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

+549新增论文
+443其中 AI 相关
+62已有论文修订
+95新增 AI 导读

arXiv 最近完成扫描:2026-11-21 12:19 · 全站最近入库:2026-11-21 12:19各信源进度 →

统计口径与数据新鲜度

统计 2026-11-20 17:54 — 2026-11-21 17:54(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:2026-11-21 12:19最近发现新增版本:2026-11-21 12:19库内论文最新更新时间:2026-11-20 02:59

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。

29,158 已收录论文24,181 AI 相关8,992 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 6,976 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-21 12:17(北京时间)
版本更新 · v3arXiv:2606.07379

Do Coding Agents Deceive Us? Detecting and Preventing Cheating via Capped Evaluation with Randomized Tests

Thanawat Lodkaew, Johannes Ackermann, Soichiro Nishimori, Nontawat Charoenphakdee, Masashi Sugiyama, Takashi Ishida

原文摘要

A growing failure mode in agent evaluation and training is that models can achieve high evaluation scores (e.g., test pass rates) by exploiting shortcuts instead of solving the intended task, making these scores deceptive. This makes evaluation scores unreliable as measures of true task-solving ability. We propose CapCode, a framework for constructing coding datasets with randomized tests whose best achievable non-cheating score is deliberately capped below 100%. This capped-score design gives evaluation scores a clearer interpretation: scores substantially above the cap are implausible and therefore provide evidence of cheating. To prevent cheating, we propose CapReward, a reward design based on the CapCode principle to discourage optimization beyond the cap. Experiments across multiple datasets show that CapCode detects cheating while preserving performance ranking of models, and CapReward reduces cheating behavior, yielding models that better follow the intended task specification.

cs.LG · cs.AI · cs.CL · stat.ME

版本更新 · v2arXiv:2605.09636

PDEAgent-Bench: A Multi-Metric, Multi-Library Benchmark for PDE Solver Generation

Zhen Hang, Yushan Yashengjiang, Junhui Li, Huanshuo Dong, Yang Wei, Zhezheng Hao, Jiangtao Ma, Songlin Bai 等 24 位作者

原文摘要

PDE-to-solver code generation aims to automatically synthesize executable numerical solvers from partial differential equation (PDE) specifications. This task requires not only understanding the mathematical structure of PDEs, but also selecting appropriate discretization schemes and solver configurations, and correctly implementing the resulting formulations in finite-element method (FEM) libraries. Existing code generation benchmarks mainly evaluate syntactic correctness, or success on predefined test cases. To our knowledge, there is currently no publicly available benchmark specifically for PDE-to-solver code generation, and general-purpose code benchmarks do not fully capture the unique challenges of numerical PDE solution, such as ensuring solver accuracy, efficiency, and compatibility with professional FEM libraries. We introduce PDEAgent-Bench, to the best of our knowledge, the first multi-metric, multi-library benchmark for PDE-to-solver code generation. PDEAgent-Bench contains 645 instances across 6 mathematical categories and 11 PDE families, with common FEM libraries for DOLFINx, Firedrake, and deal.II. Each instance provides an agent-facing problem specification, a reference solution on a prescribed evaluation grid, and case-specific accuracy and runtime targets. PDEAgent-Bench adopts a staged evaluation framework in which generated solvers must sequentially pass executability, numerical accuracy, and computational efficiency checks. Experiments with representative LLMs and code agents show that models can often produce runnable code, but their pass rate drops substantially once accuracy and efficiency requirements are enforced. These results indicate that current agents remain limited in producing numerically reliable and efficient PDE solvers, and that PDEAgent-Bench provides a reproducible testbed grounded in the practical requirements of numerical PDE solving.

cs.AI

新论文 · v1arXiv:2611.20726

PEETS: Preference-aware Embeddings with Explicit Tree Structure for Auxiliary Product Bundling

Maolei Huang, Jing Tian, Dexuan Duan, Yan Chen, Huawei Cao, Shuhan Song

原文摘要

Auxiliary products such as flight protection and flexible-change services are recommended alongside primary travel purchases. This setting creates a granularity ambiguity: SKUs within one product type share semantics but differ in price and coverage, while rejecting one offer does not imply disinterest in the entire type. We present \textbf{PEETS} (\textbf{P}reference-aware \textbf{E}mbeddings with \textbf{E}xplicit \textbf{T}ree \textbf{S}tructure), a training-time framework for auxiliary product bundling. PEETS regularizes an existing click--conversion predictor through two conditional comparisons. Hierarchical Contrastive Loss uses the pair's common ancestor as a reference; Negative Contrastive Loss conditions structural sampling and comparison shifts on logged rejection events. Both act on shared SKU representations during training. This separates catalog relatedness from individual preference supervision and leaves the predictive serving path intact. PEETS achieves 78.20\% AUC, exceeding TWIN V2 by 0.90 percentage points and the matched base model by 1.15 points. In a 14-day online evaluation, gross auxiliary GMV per exposed user-day improves by 8.20\% over Policies, 3.06\% over Base, and 0.84\% over TWIN V2, with all three user-bootstrap 95\% intervals above zero. Bundling rate is 0.51 points below Policies. The cancellation interval's upper limit exceeds the prespecified margin. PEETS adds no auxiliary-loss computation to serving and records 42.1~ms p95 latency, compared with 42.0~ms for Base.

