跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1084新增论文
+883其中 AI 相关
+130已有论文修订
+94新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-18 19:45各信源进度 →

统计口径与数据新鲜度

统计 2026-11-18 03:46 — 2026-11-19 03:46(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-18 04:28最近发现新增版本:2026-11-18 19:45库内论文最新更新时间:2026-11-18 08:00

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

26,870 已收录论文22,374 AI 相关8,156 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 49 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
版本更新 · v2arXiv:2611.19273

B-OPSD: On-Policy Self-Distillation with Self-Generated Problem Briefs

Zhi Xu, Yinqi Wu, Feng Zhang, Xiaotong Zhang, Hong Yu, Han Liu

原文摘要

On-Policy Self-Distillation (OPSD) has recently attracted interest because it provides dense teacher guidance without a separate, often larger teacher model. However, OPSD conditions its teacher on ground-truth solutions, which limits its use when such solutions are unavailable. We propose On-Policy Self-Distillation with Self-Generated Problem Briefs (B-OPSD). Using a prompt template, the initial model or a saved student checkpoint generates a problem brief from the question alone. The prompt asks for relevant principles, applicability conditions, and likely pitfalls, but not a final answer or a complete solution. A frozen teacher uses the brief as privileged context to guide the current student's on-policy rollouts through token-level distribution matching. B-OPSD uses no ground-truth solutions during training. Results on AIME 2024, AIME 2025, and HMMT 2025 show that both B-OPSD variants achieve higher average accuracy than OPSD at each of three Qwen3 model sizes. Code is available at https://github.com/ZevineXu/B-OPSD

cs.CL

版本更新 · v2arXiv:2605.25267

Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning

Minjae Kwon, Amir Moeini, Shangtong Zhang, Lu Feng

原文摘要

Safe in-context reinforcement learning (ICRL) adapts online from interaction history without test-time parameter updates while controlling episode cost under a safety budget. Under out-of-distribution (OOD) deployment shifts, pretraining-only safe ICRL can give poor reward-safety tradeoffs because the remaining budget affects behavior only through frozen policy conditioning, not an explicit action-level check against predicted future cost. We propose a latent Q-Barrier shield that learns a context representation, latent dynamics, and an ensemble cost critic before deployment. Without parameter updates, the shield infers context from history and filters or softly reweights candidate actions using the remaining budget and predicted future cost. We prove a conditional one-step barrier-margin bound over realized candidate sets, with explicit Bellman and latent-prediction error terms. Across five safe ICRL benchmarks, the shield improves deployment-time reward-safety tradeoffs over a strong safe-ICRL baseline: after a short context window, it achieves higher return in four of five benchmarks while matching or lowering average episode cost in all five. Code is available at https://github.com/kmj1122/q-barrier.

cs.LG · cs.AI

新论文 · v1arXiv:2611.21791

Regularized Score Centering

Hao Bai, Chenlu Ye, Xuanchang Zhang, Aviral Kumar, Tong Zhang

原文摘要

Asynchronous reinforcement learning increases training throughput, but data from stale policies and numerical mismatch between training and inference can destabilize learning. To stabilize off-policy learning, score centering (SC) removes the effect of reward baselines on the expected update. Our gradient diagnostics show less extreme gradients with SC than with importance sampling (IS), but SC can still fail when the learner moves far from the behavior policy that generated the data. To limit this policy mismatch, we introduce Regularized Score Centering (RSC), which adds a reverse Kullback-Leibler (KL) penalty to encourage the learner's next-token distribution to stay close to the behavior policy's next-token distribution. RSC works with or without truncated importance sampling (TIS), which caps token importance weights. Our theoretical analysis supports this approach, showing that KL regularization improves convergence guarantees. We evaluate dense and mixture-of-experts models on mathematical reasoning and agentic tasks, including online shopping, tool use, games, and fully asynchronous software engineering. RSC with TIS substantially improves sample efficiency across all agentic benchmarks, attaining higher training rewards with the same number of sampled responses. In final evaluations, it achieves top performance across all six agentic tasks, improving by 14.3% on AppWorld and reaching 9.9 times the strongest baseline's score on BabyAI.

