the otherwise stable peer chains. At Ethereum's update rate, which leaves a monolithic chain stable, the multi-chain system exhibits extreme load oscillation between full and near-empty blocks, with over half of all blocks at full capacity. Mainstream Cosmos chains already exceed their safe bounds under their default configurations, and the resulting load oscillation raises their block failure rate by an order of magnitude. Guided by this mechanism, we design a lightweight protocol-level mitigation that removes the over-correction and stabilizes dynamics in multi-chain environments.
cs.CR / 18 / 2610.23534
From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation
Abstract
Small Language Models (SLMs) are increasingly attractive for resource-constrained environments due to their low computational footprint. These characteristics also make them candidates for future autonomous malware, where connectivity, latency, privacy, or operational constraints may preclude cloud-based inference. Yet the ability of ultra-compact models to support malware during post-exploitation remains poorly understood. We systematically evaluate seven open-source SLMs (20M-1.5B parameters) on two post-exploitation code-generation tasks: Directory Enumeration and Backdoor Generation. We compare Python (high-resource) and Lua (low-resource) across prompting strategies, task formulations, iterative reasoning, grammar-constrained decoding, and execution-feedback loops. Results show that offensive code-generation capabilities do not increase linearly with model size. Successful execution depends on the interaction of model capacity, task formulation, and prompting strategy, with task formulation exerting a greater influence than prompt engineering alone. Algorithmic formulation combined with few-shot prompting enables Qwen2.5-Coder 1.5B to reach Pass@5 of 100% in both Python and Lua for Directory Enumeration, 100% in Python for Backdoor Generation, and 99.2% in Lua. Although the evaluated models remain too computationally demanding for widespread deployment in contemporary malware, rapid progress in compact models may reduce this barrier. These findings help anticipate AI-assisted malware and inform future countermeasures.
cs.CR / 19 / 2610.23656
ABF-Bench: A Verifiable Benchmark for Agentic Blockchain Forensics
Abstract
Blockchain forensics is adversarial and open-ended: money launderers split, swap, and bridge funds, cut their trail at mixers or custodial services, and change tactics from one incident to the next. Investigators must keep adapting their tracing in response. Yet the immutability of public ledgers preserves every recorded movement in public view, making much of the investigation independently verifiable. We introduce ABF-Bench, an interactive benchmark that makes evidence-backed fund tracing a verifiable agent task, in which every claimed movement must cite ledger records the agent retrieved. A case-to-task pipeline builds well-posed tasks across ledgers: conservative rules halt tracing wherever the records leave a continuation unresolved, and each task's reference graph is the answer those rules determine. Built from real laundering activity, the tasks combine multi-hop transfers, DeFi exchanges, cross-chain round trips, custodial and mixer deposits, and address poisoning. We also propose the ABF agent harness, built around a working memory, and evaluate it alongside general-purpose agents and fixed-order tracers. The harness reaches 75% edge recall, against 47% for Codex with the same model and 10% for Claude Code, which loses recall to overconfidence: it cites records it left unretrieved. Compared with fixed-order tracers, the harness acquires evidence faster than the best of them early in the budget and recognizes undisclosed custodial services that they walk past (89% vs. 0%), with 98% fidelity. However, agents tend to stop early and lose recall on larger tasks, which holds their final recall below the best tracer. Beyond exposing these gaps, ABF-Bench opens blockchain forensics to LLM agents as a verifiable domain, where the verdicts that score agents also offer potential training signals.
cs.CR / 20 / 2610.23668
TUPLE-Bench: Evaluating and Mitigating Privacy-Utility Trade-offs in Typed Decision Models
Abstract
Typed decision models often return more than a selected category. They also expose probabilities for each candidate and, sometimes, a separate confidence score. These fields support downstream decisions but may reveal information beyond a recipient's authorization. We introduce TUPLE-Bench (Typed decision models' Utility and Privacy Leakage Evaluation Benchmark), a benchmark of paired synthetic cases in which recipients may learn a coarse operational category but not its fine-grained cause. Across nine models, instructions to withhold the cause fail to protect it: 91.88%-99.93% of responses reveal the cause through direct selection or residual scores. Even when causes are excluded from the candidate set, full outputs enable 62.85%-75.35% recovery on held-out cases, compared with 51.04%-57.64% using only the selected category (50% chance). Full outputs also reduce mean downstream cost in all 27 model-task comparisons. However, confidence fields increase leakage without consistently reducing cost. We further introduce workflow-constrained privacy quantization (WCPQ), which reduces secret recovery while preserving the behavior and cost of specified downstream workflows. Our findings highlight the need to audit decision interfaces field by field and workflow by workflow.