cs.AI

版本更新 · v2arXiv:2505.15275

Learning-based Autonomous Oversteer Control and Collision Avoidance

Seokjun Lee, Seung-Hyun Kong

原文摘要

Oversteer, wherein a vehicle's rear tires lose traction and induce unintentional excessive yaw, poses critical safety challenges. Failing to control oversteer often leads to severe traffic accidents. Although recent autonomous driving efforts have attempted to handle oversteer through stabilizing maneuvers, the majority rely on expert-defined trajectories or assume obstacle-free environments, limiting real-world applicability. This paper introduces a novel end-to-end (E2E) autonomous driving approach that tackles oversteer control and collision avoidance simultaneously. Existing E2E techniques, including Imitation Learning (IL), Reinforcement Learning (RL), and Hybrid Learning (HL), generally require near-optimal demonstrations or extensive experience. Yet even skilled human drivers struggle to provide perfect demonstrations under oversteer, and high transition variance hinders the accumulation of sufficient data. Hence, we present Q-Compared Soft Actor-Critic (QC-SAC), a new HL algorithm that effectively learns from immature demonstration data and adapts rapidly to new conditions. To evaluate QC-SAC, we introduce a benchmark inspired by real-world driver training: a vehicle encounters sudden oversteer on a slippery surface and must avoid randomly placed obstacles ahead. Experimental results show that QC-SAC attains near-optimal driving policies, significantly surpassing state-of-the-art IL, RL, and HL baselines. Hardware-in-the-loop simulation (HILS) validation with physical steering hardware further demonstrates that QC-SAC can operate beyond a software-only simulation, providing an intermediate validation step toward real-vehicle implementation. To the best of our knowledge, this is the first work that achieves safe autonomous control of vehicles in oversteer while simultaneously avoiding obstacles.

cs.RO · cs.AI · cs.LG

新论文 · v1arXiv:2611.20711

AgentFaultBench: State-Verified Evaluation of Tool-Fault Recovery

Sachin Gupta

原文摘要

Does recovery advice help an agent finish a task after a tool fails? AgentFaultBench evaluates LangGraph and CrewAI agents performing specified tasks in an isolated GitLab installation. We inject tool errors and check both the resulting files and issues and the required reads, verification and final response. The prompt supplies the expected issue and report text, so the task tests execution rather than content generation. In the primary comparison of 1,944 planned runs, advice changes success by -8.33, 0.00 and +1.06 percentage points relative to a neutral reminder; missing outcomes leave the latter two comparisons inconclusive. Later experiments test a revised recovery policy. GPT-4o mini improves, while Sol and Claude already complete the recoverable cases. In a separate 624-run study on six related tasks not previously used by that policy, the guided condition records 161 correct final states and one unknown outcome in 192 planned runs, versus 142/192 with neutral feedback. The gain lies between +9.90 and +10.42 pp when missing outcomes are included. Removing issue IDs gives 47/48 successes versus 46/48 with IDs; other execution information remains. A second checker agrees on 1,879 observed primary outcomes and 241 constructed test cases. No human validation was collected. The findings support testing advice against explicit recovery controls rather than assuming it helps.

cs.AI · cs.LG

版本更新 · v2arXiv:2606.30953

Neuro-Bayesian-Symbolic Residual Attention Shallow Network: Explainable Deep Learning for Cybersecurity Risk Assessment

Nicolaie Popescu-Bodorin, Madeleine Togher

原文摘要

We introduce NBS-RASN, a novel hybrid neural architecture for explainable cybersecurity risk evaluation in open-source software ecosystems. Unlike conventional deep models that trade interpretability for predictive power, it uses a meticulously designed shallow network that encodes domain knowledge, causal reasoning, and expert judgment as differentiable components. The network has 80 interpretable neurons across 12 layers, including a gatekeeper layer enforcing five epistemological axioms-precision, causality, falsifiability, transparency, and completeness-as hard constraints before propagation. Despite limited depth, non-linear residual attention and feedback loops give it deep-learning characteristics, enabling complex context-dependent risk patterns without becoming a black box. It produces fully decomposable risk scores: a deterministic weighted component plus an expert adjustment component, with every adjustment traceable to named amplifiers (blast radius, propagation speed, structural nature, default exposure, exploitation pattern, institutional criticality). We validate on 20 open-source projects spanning all OWASP Top 10:2025 categories and language risk classes, achieving confidence scores of 0.79-0.97 - precision-recall-derived certainty estimates from bootstrap validation against confirmed CVE incidents in CISA KEV and NVD, not softmax probabilities - and show explainability is guaranteed by architectural design. This addresses the assumed link between network depth and representational power, showing a shallow network with deep reasoning can outperform opaque models in high-stakes cybersecurity where interpretability is non-negotiable.

cs.AI · cs.LG

版本更新 · v2arXiv:2610.22872

Look Back, Correct Forward: Error Feedback Calibration for Time Series Forecasting