cs.LG

新论文 · v1arXiv:2611.21524

面向安全上下文强化学习的算法蒸馏

Algorithm Distillation for Safe In-Context Reinforcement Learning

Minjae Kwon, Amir Moeini, Shangtong Zhang, Lu Feng

AI 速读 基于摘要

作者重新审视安全算法蒸馏(Safe AD),通过两阶段预算感知教师和测试时A-Lag调整,在五个安全上下文强化学习基准上改善成本与回报表现。

cs.LG · cs.AI

查看原文摘要

Safe in-context reinforcement learning (Safe ICRL) asks a policy with fixed parameters to improve on an unseen constrained task from its interaction history while aiming to keep its cost within a budget. Prior work studied Safe ICRL on five benchmarks in three families: a grid world with a hidden goal and obstacles (DarkRoom), continuous navigation where goals and hazards stay invisible until reached (PointGoal, CarGoal), and locomotion with a hidden velocity limit (HalfCheetah, Ant). Safe Algorithm Distillation (Safe AD), trained on the learning histories of a safe RL algorithm, has shown adaptation failures in prior evaluations. For example, in the locomotion tasks its return stays below zero, and it rarely reaches the goal in CarGoal. We revisit two parts of Safe AD. First, we distill a budget-aware teacher with two phases. It searches for the hidden task, exploring more when more budget remains. After its first reward or cost, an expert policy exploits the task within the remaining budget. We collect the teacher's histories under fixed or dynamic budgets. Dynamic budgets increase the occupancy entropy of the return and budget inputs in every environment, and at test time they improve return and lower cost in DarkRoom and HalfCheetah. Second, we propose A-Lag, a test-time adjustment that reweights the distilled policy's actions by their predicted cost, with an adaptive Lagrange multiplier that is updated from the costs observed. In our experiments, A-Lag had lower mean episode cost than Safe AD in all five environments, with lower mean final return in four.

新论文 · v1arXiv:2611.21382

VisualArtifactSpec 1.0: Evaluating Visual Artifacts under Task Specifications

Shuyi Ouyang, Xiaojun Tang, Zijian Zou, Xintong Zhang, Mingli Song, Guannan Zhang, Hongwei Xue

原文摘要

Visual artifact tasks test agents' ability to translate task specifications and source materials into outputs that meet practical needs. Evaluation across artifact families helps distinguish broadly applicable authoring capabilities from strengths confined to a particular medium. We introduce VisualArtifactSpec 1.0 (VAS), a benchmark of 211 creation and modification tasks spanning presentations, static graphics, animations and videos, and web interfaces. Each task specifies authoring objectives, source materials, constraints, and deliverables. Existing benchmarks provide limited support for systematically evaluating how authoring choices suit different audiences, purposes, and media across artifact families. VAS evaluates two aspects separately: Task Fulfillment verifies explicit requirements and applicable functionality, while Artifact Performance assesses Content, Organization, Visuals, and Delivery. Task-specific rating anchors instantiate shared criteria for the intended audience, purpose, and medium, with judgments grounded in source materials and artifact inspection records. We evaluate nine model-harness configurations on all 211 tasks. Overall scores range from 42.12 to 70.52, with Artifact Performance accounting for 73.4% of the gap between the highest- and lowest-scoring configurations.

cs.AI

新论文 · v1arXiv:2611.20882

Accelerate Bench: Evaluating Agents for Application-Level Program Optimization

Bowen Liu, Yuhan Li, Nuo Chen, Weile Luo, Miao Peng, Qintong Zhang, Qifan Zhang, Jia Li 等 10 位作者

原文摘要

Application optimization requires coding agents to discover useful changes and coordinate them across a complete program while preserving the required computation. We introduce Accelerate Bench, a benchmark of 32 CPU optimization tasks adapted from scientific applications and supercomputing competitions. Each task provides a working program and specifies the workload, correctness requirements, timing boundaries, and permitted edits, leaving agents to choose where and how to optimize. Public testing supports iterative development, and a separate verifier scores the final submission. On benchmark version 1.1, we evaluate 16 coding agents, each pairing a model with a harness. Their Score@1 ranges from 25.5% to 67.4%. Most of their programs are correct, and the size of the speedup separates them: the two leading agents are at least 2x faster on 22 or more of the 32 tasks, while the other fourteen reach 2x on 8 to 17 and leave 125 of their 412 correct outcomes below 1.25x. Controlled replays show why coordination across an application matters: identical source code runs between 0.29x and 2.12x depending only on build and runtime settings, and one code change is worth 1.01x or 6.4x depending on another. Accelerate Bench pairs executable workloads with agents' programs and trajectories to evaluate both their gains and the process behind them.

cs.AI

版本更新 · v2arXiv:2608.27909

Low-Altitude Fluid Antenna Network with Multi-Agent Reinforcement Learning

Tong Zhang, Yanfei Su, Shuai Wang, Wanli Ni, Chengzhong Xu, Huseyin Arslan

原文摘要

Low-altitude wireless networks (LAWNs) integrate terrestrial and aerial platforms to provide ubiquitous communication, sensing, and localization services for unmanned aerial vehicles (UAVs) and electric vertical takeoff and landing (eVTOL) aircraft. However, dynamic air-ground and air-air channels, abrupt blockages, and heterogeneous interference hinder the realization of this goal. Nevertheless, fluid antenna (FA), a cutting-edge multiple-input multiple-output (MIMO) technique, overcomes these challenges by reconfiguring antenna positions to unlock additional spatial degrees-of-freedom. In this paper, towards bringing low-altitude FA networks into reality, we study the fast and high-performance FA reconfiguration for low-altitude FA networks with multi-agent reinforcement learning (MARL). Specifically, we present an electromagnetic digital twin (EM-DT)-assisted MARL framework. To fill the sim-to-real gap, we introduce a two-stage transfer learning framework. Our case study shows that joint FA positions and beamforming optimization can enhance the system sum-rate by 118.5%, compared to the fixed position baseline. This gain comes from the dynamic millisecond timescale reconfiguration of FA arrays and the adaptive steering of beams toward aerial users with mobility.

