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 18:21 — 2026-11-21 18:21(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 · v2arXiv:2607.07984

Agentic Neural Architecture Search

Seokhoon Jeong, Mijung Kim, Taehwan Kim

原文摘要

Neural architecture search (NAS) methods have grown increasingly efficient, but they still operate within manually designed search spaces that require domain expertise and must be rebuilt for each new task. Large language models (LLMs) can write architectures in an open-ended code space, yet it is unclear how the work should be divided between LLM-driven design and NAS-driven search. We propose a way to combine the two. An LLM first designs a seed architecture and then decomposes it into a slotted architecture: a scaffold with named, interchangeable module slots that defines a bounded, task-specific search space, which a conventional NAS algorithm then explores without any manual engineering. We implement this in AgentNAS, a three-phase pipeline whose phases can be evaluated separately. On 17 tasks from NAS-Bench-360 and Unseen NAS, covering classification, dense regression, segmentation, and multi-label tagging, AgentNAS outperforms published baselines, including task-specific expert designs, on 11 tasks. Extensive ablation studies show that LLMs find strong architectures quickly but soon plateau, and that switching to NAS at this point yields further gains, indicating that the two approaches are complementary. Experiments with five LLMs show that seed architecture design is effective across all of them, whereas constructing a usable search space requires a minimum level of model capability. Our code is available at https://github.com/alroimfebruary/AgentNAS.

cs.AI

新论文 · v1arXiv:2611.20579

Language Models Meet Vibration Testing for LEO Constellation Satellites with Calibrated Response Prediction from Few-show Sweeps

Jun Wu, Deming Li, Zhenming Zheng, Honghao Yue, Fei Yang, Hao Sun

原文摘要

Large language models (LLMs) increasingly drive engineering analysis end to end, turning textual records into the inputs of numerical programs, but a plausible and wrong value can pass silently into a decision. We propose deterministic validation gates for LLM-driven engineering analysis, a general paradigm in which a typed configuration produced by LLMs is consumed only after it passes deterministic checks on units, physical consistency, verbatim quotes, agreement between independent models, consistency with measurement exports and the applicability of trained models. The gates target hallucinations from two directions, since quote checking catches fabricated content while agreement between models catches errors of omission that quote checking misses. We instantiate the paradigm on sine vibration acceptance testing for low Earth orbit mega-constellation satellites, where DeepSeek V4.1 Flash and Qwen-plus read Chinese test records and extract a configuration of 76 fields. Fact-level F1 ranged from 0.98 to 1.0 on five input variants, with all repairs frozen on development inputs and evaluated on held-out inputs. With all gates, 20 of 20 configurations produced notching results bit for bit identical to the gold run, and all 17 real and injected errors in consumed fields were stopped, of which 16 would have run without gates. Retraining the bound model on a dataset built from an LLM configuration reproduced its results within GPU nondeterminism. We also report a negative result of practical value. Direct LLM curve forecasting only matches linear extrapolation, whereas a small physics-structured network reduces the over-notched bandwidth by 86% to 89% at calibrated coverage c = 0.90, which delineates where direct LLM prediction should give way to a trained model.

cs.CL · cs.AI · cs.LG

新论文 · v1arXiv:2611.20573

Dynamic Shaping: Acquiring Interfaces and Constructing Execution over Frozen Pretrained Components

Nathaniel Oh

原文摘要

We introduce Dynamic Shaping, a method for constructing executable neural arrangements through acquired response and partial-derivative interfaces over frozen pretrained components. The method learns local descriptions of available computation, ranks component rebindings using task-gradient alignment, transports the current interface across a rebinding, and retains changes through behavioral feedback. Continuous interface acquisition uses proximal policy optimization; the source coefficients remain fixed. We derive an exact identity relating transport error in component responses and partials to changes in action logits and their sensitivities. A pixels-only ViZDoom Basic study evaluates first task acquisition without a competent reference controller, demonstrations or intact response targets. Across three initializations and 100 shared held-out game seeds, initial arrangements succeed on 86/300 controller/episode pairs; joint interface-and-binding construction succeeds on 295/300. Fixed-binding interface acquisition also reaches 295/300. Each policy executes four source fragments, with 110 acquired interface values and fixed random visual and action projections. The artifact includes all candidate trials, interventions and costs; all 1,200 recorded initial/retained outcomes replay exactly. The contribution is an explicit profile, rebind, transport and retain procedure and a reproducible reference-free acquisition case over frozen pretrained computation.