Haonan Yang, Yulin He, Zhuo Li, Jing Zhao, Huibin Tan

原文摘要

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

新论文 · v1arXiv:2611.20699

A First Look at System One Decision Models for Malware Detection

Pier Mattia Basciano, Giovanni Farina, Salvatore Monteleone

原文摘要

Malware detection needs scores calibrated at a fixed false positive rate (FPR) and decisions an analyst can read: feature-based detectors give the first, large language models (LLMs) on decompiled code the second, through generated text of uncalibrated confidence. System One decision models answer a typed question about a text with a probability over its declared answers in one forward pass. We present the first evaluation of these models for Windows Portable Executable malware detection, analyze how they decide, and propose S1Mal, an ensemble on the typed answers and the embedding of CLM-8B. In an isolated virtual machine, Ghidra and capa decompose each binary, and the host renders the records into a canonical textual state, a digest of headers, imports, MITRE ATT&CK capabilities and strings, optionally with decompiled functions, without class shortcuts. Three readouts score each model on the state: A0, the probability of one verdict question; A1, 32 atomic behavioral questions weighed by gradient-boosted trees; A2, heads on the model's state embedding. S1Mal feeds gradient-boosted trees with the atomic answers and the A2 score of CLM-8B. We evaluate CLM-8B, ten open decision models and two LLMs read through logits on BODMAS temporal splits, SOREL-20M out-of-distribution malware and 14,534 matched benign files. On BODMAS, A0 wastes the static evidence: its area under the receiver operating characteristic curve (AUROC) spans 0.4025-0.8590 across 13 models. A1 recovers it, raising every model's AUROC by +54.61% on average; A2 lifts CLM-8B by +0.2749 to 0.9943, above a combiner on digest facts (0.9901), and reaches 0.9853-0.9948 on nine open models. S1Mal reaches 0.9946 and detects 93.47% of the malware at a realized FPR of 1.45%, and 59.15% at 0.1% FPR on the test ROC curve, against 46.59% with A2. S1Mal gives analysts calibrated decisions, each with the atomic answers behind it.

cs.CR · cs.AI · cs.LG

新论文 · v1arXiv:2611.20695

IndustrialKnowEval: A Benchmark for Evaluating Multi-Level Industrial Knowledge in LLMs

Haiteng Wang, Haoze Li, Jingwei Guo, Jingbo Zhou, Si Liu, Lei Ren

原文摘要

Large language models (LLMs) are increasingly explored for industrial tasks, yet few benchmarks evaluate their mastery of industrial knowledge across different types of cognitive demand. We introduce IndustrialKnowEval (IndKnow), a benchmark for industrial knowledge and engineering question answering. It contains 43,282 samples covering 21 task types across seven industrial sectors, including aerospace, electric power, and metallurgy, organized into four levels of cognitive demand: memory, comprehension, reasoning, and engineering application. Enterprise-based domain specialists and experienced researchers participate in all stages of dataset construction and quality control. We conduct a series of experiments on 18 representative LLMs, with domain experts deeply involved in the evaluation. Building upon IndKnow, we propose the Industrial Cognitive Capability Gap (ICCG) metric to characterize the performance differential between foundational industrial knowledge and higher-order reasoning and application, and we analyze it alongside absolute performance to provide a more complete view of model capabilities. We anticipate that IndustrialKnowEval will support the benchmarking of LLMs in industrial scenarios and promote the development of robust industrial intelligence. Code is available at: https://github.com/J4cksonLee/IndKnow

cs.CL · cs.AI

新论文 · v1arXiv:2611.20691

SpeCal: Spectral Calibration for Unsupervised RLVR

Kai Tang, Xing Zheng, Haobo Wang, Guangcheng Zhu, Shenzhi Yang, Zhengqing Zang, Yingfan Ma, Cheng Peng 等 10 位作者

原文摘要

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning abilities of large language models, but its dependence on ground-truth answers or external verifiers limits its ability to exploit abundant unlabeled reasoning data. Unsupervised RLVR alleviates this bottleneck by deriving scalar rewards from the model's own rollouts. However, the reliability of such self-generated supervision varies substantially across rollout groups and individual responses, while existing methods construct or refine their rewards mainly from decoded outcomes and reward statistics. These endpoint signals reveal what answers the model produces, but provide limited information about whether the underlying rollout representations form a reliable structure for learning. Our key observation is that hidden-state spectra offer a complementary view of supervision reliability by characterizing both the organization of a rollout group and the support received by each rollout within that group. Based on this observation, we propose SpeCal, a bi-level spectral calibration framework for unsupervised RLVR. At the group level, SpeCal combines spectral mass with a Gaussian band-pass gate over normalized entropy-based effective rank, favoring rollout groups with substantial representation variation whose effective dimensionality lies in a preferred regime. At the rollout level, the same group effective rank determines how many dominant directions are retained in a leave-one-out spectral subspace formed by the remaining rollouts, within which the conditional support of each held-out rollout is measured. The resulting signals calibrate the original reward-induced advantages at the group and rollout levels without redefining the underlying unlabeled reward.

cs.LG · cs.AI

新论文 · v1arXiv:2611.21691

AdaFlowLite: Scalable and Non-blocking Inference on Asynchronous Mobile Data

Sicong Liu, Fengmin Wu, Yuan Gao, Bin Guo, Zimu Zhou, Hongkai Wen, Zhiwen Yu

原文摘要

The rise of mobile devices equipped with numerous sensors, such as LiDAR and cameras, has driven the adoption of multi-modal deep intelligence for distributed sensing tasks, such as smart cabins and driving assistance. However, the arrival time of mobile sensory data vary due to modality size and network dynamics, which can lead to delays (if waiting for slow data) or accuracy decline (if inference proceeds without waiting). Moreover, the diversity and dynamic nature of mobile systems exacerbate this challenge. In response, we present a shift to \textit{opportunistic} inference for asynchronous distributed multi-modal data, enabling inference as soon as partial data arrives. While existing methods focus on optimizing modality consistency and complementarity, known as modal affinity, they lack a \textit{computational} approach to control this affinity in open-world mobile environments. % \sysname pioneers the formulation of structured cross-modality affinity in mobile contexts using a hierarchical analysis-based normalized matrix. This approach accommodates the diversity and dynamics of modalities, generalizing across different types and numbers of inputs. Employing an multi-modal lightweight Swin Transformer (MMLST), \sysname facilitates real-time and flexible data imputation, adapting to various modalities and downstream tasks without retraining. Experiments show that \sysname significantly reduces inference latency by up to 80.4\% and enhances accuracy by up to 62.1\%, while achieving nearly a 50\% reduction in energy consumption, outperforming status quo approaches. Also, this method can enhance LLM performance to preprocess asynchronous data.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20688