cs.IT · cs.AI

版本更新 · v2arXiv:2503.18998

FACE: Few-shot Adapter with Cross-view Fusion for Cross-subject EEG Emotion Recognition

Haiqi Liu, C. L. Philip Chen, Tong Zhang

原文摘要

Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability. Existing works have primarily addressed these challenges through domain adaptation or generalization strategies. However, they typically require extensive target subject data or demonstrate limited generalization performance to unseen subjects. Recent few-shot learning paradigms attempt to address these limitations but often encounter catastrophic overfitting during subject-specific adaptation with limited samples. This article introduces the few-shot adapter with a cross-view fusion method called FACE for cross-subject EEG emotion recognition, which leverages dynamic cross-view fusion and effective subject-specific adaptation. Specifically, FACE incorporates a cross-view fusion module that dynamically integrates global brain connectivity with localized patterns via subject-specific fusion weights to provide complementary emotional information. Moreover, the few-shot adapter module is proposed to enable rapid adaptation for unseen subjects while reducing overfitting by enhancing adapter structures with meta-learning. Experimental results on three public EEG emotion recognition benchmarks demonstrate the effectiveness of FACE under few-shot cross-subject evaluation.

eess.IV · cs.AI · cs.CV

新论文 · v1arXiv:2611.15928

Runtime Safety and Emergency Stops on a Full-Size Humanoid: Deployment Lessons and a Case for Graduated Force Authority

Max Conway, Yutong Zhang, Allen Devaraj, William Xie, Alessandro Roncone, Nikolaus Correll

原文摘要

Full-size humanoid robots are being considered for hazardous industrial work such as disassembling electric-vehicle batteries. We describe the safety mechanisms of an open-source autonomy stack for the Unitree H1-2 humanoid and what deploying them taught us. All joint commands pass through a safety layer that runs separately from the control software. It enforces joint position, velocity, and torque limits, and it stops the robot when a state estimate becomes stale or when the hard-wired emergency stop (e-stop) is pressed or disconnected. We pushed the standing robot until it lost balance. With a model-predictive lower-body controller it never recovered, and the operator had to press the e-stop in every trial. Pressing it removes balance control, and the robot falls. The safety layer's own limits trigger the same stop, and limits chosen for manipulation were exceeded during normal walking. Near an operator or a battery pack, the fall can be more dangerous than the cause of the stop. Joint limits bound how much force a robot can apply, but not when applying it is appropriate. Drawing on research on responsibility and on child and adolescent development, we therefore make a case for granting robots force authority in stages, with a person or organization explicitly responsible for its use.

cs.RO

版本更新 · v2arXiv:2610.02274

Awomo-SimDataEngine: Agentic Simulation-ReadyWorld Generation

Awomo-PhysicalRSI Team, Danjiao Ma, Enhui Ma, Haohan Liu, Heng Jia, Hui Shan, Jianhua Xu, Jiahuan Zhang 等 24 位作者

原文摘要

Generating useful robot-training data requires more than visually plausiblescenes: objects must support interaction, placements must remain physicallyvalid, and tasks must admit repeatable execution. We present\textbf{Awomo-SimDataEngine}, an agentic system that connects asset and scenegeneration to robot demonstration synthesis. Shared asset services providerigid and articulated objects, including structure-grounded part and jointgeneration with ISArt. Scene generation supports two complementary routes:Unravel reconstructs editable scenes from images, while SimForge buildssingle-room and multi-room environments from text. A graph-native harnesscoordinates construction, validation, andbounded repair, routing failures to the responsible module while retainingunaffected scene state. PolicyForge binds validated worlds to tasks and robotembodiments to produce replayable demonstrations. Evaluations cover assetgeometry, scene quality, and downstream policy learning. On MuJoCo-basedLIBERO-Plus, co-training with Isaac Sim demonstrations improves the overallsuccess rate of a World-Action Model (WAM) from $77.17\%$ to $89.43\%$. Goal and spatialsuccess improve by $31.66$ and $6.25$ percentage points, respectively . These results support the utility of the generated data for cross-simulatorpolicy training, with more limited gains on long-horizon tasks.

cs.RO

新论文 · v1arXiv:2611.13524

The RSI Tax: When Recursive Self-Improvement Is Not Worth the Cost

Yitong Zhang, Ang Li, Yisen Wang

原文摘要

Recursive self-improvement (RSI) seeks to improve an agent using its own task experience. Its practical value depends on whether performance gains justify the full cost of self-improvement and task execution. We evaluate this trade-off on continual task streams, scoring each task before its feedback is used for subsequent updates. We define the \emph{RSI Tax} as the token cost of the self-improvement process and include it alongside task-execution costs in end-to-end evaluation. Across six methods and two language models, a bounded-note baseline matches or exceeds the observed accuracy of Meta-Harness, EvoSkill, and RRSI, with mean total-token savings of 65.6-71.0\%. We decompose the RSI Tax into verification, exploration, and writing costs, and track subsequent use of the associated state or artifact. For these three RSI methods, verification accounts for 74.8--93.6% of the aggregate RSI Tax. A substantial share of the RSI Tax is also associated with states or artifacts that are not used later in the evaluated streams. These results expose a practical limitation of costly RSI procedures: improvements in the deployed agent need not justify the expense of the improvement process. They motivate RSI objectives that jointly optimize the RSI Tax and expected task-execution cost while ensuring continued growth in agent capabilities across the task stream.