cs.AI

新论文 · v1arXiv:2611.20569

ToolGraph as a Q-Function for Test-Time Scaling of Tool-Use Agents

Ke Shen, Ankita Bhaumik, Li Dong, Shengjie Liu, Shuzhen Li, Naveed Ahmed Janvekar

原文摘要

Tool-use agents must chain interdependent calls, yet which outputs satisfy which later inputs in a deployed catalog is not something a general-purpose language model reliably infers from tool descriptions alone. We compile that structure once into a reusable external representation. ToolGraph is a directed, weighted dependency graph over a tool catalog: parameter matching and interventional probes, which execute a candidate chain with and without its proposed predecessor, propose edges whose evidence is combined into an edge confidence. Prior work uses such graphs to retrieve tools, search for tool paths, or score the next call. We ask instead whether the same structure can supply a \emph{multi-step} value by composing graph paths into a goal-conditioned action value \(Q(s,a,g)\). We study three estimators (dynamic programming over entity sets, a GraphSAGE estimator distilled from it, and an LLM scorer conditioned on graph evidence) plus a verbalizer that returns action--goal feedback to the agent, which keeps final action selection. Across two tool-use benchmarks and three backbones, we compare value ranking against personalized-PageRank retrieval over the identical graph, and against direct and reflective agents. Composition pays off only when the estimator also reads the concrete state: the LLM estimator improves on PageRank in every benchmark--backbone setting, by \(8.3\) points on average, while the two purely structural estimators fall below it, because a value defined over tool types cannot separate a correct tool called with a wrong argument. Graph construction is a one-time cost; we report it separately rather than assume it amortized away (at our scale it is comparable to request-time cost) and sweep four knobs controlling how much graph evidence to expose against success and tokens.

cs.AI

新论文 · v1arXiv:2611.20566

Weight-Space Programming Changes How an LLM Uses Python for Computational Biology

Makoto Sato

原文摘要

Weight-space programming allows a language model to edit another model's weights without gradient-based training. We used Claude Opus 5.5 to edit Qwen3.5-0.8B and investigated how the edits changed its use of Python on computational biology problems. We compared four initial edited checkpoints with the original model on 50 test problems across three test seeds, yielding 600 paired evaluations. Strictly correct answers increased from 44 to 77, and technically correct answers, which included correct final values rejected for formatting, increased from 64 to 100. Repeated Python calls decreased from 1,834 to 1,044. Among the 64 pairs in which only the edited response was technically correct, the original model had already computed the correct value in 36 but did not submit a technically correct final answer. The edited model corrected the computation in 13 of the 64 pairs. With six additional checkpoints on the same 50 problems, technically correct answers rose from 96 to 154 across 900 pairs. We explored three selected problems with ten test seeds for each of ten checkpoints. On the melting-temperature problem, correct Python outputs decreased from 60 to 22 while technically correct final answers increased from 0 to 18. Weight editing changed both the calculations Qwen ran and whether it submitted a correct answer.