VIVID: A Visual Harness for Agentic Video Editing

Wenhan Yu, Zhaoxi Zhang, Zhilong Tan, Weikang Li, Lei Sha

原文摘要

Video editing requires agents to make editing decisions grounded in audiovisual observations and precise temporal localization, and to evaluate their effects in rendered outputs. General-purpose coding agents possess strong comprehension, reasoning, and code execution capabilities, while media interaction and evidence management mechanisms tailored to video tasks can further unlock their editing potential. We introduce VIVID, a visual harness that supports multiple native coding agents in video post-production without additional model training. VIVID provides tools for on-demand video inspection, audio analysis, and transcription, and preserves observations alongside their corresponding media versions and temporal coordinates. Agents can actively retrieve these observations as they reason, perform edits using their native coding and shell tools, and inspect rendered outputs to guide subsequent revisions. On AgenticVBench, comprising 100 tasks across Assembly, Repair, Sequencing, and Repurpose, Codex with VIVID achieves an average score of 58.6%, an improvement of 20.2 percentage points over native Codex. The same harness also integrates with Claude Code and OpenCode, achieving average scores of 55.1% and 37.1%, respectively, both substantially exceeding their native baselines. These results highlight VIVID's potential as a reusable visual harness for extending the capabilities of general-purpose coding agents to video editing.

cs.AI · cs.CV · cs.MM

新论文 · v1arXiv:2611.20686

UbiHearo: Bringing Scenario-Aware Sound Guidance to DHH Users with Mobile Agents

Fengmin Wu, Sicong Liu, Zimu Zhou, Chenren Xu, Wanbing Zhao, Peiwen Sun, Zhiwen Yu

原文摘要

For Deaf and Hard-of-Hearing (DHH) users, effective sound awareness goes beyond detecting \emph{what} sound occurs; it requires understanding \emph{what it means}, \emph{how urgent it is}, and \emph{what to do}. However, existing mobile assistants largely reduce acoustic scenes to isolated labels or generic alerts, failing to provide contextual guidance when the same sound implies different risks across scenarios. We present \sysname, a mobile hearing assistant agent that closes the sensing--reasoning--action loop for \textit{scenario-aware} and \textit{personalized} DHH assistance. \sysname adopts a \textit{neuro-symbolic} design that transforms continuous acoustic and physical context streams into structured intermediate representations (IRs) for controllable reasoning and guidance. It introduces three key mechanisms. \textit{First}, a perception-thresholded sensing mechanism enables energy-efficient always-on awareness by activating processing only for perceptually meaningful events while preserving safety-critical cues through deterministic overrides. \textit{Second}, a staged neuro-symbolic reasoning framework performs promptless scenario understanding by conditioning an on-device audio-language model on structured IRs instead of user-authored context descriptions. \textit{Third}, a bounded human-in-the-loop personalization mechanism adapts user-specific guidance through lightweight on-device decision-boundary updates while preserving population-aligned reasoning. We evaluate \sysname on smartphones and in-vehicle systems across diverse scenarios. It achieves favorable accuracy--latency--energy tradeoffs and personalized guidance.

cs.AI · cs.CY

版本更新 · v2arXiv:2608.27259

LEON: Operator-Structured Latent Evolution for World Modeling

Xiaoxiao Lu, Jiahao Shi, Yunlong Dong, Ye Yuan

原文摘要

Modeling action-conditioned physical evolution is fundamental to latent world models. World Action Models (WAMs) couple such models with policy learning through predictive supervision or future-conditioned action generation. Yet their latent transitions typically rely on transformer-based predictors, leaving dynamical structure to be learned implicitly from data. We derive a latent-transition formulation from controlled mechanical evolution. A common finite-dimensional projection of observable dynamics yields a shared operator family affine in physical control and an explicit projection residual. Guided by this principle, we introduce the Latent Evolution Operator Network (LEON), which adapts pretrained visual features into learned observables and evolves them through an affine family of shared operator matrices and complementary additive updates. Coefficients inferred from evolving observables and action-related context modulate the operators. We instantiate LEON in two WAM architectures while preserving their distinct original prediction targets and policy interfaces. Across single-arm and bimanual manipulation, LEON improves average policy success over corresponding WAM baselines, including under diverse perturbations. Ablations support its observable adaptation and operator-structured evolution, while controlled-system evaluations demonstrate improved extrapolation beyond training conditions.

cs.LG · cs.AI · eess.SY

新论文 · v1arXiv:2611.21690

EgoLink: Egocentric Language-Vision Interactive Network Knowledge Challenge Summary

Renhe Sun, Yueying Feng, Bohan Yu, Tong Niu, Yunqi Liu, Zitong Wang, Haolin He, Kexin Shi 等 15 位作者