cs.LG

新论文 · v1arXiv:2611.11994

EmoHarness: Boosting Multimodal Emotion Understanding with Meta-Harness Evolving

Fenghai Li, Bingxuan Li, Jiawen Zhang, Yifan Shen, Jiateng Liu, Yutong Zhang, Yue Guo

原文摘要

Small multimodal models can misjudge emotion when words, voice, and visual cues conflict. We present EmoHarness, which runs such a model, the frozen student, inside a harness: the prompts, settings, and verification code around its calls. A human-written starting harness, the scaffold, numbers the student's visual observations, transcript, and prosody descriptions as citable evidence and has the student weigh candidate emotions against it. In Meta-Harness Evolving (MHE), a stronger LLM, the Builder, reads the student's errors and execution traces offline and edits the harness; static checks and validation gates decide which edits are kept, and the Builder never sees evaluation items. On 5,507 MME-Emotion items, EmoHarness raises task-macro Recognition by 7.6-13.4 points over a single video-only call for all four students and both Builders. Transcripts and prosody descriptions supply most of this boost (+5.38 to +10.71); the scaffold changes the audio-augmented call by -0.79 to +2.54, partly through extra multi-label predictions, and evolved edits add at most +0.95, although no evolved increment survives Holm correction. The harness's emotion context also transfers to IntentQA, with small gains for three of four students, none of which survives correction. EmoHarness thus boosts small multimodal students chiefly by supplying and organizing audio-derived evidence.

cs.CL · cs.AI · cs.CV

版本更新 · v3arXiv:2603.16861

MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation

Abhay Deshpande, Maya Guru, Rose Hendrix, Snehal Jauhri, Ainaz Eftekhar, Rohun Tripathi, Max Argus, Jordi Salvador 等 26 位作者

原文摘要

A prevailing view in robot learning is that simulation alone is not enough; effective sim-to-real transfer is widely believed to require at least some real-world data collection or task-specific fine-tuning to bridge the gap between simulated and physical environments. We challenge that assumption. With sufficiently large-scale and diverse simulated synthetic training data, we show that zero-shot transfer to the real world is not only possible, but effective for both static and mobile manipulation. We introduce MolmoBot-Engine, a fully open-source pipeline for procedural data generation across robots, tasks, and diverse simulated environments in MolmoSpaces. With it, we release MolmoBot-Data, a dataset of 1.7 million expert trajectories for articulated object manipulation and pick-and-place tasks. We train three policy classes: MolmoBot, a Molmo2-based multi-frame vision-language model with a flow-matching action head; MolmoBot-Pi0, which replicates the $π_0$ architecture to enable direct comparison; and MolmoBot-SPOC, a lightweight policy suitable for edge deployment and amenable to RL fine-tuning. We evaluate on two robotic platforms: the Franka FR3 for tabletop manipulation tasks and the Rainbow Robotics RB-Y1 mobile manipulator for door opening, drawer manipulation, cabinet interaction, and mobile pick-and-place. Without any real-world fine-tuning, our policies achieve zero-shot transfer to unseen objects and environments. On tabletop pick-and-place, MolmoBot achieves a success rate of 79.2% in real world evaluations across 4 settings, outperforming $π_{0.5}$ at 39.2%. Our results demonstrate that procedural environment generation combined with diverse articulated assets can produce robust manipulation policies that generalize broadly to the real world. Technical website: https://allenai.github.io/MolmoBot

cs.RO

新论文 · v1arXiv:2611.11450

How Much Training-Inference Mismatch Can On-Policy Distillation Tolerate?

Cong Yu, Ziting Wang, Songlin Jiang, Sitong Zhang, Chengliang Chai, Youhe Jiang, Eiko Yoneki, Tuo Shi

原文摘要

On-policy distillation (OPD) trains a student language model on its own responses, guided by a fixed teacher. Generating these responses is expensive, so systems use quantized inference and asynchronous rollouts to reduce the cost. Both, however, introduce training-inference mismatch between the rollout engine and the model being trained. Such mismatch can destabilize reinforcement learning (RL), but its effects on OPD remain unclear. This paper investigates how much mismatch OPD tolerates, where its effects arise, and how training systems can use this tolerance. Measuring all mismatch sources on one scale, we examine quantization across three student-teacher pairs, with additional tests of weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful loss of training quality but degrades at higher levels, revealing an empirical mismatch budget. Both the quality cost of mismatch and the mismatch introduced by quantization depend on the model, with some layers particularly sensitive to reduced precision. In controlled comparisons, OPD remains stable under rollout configurations that destabilize RL methods. Under weight noise, mismatch also shifts the prefixes reached during generation, introducing errors that per-token corrections cannot address. To put these findings into practice, we introduce a lightweight admission check that screens rollout configurations against a calibrated budget without requiring a full training run for each candidate. Model-specific FP8 recipes preserve training quality while providing capacity for at least $1.93\times$ as many KV-cache tokens as BF16. Combining FP8 with asynchronous rollouts reaches the same training quality $1.62\times$ sooner than synchronous BF16 training. Code is available at https://github.com/benyucong/train-inference-mismatch-opd

cs.LG · cs.AI

版本更新 · v3arXiv:2602.06643

HuMI: Humanoid Whole-Body Manipulation from Humanoid-Free Demonstrations

Ruiqian Nai, Boyuan Zheng, Junming Zhao, Haodong Zhu, Sicong Dai, Zunhao Chen, Yihang Hu, Yingdong Hu 等 11 位作者