cs.CR / 21 / 2610.23685
HardEval: An Execution-Verified Benchmark for AI-Driven Security Hardening and Compliance Remediation
Abstract
Implementing Security Technical Implementation Guide (STIG) controls remains a demanding operating-system administration task. Each natural-language security requirement must be translated into a precise and persistent configuration change. These changes span Secure Shell (SSH) configuration, authentication policy, audit rules, and kernel parameters. Existing benchmarks do not measure whether a model can perform this task on a live system. Code benchmarks evaluate isolated programs without persistent system state, while agent benchmarks largely focus on applications rather than operating-system configuration. In this paper, we introduce \textbf{HardEval}, an execution-verified benchmark for evaluating language-model-assisted security hardening. HardEval is built from Red Hat Enterprise Linux 8 STIG controls that fail on an unhardened baseline system. Expert-written reference remediations and Open Vulnerability and Assessment Language (OVAL) checks serve as hidden ground truth for grading; they are never shown to the model, and each task states only the required security outcome while withholding the implementation. The model produces a Bash script, which is executed on a live virtual machine and graded from fail to pass using the corresponding OpenSCAP/OVAL compliance check. Evaluation of fourteen frontier and open-weight configurations reveals large and uneven capability differences across security domains. A recurring failure mode is plausible security code that does not satisfy the exact compliance condition. For four open-weight configurations, multi-run Wilson confidence intervals preserve the observed ordering. HardEval is publicly available.
cs.CR / 22 / 2610.23851
Bad Diagnosis: Analyzing Consent and Data Collection Practices of Mobile Medical Apps
Abstract
Mobile health (mHealth) apps collect personally-identifiable information while handling uniquely sensitive medical data, ranging from mental health records to real-time glucose or heart rate monitoring -- yet they remain poorly understood at scale. We present the first comprehensive analysis of consent and data collection practices across 10,000+ mHealth apps on the Google Play Store, combining static analysis and privacy policy examination. We develop a rigorous operational definition to distinguish health apps from general wellness programs, and design a measurement framework to detect misuse of permissions and violations of consent. Our analyses reveal a troubling ecosystem: 69.15% of apps request at least one permission unrelated to their stated purpose, with over 52.9% apps requesting location data, and 50.5% and 34.9% apps requesting camera and microphone access, respectively -- despite no functional justification based on app descriptions and stated purposes. Most alarmingly, nearly 93% of apps fail to provide users with a consent form, directly violating regulatory requirements such as GDPR; and even among the remaining 7%, 1 in 14 demonstrate a gap between the obtained consent and collection of personally identifiable information. These findings demonstrate an urgent need for stronger enforcement of privacy regulations in mHealth apps. We plan to release our dataset and measurement tools for continued monitoring of this critical space.
cs.CR / 23 / 2610.23090
Concrete Functions in Private Simultaneous Quantum Messages: Quantum Advantages and Structural Limitations
Abstract
Private simultaneous quantum messages (PSQM) provide a minimal model of noninteractive secure multiparty quantum computation, yet their communication complexity for concrete functions remains poorly understood. We show that familiar total functions already exhibit both genuine quantum advantages and strong structural limitations. For three-party equality on $n$-bit inputs, which determines whether the players' $n$-bit strings are identical, we construct an entanglement-assisted protocol with asymptotic communication $(1.633\ldots)n$ qubits, whereas every perfectly correct protocol without prior entanglement requires at least $3n$ qubits. For three-bit majority, one shared GHZ state and one transmitted qubit per player suffice for perfect correctness and perfect privacy, reducing the communication from the optimal classical cost of six bits to three qubits. Three-bit AND behaves differently. No perfectly correct and perfectly private protocol exists when the players share an arbitrary three-qubit state, apply arbitrary local unitary encodings, and transmit one qubit each. However, increasing the local dimension overcomes this obstruction: one shared three-qutrit GHZ state and one transmitted qutrit per player yield a one-shot protocol with total communication $3\log_2 3$. Together with the six-bit classical lower bound, this gives a strict quantum communication advantage. More generally, for every $k\geq 3$, we construct a perfectly correct and perfectly private one-shot protocol for $k$-bit AND using one qudit of dimension at most $k$ per player, and hence at most $k\log_2 k$ qubits in total.
cs.CR / 24 / 2610.23798
Task-aware Kernel Stein certificates for Rényi differential privacy
Abstract
Rènyi differential privacy, based on Rènyi divergence, has become a notion of choice for privacy guarantees. However, outside the Gaussian setting, Rènyi divergence can be difficult to compute. In particular, Rènyi differential privacy guarantees that exploit task-specific information, as in Rènyi Pufferfish privacy, can be challenging. Here we use kernel Stein discrepancies based on escort distributions to provide certificates for Rènyi differential privacy. By designing the kernel to focus on the relevant privacy directions, we also derive task-aware privacy certificates. We provide exact identities for Gaussian and Laplace mechanisms and bounds for exponential mechanisms, with separate task-aware guarantees. A real data set example illustrates how the protected comparison set and the released subspace jointly determine the resulting certificates.