原文摘要

The EgoLink Challenge at ACM Multimedia 2026 benchmarks egocentric multimodal intelligence linking first-person perception, social reasoning, and interactive action. Unlike benchmarks centered on visible objects and activities, EgoLink evaluates whether models can interpret emotional and social cues, resolve visually grounded references, and convert user requests into verified tool-mediated actions. Its two tracks cover social reasoning in egocentric video and interactive agents that complete service tasks through dialogue, tool use, and closed-loop execution. Together, they test whether perception, reasoning, and decision-making remain coupled under both offline question answering and state-changing interaction. We highlight a multi-stage Track 1 data pipeline for multiple-choice question (MCQ) generation, chain-of-thought generation, automatic quality grading, and video-contribution filtering. We summarize the task design, evaluation metrics, rules, leaderboards, and representative solutions. The benchmark combines objective answer scoring with process- and result-based evaluation, enabling analysis of both final correctness and execution quality. The best Track 1 system achieved 0.9263 accuracy, while the best Track 2 system achieved 0.844 joint success and 0.916 micro accuracy. Participant reports reveal a shared principle: strong systems make intermediate evidence explicit through candidate verification, option-robust reasoning, scenario-aware planning, tool canonicalization, and state-aware correction before answering or acting. These findings highlight the importance of process verification and evidence-grounded reasoning for reliable egocentric multimodal intelligence.

cs.CV · cs.AI

新论文 · v1arXiv:2611.20671

MASTA: A Feedback-Scheduled Multi-Agent System for End-to-End Tamarin Modeling and Analysis

Yongkang Xiao, Qiyi Deng, Jing Chen, Min Shi, Ju Ma, Ruiying Du

原文摘要

Tamarin-based formal verification of security protocols requires expert-authored multiset rewriting rules and temporal-logic lemmas. Translating a natural-language protocol description into a complete formal model with an LLM is hard because syntactic, reachability, and property-level errors compound across the multi-step modeling process, and a single agent that both authors and self-evaluates cannot reliably localize them. We present MASTA, a multi-agent system that addresses these difficulties through three innovations. First, an operational--evaluative role decomposition assigns rule modeling to an Architect, property modeling to an Auditor, and pure dispatch to a dedicated Arbiter that holds no authoring duties. Second, the Arbiter implements a Tamarin-anchored scheduler that dispatches the next agent based on the latest verifier verdict and routes problem-tagged fixes through a modification list, with a six-class problem taxonomy and an LLM deduplication check preventing routing deadlock. Third, we construct a benchmark of 10 protocols paired with 139 mutation variants and five fine-grained metrics that jointly assess model correctness ($M_1$--$M_3$) and completeness ($M_4$--$M_5$). On this benchmark, MASTA on Qwen 3.5-plus improves over OpenCode on the same backbone by up to $+0.27$ on individual metrics and ranks first on both the correctness and completeness averages, surpassing Codex (GPT-5.5) and Claude Code (Opus 4.7) despite using a weaker LLM. Ablations on five protocols confirm that removing the Arbiter collapses property and verification scores to zero.

cs.CR · cs.AI

版本更新 · v5arXiv:2605.03940

Reentrant value fields as delayed coupled reaction-diffusion systems on finite graphs

Karsten Bohlen

原文摘要

We describe a dynamical system in which a symbolic field is coupled to a geometric field via a bipartite Hilbert-Schmidt kernel. The system is fully described by a retarded functional differential equation (RFDE) on the history space, subject to Lipschitz and small gain conditions. We show that the RFDE is well-posed under constant input and that it admits a compact global attractor. The principal subsystem $(H_L, X_R, P)$, which is comprised of the two primary fields as well as an executive field, is shown to be globally stable independent of delay, provided that the interfield coupling satisfies $C_{\mathcal{K}}^2<μ_Lμ_R$. In addition, we describe design specifications that fulfill the hypotheses of the main Theorem.

math.DS · cs.AI

新论文 · v1arXiv:2611.20644

Instrumenting Failure: Scoring Language Model Agents at Every Decision Node in Imperfect-Information Games

Michael Kingston, Sonit Singh, Alan Blair

原文摘要

Agent evaluations report task success and rarely say what an agent did at each decision. We score every decision node of three 12B to 14B instruct-tuned language models (LMs) playing five zero-sum imperfect-information games against four opponents on eight properties from decision and game theory, five of them against a CFR+ equilibrium reference. At each node, the LM also states a belief about the opponent's hidden state in a separate call. Without memory and at temperature 0, stated beliefs sit 0.250 to 0.288 in total variation distance from the equilibrium posterior, and where the action responds to the state its agreement with that belief exceeds chance by -0.04 to 0.19. The two calls frame the opponent differently, so the result is a behavioural inconsistency and makes no claim about internal beliefs. Under one prompt and greedy decoding, the LM's action never varies with the state in 10 of the 38 cells where the reference varies its own. Memory lowers the constant-action share in 9 of 13 testable ladders at temperature 0; at temperature 0.2 the cells at 1 fall from 16 to 5 of the 46 informative cells, while the belief results hold. A gate withholds, with a stated reason, any belief score the run cannot support, and seven of the eight non-zero defect rates are attributable to the instrument, a representation, or the LMs.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20643

Safety-Preserving Medical LLM Fine-Tuning via Bilevel Data Selection

Ruoxuan Shi, Sheng Yang, Zhengxing Su, Xiaoyang Hou, Yating Liu

原文摘要

As general large language models (LLMs) are increasingly adapted for patient-facing medical applications, improving medical capability while controlling patient-facing risks remains a critical challenge. Large-scale medical instruction data are essential for acquiring domain knowledge through supervised fine-tuning (SFT), but unsafe responses can undermine existing safety alignment. To address this problem, we propose Scalable Bilevel Data Selection (S-BDS) for safety-preserving medical SFT. By controlling selection-weight concentration, S-BDS enables reliable sample ranking over large candidate pools and identifies training examples that improve medical capability while controlling patient-facing risks. We combine clinician-validated safe data and selected data with unscreened medical instruction data for SFT, and jointly evaluate medical capability and patient-facing safety on HealthBench and PatientSafetyBench, respectively. Experiments across Llama and Qwen models show that S-BDS achieves the highest safe rates among all compared methods, outperforming the strongest baselines by 4.64 and 3.00 percentage points, respectively, while maintaining competitive medical capability. Further experiments show that varying the safety-oriented weight during SFT controls the trade-off between medical capability and patient-facing safety.