原文摘要

Current approaches for humanoid whole-body manipulation, primarily relying on teleoperation or visual sim-to-real reinforcement learning, are hindered by hardware logistics and complex reward engineering. Consequently, demonstrated autonomous skills remain limited and are typically restricted to controlled environments. In this paper, we present the Humanoid Manipulation Interface (HuMI), a portable and efficient framework for learning diverse whole-body manipulation tasks across various environments. HuMI enables humanoid-free data collection by capturing rich whole-body motion using portable hardware. This data drives a hierarchical learning pipeline that translates human motions into dexterous and feasible humanoid skills. Extensive experiments across five whole-body tasks--including kneeling, squatting, tossing, walking, and bimanual manipulation--demonstrate that HuMI achieves a 3x increase in data collection efficiency compared to teleoperation and attains a 70% success rate in unseen environments.

cs.RO · cs.AI · cs.LG

新论文 · v1arXiv:2611.09036

ToolDex: An Ego-Exo Human Demonstration Dataset for Creative Physical Tool Use

Mohamed Elmoghany, Kehan Li, Ruoqu Chen, Tarun Chiruvolu, Shutong Zhang, Ryan Wang, Stephen Tian, Yaru Niu 等 14 位作者

原文摘要

Creative tool use is a fundamental human capability and an important step toward general-purpose embodied intelligence. It requires agents to reason about object affordances, physical constraints, and manipulation strategies, especially when standard tools are unavailable. However, progress in this area is limited by the lack of real-world datasets that capture human tool-use behavior with sufficient scale, fidelity, and diversity. Existing benchmarks are often simulation-based, focus on simplified interactions, or lack detailed observations and annotations. We introduce ToolDex, a large-scale egocentric-exocentric (ego-exo) human demonstration dataset for creative physical tool use. ToolDex captures scenarios in which humans select, adapt, and repurpose available objects as tools. Using synchronized egocentric cameras and eight exocentric cameras, the dataset provides high-fidelity multi-view observations and rich annotations for hands, tools, target objects, and their deformable interactions. Beyond data collection, we establish a benchmark suite characterizing three core capabilities: physical tool-use reasoning, sparse-view 3D and 4D reconstruction, and robot policy learning. ToolDex provides a foundation for studying human tool-use behavior and developing embodied agents that can adaptively use tools in complex real-world environments. Dataset visualizations and benchmarking results are available at https://tooluse-dataset.github.io/.

cs.RO · cs.AI

版本更新 · v8arXiv:2509.24102

Pragmatic Inference for Improving Social Reasoning Generalization in Large Language Models

Guangliang Liu, Xi Chen, Bocheng Chen, Han Zi, Xitong Zhang, Yuda Chen, Kristen Johnson

原文摘要

Achieving social reasoning, including moral reasoning, cultural awareness, and politeness, is critical for the socially reliable deployment of large language models (LLMs). A key bottleneck is limited generalization: social situations that convey similar social implicatures can be expressed through highly diverse linguistic forms, creating an enormous space of possible situations and posing a substantial computational challenge even for LLMs. From a linguistic perspective, this challenge arises because social reasoning operates at the pragmatic level, requiring a pragmatic inference process to infer social implicatures from a given situation and its underlying social context. Although moral values such as moral foundations are generally considered as the representations of the social context, how to construct the corresponding pragmatic inference steps remains underexplored. Taking moral reasoning as an instance of social reasoning, this paper develops a pragmatic inference approach that enables LLMs to infer moral implicatures by (1) specifying the inferential steps through which LLMs derive moral implicatures from language, and (2) explicitly modeling the metapragmatic links among these inferential steps to bridge the pragmatic gaps between them and make such inferences learnable for LLMs. Experimental results demonstrate that our approach substantially improves LLMs' generalization in moral reasoning, highlighting the potential of pragmatic inference for future social reasoning research.

cs.CL

版本更新 · v7arXiv:2601.03079

Learning to Diagnose and Correct Moral Errors in Large Language Models

Bocheng Chen, Xi Chen, Han Zi, Haitao Mao, Zimo Qi, Xitong Zhang, Yuda Chen, Kristen Johnson 等 9 位作者

原文摘要

Enabling large language models (LLMs) to diagnose and correct moral errors is crucial not only for ensuring their reliable deployment but also for charting a pathway toward continuous improvement. Existing research has primarily followed two directions: fine-tuning LLMs on corpora of morally appropriate texts and developing carefully designed prompting strategies to promote moral behavior, including moral self-correction and self-improvement. However, both lines of work often remain brittle, relying heavily on shallow heuristics rather than enabling LLMs to reason about why and how a given moral discourse is morally (in)appropriate. In this paper, we address this challenge by developing pragmatic inference-driven approach that help LLMs learn to diagnose and correct moral errors. Pragmatic inference is the reasoning process through which implied meanings are derived, a central concept in linguistics. To improve the reasoning efficiency, our approach adapt the inference procedure according to the inferential load required by different moral discourses. Empirical results demonstrate that our method not only improves LLMs' ability to diagnose and correct moral errors, but also generalizes effectively across tasks without relying on shallow heuristics.