cs.CL · cs.AI · cs.LG

新论文 · v1arXiv:2611.20631

Instructions Before Actions: Prospective and Retrospective Language Supervision for Driving Vision-Language-Action Data

Angel Martinez-Sanchez, Ross Greer

原文摘要

Driving vision-language-action (VLA) models can accept navigation instructions, yet predicting a driver's trajectory does not establish that a model follows the instruction. We study how driving-language supervision is collected and evaluated, distinguishing prospective instructions issued before an action from retrospective labels written after observing it. We develop ADVLAT-Engine to synchronize spoken instructions, road-scene video, and vehicle trajectories, and compare three navigation apps on 48 replayed routes. On 87 maneuvers announced by both Google Maps and a passenger, the passenger uses more scene-dependent language. We evaluate four open driving VLAs before and after fine-tuning on prospective navigation-app drives. Fine-tuning reduces passenger-drive trajectory error by 23-59%, but instruction effects remain inconsistent. At real intersections, counterfactual instructions reveal stronger responsiveness: fine-tuned Alpamayo~2 Super gains 31 and 18 accuracy points at 15 and 5~m, while Qwen-Drive follows alternative maneuvers more reliably with a discrete command than with a sentence. We additionally collected retrospective annotations from four people for the driving recordings and describe controlled comparisons of annotation content, timing, and downstream learning; those analyses are pending. Our findings separate maneuver prediction from instruction-following and establish a framework for comparing prospective and retrospective supervision.

cs.CV · cs.AI · cs.LG · cs.RO

版本更新 · v3arXiv:2605.21061

Grounding Driving VLA via Inverse Kinematics

Junsung Park, Hyunjung Shim

原文摘要

Counterfactual obstacle insertion exposes blind planning in representative Driving VLAs: their trajectories remain largely unchanged despite a vehicle appearing directly ahead. An empirically calibrated bias-variance analysis identifies how ego-predictable training examples admit a shortcut through ego status and commands. We address this failure with an inverse-kinematics-inspired architecture that anchors trajectory decoding on current and predicted future visual states. First, a next visual state prediction objective that requires the LLM to predict the future visual scene provides dense visual supervision and suppresses shortcut paths. Second, a separate Inverse Kinematics Network (a cross-attention-based conditional diffusion model) that takes only the current and future visual states as input is designed to suppress reliance on ego status and textual shortcuts during trajectory decoding. Our 0.5B model achieves state-of-the-art trajectory planning performance on NAVSIM-v1, NAVSIM-v2, and nuScenes using cameras alone, without LiDAR or a BEV encoder. Against the matched OpenDriveVLA* baseline, it improves NAVSIM PDMS and EPDMS by 19.0 and 20.4 points, respectively. Controlled ablations and intervention analyses support improved use of visual features, with future-state prediction sharpening hazard-specific responses.

cs.CV · cs.AI · cs.RO

版本更新 · v2arXiv:2611.02688

Accelerating LLM Pretraining with Zeroth-Order Curvature on Idle Serving Compute

Mingchuan Ma, Dongqi Cai, Daqian Yang, Zhenyan Lu, Chen Chen, Teresa Lv, Yan Gao, Shangguang Wang 等 11 位作者

原文摘要

LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.

cs.LG · cs.AI

版本更新 · v2arXiv:2611.16424

PhysCapBench: A Cross-Domain Benchmark for Physically Grounded Video Captioning

Haofeng Wang, Xuanyi Liu, Pengcheng Wen, Ruiqi Li, Yuhao Zhang, Tiange Zhang, Boren Zheng, Siyu Tao 等 15 位作者

原文摘要

Physical AI aims to learn transferable physical knowledge from real-world experience to understand, predict, and act in the physical world. Internet videos and embodied demonstrations provide complementary observations of motion, interaction, and state evolution. A foundational data challenge is to measure and preserve this physical information as training supervision. In video-language pretraining, captions express observed processes in language, yet a semantically correct description can omit the interactions and transitions that define those processes. Rich physical experience in video does not automatically translate into sufficient language supervision. This gap calls for a yardstick to assess annotation quality, compare captioning strategies, and guide training-data refinement. Motivated by our data construction for world-model pretraining, we introduce PhysCapBench, to our knowledge the first unified physical-caption benchmark spanning all three domains: third-person scenes, egocentric manipulation, and handheld-gripper demonstrations. It provides human-verified annotations for 1,500 video sequences and evaluates accuracy and coverage through seven physical dimensions. We examine downstream utility through video reconstruction and manipulation-policy fine-tuning. Physical captions reduce reconstruction LPIPS by 8.6-9.5% on third-person videos and 26.9-30.9% on egocentric videos relative to metadata captions. Training-instruction enrichment improves both PhysCapBench scores and policy success, including a 7.5-percentage-point gain for pi0.5 on a RoboTwin 2.0 subtask. PhysCapBench makes the physical content of language supervision inspectable and comparable, providing a cross-domain basis for its evaluation and improvement.

cs.CV · cs.AI

新论文 · v1arXiv:2611.20616

GEAR$^2$: Pruned Experts Live On through Gated Reuse of Activated Experts and a Missing-Mass Residual

Linliang Chen, Yan Zhong, Wang Kang, Sai Li, Ji Wu, Ji Liu

原文摘要

Mixture-of-Experts (MoE) Large Language Models (LLMs) activate only a few experts per token while generally keeping all the experts in GPU memory for high efficiency. Expert pruning reduces this GPU memory footprint cost, yet discards the contribution of pruned experts. Existing remedies either rely on static substitutes, merge experts at the cost of altering their functions, or require costly recovery training. In this paper, we find that the functionalities of pruned experts can be efficiently reconstructed from the activated remaining experts. We propose GEAR (Gated Expert Activation Reuse), which reuses the scaled outputs of reliable activated experts without any training or additional expert computation. In addition, we propose GEAR$^2$ to further recover what reuse cannot express with an ultra-lightweight residual. On five MoE LLMs from 16B to 106B parameters with 25% and 50% of experts pruned, GEAR improves over its training-free counterparts in nearly every setting (up to 4 pp.) in terms of accuracy, and GEAR$^2$ outperforms baseline approaches (up to 2.75 pp.) in every setting on nine zero-shot benchmarks, with larger advantages (up to 2.72 pp. higher) under more aggressive pruning (up to 50%).

cs.AI

版本更新 · v2arXiv:2611.15146

Reframing Video Inverse Tone Mapping as Blind Posterior Inference via Latent Luminance Supervision

Hanwool Sul, Chang-Hoon Jeong, Kisung Shin, Bomseok Park, Donghoon Cha, Hansl von Kwon, DoHyung Ku, Byoung-Tak Zhang

原文摘要

High-dynamic-range (HDR) displays are now common, but most video content, such as archival footage, broadcast material, and the output of generative video models, exists only in standard dynamic range (SDR). The need is sharpest for generated video, which has no HDR master at all and is graded alongside live-action and visual-effects material finished in HDR. Converting SDR to HDR means undoing a tone-mapping operation that was never recorded, and clipping destroys the radiance wherever it acts, so HDR recovery is a blind inverse problem. Methods built on generative video priors fill clipped regions with plausible structure, but their reconstructions are systematically biased: in every public method we test, saturated content leans too bright or too dark, and the direction of the lean moves with the tone operator behind the input. We call this failure the luminance error, measure both its magnitude and its sign, and make it the focus of this work. Reframing ITM as posterior inference under an unknown operator locates the failure in two places and yields two corrections. A fixed affine functional of the video-diffusion latent reads log-luminance without decoding; on that coordinate an asymmetric loss selects an expectile of the posterior, applied only where clipping leaves luminance unconstrained. We also synthesize training SDR through an explicit family of tone operators, so the model marginalizes over what the observation cannot show. Trained this way, our luminance error stays flatter across operators and smaller on average than any public method, and the full system achieves the best perceptual quality on all three benchmarks.

cs.CV · cs.AI

新论文 · v1arXiv:2611.20609

Newton-PINet: A fast physics-informed neural network with Newton linearization for meta-learning nonlinear PDEs

Yuchen Fan, Chang Wei, Pao-Hsiung Chiu, Chin Chun Ooi, Heyang Wang, Jian Cheng Wong

原文摘要

Scientific machine learning has opened new avenues for solving parameterized partial differential equations (PDEs), enabling models to learn a family of PDEs and generalize to unseen instances. In this context, data-driven operator learning methods typically require large training datasets, while physics-informed neural networks (PINNs) suffer from difficult optimization and limited generalization, especially for nonlinear PDEs. We propose Newton-PINet, a physics-informed network enhanced by Newton linearization, offering an effective meta-learning framework for nonlinear PDEs. Newton-PINet (i) employs a physics-informed multilayer network with skip connections, where the output-layer weights are solved by least squares; (ii) adopts a two-stage learning strategy that first leverages gradient-based training to learn robust representations from the available training tasks, and then performs gradient-free fine-tuning on the output layer for fast task-specific generalization; and (iii) incorporates a Newton linearization method to speed up the least-squares iteration for nonlinear PDE problems. On a challenging nonlinear reaction-diffusion benchmark, Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers, while using 16x fewer training tasks and over an order of magnitude less training time (about 5 minutes versus several hours). This work advances the meta-learning of PINNs toward data-efficient, fast, and generalizable physics solvers. The code and datasets are publicly available at: https://github.com/fycAIwj/Newton-PINet.

cs.LG · cs.AI · cs.CE · physics.comp-ph

新论文 · v1arXiv:2611.20608

Same Papers, Different Verdicts: Tracing Version Drift in LLM Peer Review

Jizhou Guo

原文摘要

Venues and benchmarks read LLM review scores as measurements of a paper, while providers replace their models within months. We audit twenty frontier models from OpenAI, Google and Anthropic on ICLR 2027 submissions while they were under review, with tests registered before the decisions. Between an earlier and a later version of Claude Opus and of Gemini, the share of papers placed on the accept side falls by 52 and 53 percentage points on 183 and 60 randomly drawn submissions, and by 43 to 79 points in six further paired comparisons, with no paper moving up in any of them. Stating a 30% acceptance rate in the prompt leaves a gap of 79 points between the two Opus versions. When score-stripped critiques are swapped between versions, the later critiques alone lower the accept share by 35 points for Opus and 51 for Gemini; for Opus the drop is 35 points under a third-party scorer and 24 after a rewrite that keeps each technical point and removes evaluative wording. On 57 papers, the percentiles of Opus 4.6 and 5.5 correlate at 0.73 while their accept shares differ by 44 points. On 99 to 119 submissions, the self-estimated percentiles of three models correlate with two human reviewers more strongly than a third reviewer does (point estimates 0.17 to 0.21, 95% intervals between -0.04 and +0.42). All twenty models give a 2 or a 3 in 473 of their 475 reviews, against 60 of 72 human reviews. Retrospective tests on decided papers can draw on memory. Claude Opus 5.5 names the true first author of 27 decided ICLR 2026 papers from their titles alone, with no wrong name, and Claude Opus 4.6 accepted all 11 decided papers it said it recognized, against 5 expected. Thresholds on LLM-reviewer scores therefore have to be validated against human decisions for each model version; our tests against the ICLR 2027 decisions follow on 16 December 2026.