cs.CL

版本更新 · v3arXiv:2605.07104

Almost Sure Convergence Rates of Stochastic Approximation and Reinforcement Learning via a Poisson-Moreau Drift

Xinyu Liu, Zixuan Xie, Shangtong Zhang

原文摘要

Establishing almost sure convergence rates for stochastic approximation and reinforcement learning under Markovian noise is a fundamental theoretical challenge. We make progress towards this challenge for a class of stochastic approximation algorithms whose expected updates are contractive, a setting that arises in many reinforcement learning algorithms such as $Q$-learning and linear temporal difference learning. Specifically, for a power-law learning rate $O(n^{-η})$ with $η\in (1/2, 1)$, we obtain an almost sure convergence rate arbitrarily close to $o(n^{1 - 2η})$. For the harmonic learning rate $α_n=α/(n+1)$, the almost sure squared-error rate is $o(n^{-ζ})$ for every $ζ<\min\{1,2(1-κ)α\}$, where $κ$ is the contraction factor. Thus exponents arbitrarily close to one are available when $2(1-κ)α\geq1$. This approaches the $n^{-1}$ polynomial exponent of classical law-of-the-iterated-logarithm benchmarks. Key to our analysis is a novel Lyapunov drift construction that applies a Poisson-equation based correction for Markovian noise to the well-established Moreau-envelope smoothing for the contractive mapping.

cs.LG · math.OC · stat.ML

新论文 · v1arXiv:2611.06500

Tools Need Not Speak in Text: Latent Context Compression for Agents

Yuetong Liu, Zhehao Zhou, Yitong Zhang, Haorui Chen, Jia Li

原文摘要

Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent's attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios. In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent's token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents. Our code is available at https://github.com/THU-Agent/Coca-cola.

cs.AI

新论文 · v1arXiv:2611.06449

AmbiHTMLBench: A Benchmark for HTML-based Visually Rich Document parsing

Boyu Niu, Yingjie Xiong, Shurui Li, Linhan Pang, Yuntong Zhang, Han Wang, Hao Wu, Kai Zuo 等 11 位作者

原文摘要

Traditional document parsing systems use Markdown as their output representation, which is effective for linear text but does not fully preserve fine-grained spatial structure, table geometry, visual hierarchy, or rendering semantics. We show that HTML is advantageous for heterogeneous and layout-sensitive documents whose representations must remain editable, inspectable, and reusable, and introduce AmbiHTMLBench to systematically study this setting. In this benchmark, given a real-world document, a model must recover executable HTML that preserves its content, organization, and visual appearance. The benchmark contains 1,115 diverse pages with structured annotations and evaluates predictions in two different ways: The objective protocol aligns ground-truth and predicted rendered blocks and measures table fidelity, page-level text error, and a two-dimensional layout consistency; a complementary reference-anchored protocol measures perceived fidelity by aligning LLM judges with human preferences. Across ten primary multimodal systems, the best objective score is 0.8092, while the subjective winner is different, demonstrating that automatic decomposition and human-perceived fidelity capture complementary properties. We further fine-tune a small-parameter expert model and find that task-specific adaptation yields good performance on document-to-HTML reconstruction task. These results position AmbiHTMLBench as a diagnostic testbed for building smarter and more logical document parsing systems.

cs.AI

版本更新 · v3arXiv:2606.11817

Grammar-Constrained Decoding Can Jailbreak LLMs into Generating Malicious Code

Yitong Zhang, Shiteng Lu, Jia Li

原文摘要

Large Language Models (LLMs) are increasingly used for code generation, raising concerns that they may be misused to produce malicious code. Meanwhile, Grammar-Constrained Decoding (GCD) has been widely adopted to improve the reliability of LLM-generated code by enforcing syntactic validity. In this paper, we reveal a counterintuitive risk: this reliability-oriented technique can itself become an attack surface. We uncover a new jailbreak attack, termed CodeSpear, that exploits GCD to induce LLMs into generating malicious code. Our experiments show that simply applying a benign code grammar constraint can effectively jailbreak LLMs. To address this vulnerability, we propose CodeShield, a safety alignment approach that robustly preserves safe behavior even under attacker-controlled grammar constraints. CodeShield aligns the model in the code modality by teaching it to generate honeypot code under GCD. Such code is semantically harmless, so it does not implement the malicious request, and structurally diverse, so it is difficult to suppress through grammar tightening. At the same time, CodeShield still preserves natural-language refusals when natural language is available. Experiments on 10 popular LLMs across 4 benchmarks show that CodeSpear outperforms representative jailbreak baselines and increases the attack success rate by more than 30 percentage points on average. CodeShield also restores safety under CodeSpear while preserving benign utility. Our findings reveal a fundamental risk of GCD and call for greater attention to its potential security implications.