cs.CL · cs.AI · cs.CY · cs.DL · physics.soc-ph

新论文 · v1arXiv:2611.20605

AdMem: Expectation-Grounded Procedural Memory for Long-Horizon Task-Solving Agents

Runzhe Wang, Huilin Lu, Shengjie Liu, Li Dong, Zhihuai Zhu

原文摘要

An agent that carries memory across tasks is only as good as the entries it keeps. We begin from an observation that is easy to reproduce and uncomfortable for the field: naively accumulating cross-task experience can make an agent \emph{worse} than one with no memory at all. On AgentBoard, a workflow-induction memory solves fewer tasks than a memoryless ReAct agent in seven of eight domains, and a per-step procedural store with no surrounding structure loses half of ReAct's completions on Jericho. We argue that the missing ingredients are per-action credit assignment and an explicit life cycle for stored entries, and present AdMem, which supplies both. Before each action the actor commits to an expected outcome; a concurrently running critic scores the realized outcome against that expectation and writes a procedural entry carrying a binary reward and a reflection, so credit lands on the action that earned it and no environment reward is required. Each entry then carries a scalar utility, updated online under a noisy-OR reward model, that ranks retrieval jointly with context similarity and drives eviction and merging, giving the store a forgetting and consolidation policy instead of unbounded growth. Across 707 tasks in eight AgentBoard domains AdMem solves 441 against 388 for ReAct and 312 for the workflow-induction baseline --- a pooled gain significant under a two-sided Fisher exact test, concentrated in the domains that supply transferable cross-task structure. All memory stays human-readable text, which is what makes the failures above diagnosable --- and, as we discuss, also makes the store an auditable but attackable surface.

cs.AI

新论文 · v1arXiv:2611.20603

BiGSyn: A Bi-Level Heterogeneous Graph Contrastive Learning Framework for Interpretable Drug Synergy Prediction

Senquan Wang, Hui Fang, Haoran Nan, Xingwei Wang, Qiang He

原文摘要

Drug combination therapy has become a mainstream of modern medicine, with the potential to enhance efficacy, reduce toxicity, and overcome drug resistance. However, identifying synergistic drug combinations remains challenging, as it requires modeling complex interactions across multiple biological scales---from atom-level molecular structures to entity-level relations in biological networks. Existing approaches typically model these levels in isolation, overlooking complementary signals and long-range dependencies. To address these limitations, we propose BiGSyn, a Bi-level heterogeneous Graph contrastive learning framework for drug Synergy prediction. Specifically, BiGSyn leverages heterogeneous graph transformers to extract high-order dependencies from molecular graphs and biological networks, and introduces a bi-level contrastive learning objective to bridge the cross-level semantic gap. Extensive experiments on two benchmark datasets show that BiGSyn consistently outperforms state-of-the-art baselines. Moreover, its attention mechanism highlights key molecular substructures and biological interactions driving synergy, providing interpretable insights.

cs.AI

版本更新 · v2arXiv:2611.12961

RBLA+: Efficient Rank-Specific LoRA Adapter Extraction for Federated Fine-Tuning with Heterogeneous Clients

Shuaijun Chen, Omid Tavallaie, Xin Chen, Wentao Gao, Nguyen H. Tran, Albert Y. Zomaya

原文摘要

Federated fine-tuning with low-rank adaptation (LoRA) reduces communication and computational overhead by training and exchanging compact trainable LoRA factors. Clients may use different LoRA ranks to meet adapter communication budgets and accommodate different local adaptation needs. The server must process these differences during aggregation and construct adapters at each client's rank for local training. We distinguish pre-aggregation alignment, which handles factor dimensions and scaling conventions before aggregation and rank-specific LoRA adapter extraction, which constructs client-sized pairs from the aggregated global model. Two common approaches to rank-specific extraction are direct rank truncation and Singular Value Decomposition (SVD) of the dense update. Direct prefix truncation keeps the leading columns of one aggregated factor and the corresponding rows of the other. It is cheap but can increase reconstruction error. The dense SVD approach first forms the product of aggregated LoRA factors. It then decomposes this matrix and retains the leading singular components to construct adapters at the target rank. This gives the optimal rank-constrained approximation in Frobenius norm but incurs significant computational overhead. To address this trade-off, we propose RBLA+ as an extension of RBLA. It uses established compact QR/core-SVD decomposition to minimize Frobenius error for the extracted LoRA update without forming the dense update. It also accounts for client and server scaling conventions during upload and broadcast. Across vision and language tasks, RBLA+ improves accuracy over RBLA in multiple heterogeneous settings. In CPU benchmarks with synthetic adapters, RBLA+ has lower extraction time than the dense partial-SVD baseline. RBLA+ preserves each client's prescribed adapter size and per-round communication payload.

cs.AI · cs.DC