cs.CR · cs.AI · cs.CL · cs.SE

版本更新 · v6arXiv:2603.15159

To See is Not to Master: Teaching LLMs to Use Private Libraries for Code Generation

Yitong Zhang, Chengze Li, Ruize Chen, Guowei Yang, Xiaoran Jia, Yijie Ren, Jia Li

原文摘要

Large Language Models (LLMs) have shown strong potential for code generation, yet they remain limited in private-library-oriented code generation, where the goal is to generate code using APIs from private libraries. Existing approaches mainly rely on retrieving private-library API documentation and injecting relevant knowledge into the context at inference time. However, our study shows that this is insufficient: even given accurate required knowledge, LLMs still struggle to invoke private-library APIs effectively. To address this limitation, we propose PriCoder, an approach that teaches LLMs to invoke private-library APIs through automatically synthesized data. Specifically, PriCoder models private-library data synthesis as the construction of a graph, and alternates between two graph operators: (1) Progressive Graph Evolution, which improves data diversity by progressively synthesizing more diverse training samples from basic ones, and (2) Multidimensional Graph Pruning, which improves data quality through a rigorous filtering pipeline. To support rigorous evaluation, we construct two new benchmarks based on recently released libraries that are unfamiliar to the tested models. Experiments on three mainstream LLMs show that PriCoder substantially improves private-library-oriented code generation, yielding gains of over 20% in pass@1 in many settings, while causing negligible impact on general code generation capability.

cs.SE · cs.AI · cs.CL

版本更新 · v3arXiv:2605.07147

MathlibPR: Pull Request Merge-Outcome Benchmark for Formal Mathematical Libraries

Zixuan Xie, Xinyu Liu, Shangtong Zhang

原文摘要

The ecosystem of Lean and Mathlib has become the de facto standard for large language model (LLM) assisted formal reasoning with remarkable successes in recent years. Those successes, however, only consume Mathlib as an essential dependency but do not directly contribute to it. In the meantime, the growth of Mathlib has recently been bottlenecked by the review process, which requires human reviewers to judge whether proposed pull requests (PRs) follow the Mathlib's conventions and are worth integrating as part of a shared mathematical infrastructure. This leads to our central question: can LLMs help review Mathlib PRs? To this end, we introduce MathlibPR, a benchmark built from real Mathlib4 PR histories. We further propose a staged evaluation protocol and use it to evaluate both LLM models (e.g., DeepSeek, Qwen, Goedel, and Kimina) and LLM agents (e.g., Codex and Claude Code). Surprisingly, both LLM models and LLM agents struggle to distinguish merge-ready PRs from build-passing PRs that were revised or never merged. By turning Mathlib PR histories into a supervised signal, MathlibPR provides a step toward reviewer assistants and reward models that could help evaluate PRs and steer LLMs toward producing merge-ready Mathlib contributions.

cs.LO · cs.AI · cs.LG

版本更新 · v2arXiv:2605.06874

On the Divergence of Differential Temporal Difference Learning without Local Clocks

David Antrobius, Shangtong Zhang

原文摘要

Learning rate is a critical component of reinforcement learning (RL). This work uses global and local clocks to distinguish two types of learning rates. The former is of the standard form $α_t$ that depends only on the time step $t$ (i.e., a global clock). The latter is of the form $α_{ν(S_t, t)}$, where $ν(s, t)$ counts the number of visits to state $s$ until time $t$ (i.e., a local clock). In discounted RL, an RL algorithm that is convergent with a local clock is always also convergent with a global clock, and vice versa. We are not aware of any counterexample. The key contribution of this work is to show that this nice correspondence breaks down in average-reward RL. Specifically, we construct a counterexample showing that although differential temporal difference (TD) learning is convergent with a local clock, it can diverge with a global clock. We empirically observe this divergence in our experiments and also prove that the associated ODE of differential TD with a global clock is unstable. This counterexample closes the open problem in Wan et al. [2021], Blaser et al. [2026].

cs.LG

版本更新 · v2arXiv:2603.26841

FatigueFormer: A Physiology-Informed Model for Upper-Limb Muscle Fatigue Assessment

Tong Zhang, Hong Guo, Shuangzhou Yan, Dongkai Weng, Jian Wang, Hongxin Zhang

原文摘要

Surface electromyographic (sEMG) signals provide information about muscle activity and function during movement. Fatigue alters multiple properties of these signals, with patterns that vary across individuals and contraction intensities. However, physiological insights into fatigue remain underused in sEMG modeling, particularly for cross-subject recognition. We introduce FatigueFormer to recognize fatigue states defined by ratings of perceived exertion (RPE). The model incorporates physiologically motivated trend grouping into sEMG encoding by organizing descriptors into increasing- and decreasing-trend streams and combining inter-window context with within-window feature interactions. To assess fatigue-state discriminability across contraction intensities, we further present an sEMG fatigue dataset collected from 30 participants at four levels of maximum voluntary contraction (MVC) and benchmark representative machine learning and deep learning methods. We evaluate performance offline under leave-one-subject-out (LOSO) and within-subject protocols and use the SWIFTIES dataset for external validation. Results support FatigueFormer's effectiveness across the examined contraction intensities and both datasets, while indicating that fatigue-related cues vary across MVC levels. An exploratory adaptation analysis shows higher FatigueFormer scores with more labeled target-user data on complete target recordings, which include calibration samples.

cs.LG · cs.AI

版本更新 · v4arXiv:2610.11773

Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models

Youwei Feng, Yitong Zhang, Yuetong Liu, Jia Li

原文摘要

Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.

cs.AI

新论文 · v1arXiv:2611.04158

On the Feature Evolution of Temporal Difference Learning

Shangzhe Li, Dechen Zhang, Difan Zou, Weitong Zhang

原文摘要

Temporal-difference (TD) learning trains neural representations using bootstrapped targets generated by evolving value estimates, yet how predictive features emerge and propagate under incomplete neural updates remains poorly understood. We investigate this problem using finite-width two-layer ReLU critics on a constructed family of smooth stochastic horizon-$H$ MDPs with unknown quadratic structure and $d$-dimensional observations. We first show that an informative TD update acquires the quadratic Bellman signal and yields a representation admitting a bounded predictive readout. The simultaneous hidden-weight and head update achieves a strict finite-sample prediction-risk improvement over matched head-only SGD, with a $2(d+1)$-fold larger population quadratic response. We then establish that an imperfect learned continuation can already provide informative supervision to an earlier zero-reward stage. By tracking coupled neural updates with one SGD step per target copy, without requiring intermediate Bellman regressions to be solved, we obtain a uniform prediction-error transient $\exp[-Ω(K/(Hd))]$, up to action-independent shifts and explicit sampling and finite-width errors. On the constructed signed-action MDP family whose total reward is bounded by 1, we show that an exploration-then-commit algorithm can achieve $\widetilde{O}(d^3H^{1/3}T^{2/3})$ regret over $T$ episodes, using $K=\widetilde{O}(Hd)$ global updates. Our analysis connects bootstrapped feature acquisition, concurrent neural refinement, and finite-budget policy performance.

cs.LG

新论文 · v1arXiv:2611.03219

Can Coding Agents Sustain Continual Evolution without Forgetting?

Shangziqi Zhao, Yitong Zhang, Jia Li

原文摘要

Recently, harness self-evolution has emerged as an important paradigm for improving agent performance. Prior studies show that evolving the agent harness can effectively enhance agent capabilities without modifying the foundation model's parameters. We find that existing approaches largely study agent evolution on a single task. In practice, coding agents may need to handle multiple task types over time. Whether an agent harness can continuously evolve across multiple tasks without suffering from catastrophic forgetting remains an important open question. In this paper, we investigate this open question through a comprehensive empirical study of continual harness self-evolution. We organize three software engineering task datasets into six different evolution directions and apply two representative harness evolution methods to perform continual evolution along these directions. Our experimental results show that evolution on a subsequent dataset can substantially undermine the performance on previously seen datasets. For example, with DeepSeek-V4-Flash-0731, Meta-Harness improves its performance on SWE-Bench to 94.0% after evolution. However, after further evolution on FeatureBench, its performance on SWE-Bench drops to 32.0% by the fifth round, corresponding to a 66% relative decrease. Finally, based on our experimental findings, we identify several insights, providing directions for the continual self-evolution of agent harnesses. Code is available at https://github.com/THU-Agent/continual-harness-evolution.

cs.SE · cs.AI

版本更新 · v2arXiv:2606.22174

OpenHLM: An Empirical Recipe for Whole-Body Humanoid Loco-Manipulation

Yingdong Hu, Haodong Zhu, Boyuan Zheng, Yihang Hu, Tong Zhang, Zunhao Chen, Junming Zhao, Ruiqian Nai 等 9 位作者

原文摘要

Whole-body humanoid loco-manipulation requires coordinating the robot's entire kinematic chain. However, most existing systems typically decouple the upper and lower bodies into separate controllers, limiting such coordination and yielding behaviors similar to those of a wheeled dual-arm platform. In this paper, we ask what it takes to build a whole-body native vision-language-action (VLA) model that maps language and pixels directly to all of the humanoid's degrees of freedom. We conduct a systematic empirical study organized as a roadmap of one-variable-at-a-time experiments across three phases: whole-body teleoperation, VLA model design, and heterogeneous co-training. Our study yields several intriguing findings: a joint-based whole-body teleoperation interface outperforms alternatives that only partially expose the humanoid's degrees of freedom; a VLA pretrained on static and wheeled dual-arm platforms transfers surprisingly well to a humanoid's full action space; and co-training with HuMI, the humanoid analog of UMI, extends the policy to new objects and instructions without additional whole-body teleoperation on those targets. Following this roadmap yields OpenHLM, an open-source recipe for whole-body humanoid loco-manipulation. In a challenging long-horizon task that spans a wide vertical range of the humanoid, OpenHLM outperforms two state-of-the-art humanoid VLA baselines (GR00T N1.6 and $Ψ_0$) using less than half the total demonstration time. Our code, training data, and model checkpoints are available at [https://openhlm-project.github.io/].

cs.RO