cs.CL · cs.AI · q-bio.QM

新论文 · v1arXiv:2611.20565

Memory-Efficient Speculative Decoding with Quantized Draft KV Cache

Saksham Gupta, Michael Li

原文摘要

Speculative decoding reduces generation cost, but it stores two KV caches: one for the target model and one for the draft model. We reduce this extra memory by quantizing the draft cache. The target cache and verifier remain in BF16, so that compression cannot change the target model's final correction rule. We first study how to divide precision between keys and values. A naive per-token quantizer makes keys appear much more sensitive: splitting a 12-bit budget, K8V4 beats K4V8 in acceptance by 10.87pp. However, with grouped per-channel keys, the gap becomes -0.25pp. A controlled Qwen test confirms that changing only the key quantizer can erase or reverse the preferred bit split. Hence, following KIVI, we propose K4V4 quantization with per-channel keys and per-token values. Across six model pairs, our method reduces draft cache memory by 66.20%, and combined memory across both caches by 23.95%. Mean acceptance drops from 57.93% to 57.14%, only 1.36% relative. Overall, quantized draft caches can make speculative decoding more memory efficient; as a predictive principle for efficient AI design, key and value precision must be co-designed with the quantizer layout.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20562

One-Shot Calibration via Model Editing

Jiaxing Qiu, Luke Hakso, Roxana Daneshjou, Ahmed Alaa, Thomas Hartvigsen

原文摘要

Model editing aims to correct errors in large, pretrained models by replacing model generations with user-provided responses. Existing editors assume a single target answer for each edit. However, many real tasks inherently entail uncertainty, resulting in multiple plausible answers with varying likelihoods in user feedback. Such feedback forms a target distribution to not only correct but also calibrate model generations. We therefore propose SoftEdit, the first model editor to leverage uncertainty in user responses and perform one-shot calibration during editing. To generalize calibration to unseen samples, SoftEdit uses co-plausibility relationships between answers while preserving real distributional shifts, balancing the new distributional generality-locality tradeoff. Across three language models, three vision-language models, and two large datasets, SoftEdit achieves state-of-the-art performance, ultimately showing that uncertainty in user responses enables generalizable calibration of pretrained models beyond single-answer corrections.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20557

AsymEEF: Bimanual Manipulation with Asymmetric End Effectors

Xiaokang Sun, Ningyue Wang, Kunqi Xu, Guangqi Jiang, Jianglong Ye, Yong Wu, Xueyan Zou

原文摘要

Teleoperated demonstrations are costly to collect and rarely reusable across end-effectors with different kinematics and action spaces. This burden grows in bimanual manipulation, where each end-effector pair may require new data. We present AsymEEF, a framework for executable cross-end-effector demonstration generation and unified bimanual policy learning. Given a small source dataset, AsymEEF transfers demonstrations to heterogeneous combinations of two-, three-, and five-fingered end-effectors without additional teleoperation. Kinematic teacher anchors provide geometric supervision, while sparse real-world X-Calib anchors correct discrepancies in grasp execution. A shared encoder maps wrist poses and actuator states to a morphology-agnostic latent space, and a morphology-conditioned decoder reconstructs target commands. Generated trajectories undergo replay filtering for task-level executability. The representation also supports a unified bimanual visuomotor policy trained across multiple pairs. Across four simulated and six real-world tasks, AsymEEF improves replay success over adapted transfer baselines, achieves a mean replay-valid expansion gain of 16.6 times in simulation, and is evaluated through downstream policy training on real robots. Project website: https://asymeef.vercel.app/.

cs.RO · cs.AI · cs.CV · cs.LG

版本更新 · v2arXiv:2603.19615

CAF-Score: Calibrating CLAP with LALMs for Reference-free Audio-Text Alignment Evaluation

Insung Lee, Taeyoung Jeong, HaeJun Yoo, Du-Seong Chang, Myoung-Wan Koo

原文摘要

While Large Audio-Language Models (LALMs) have advanced audio understanding and generation, robustly evaluating audio-text alignment remains difficult: reference-based metrics require costly annotations, while Contrastive Language-Audio Pretraining (CLAP)-based reference-free metrics frequently overlook syntactic errors and fine-grained details. We propose CAF-Score, a reference-free metric that calibrates CLAP's coarse-grained semantic alignment with the fine-grained comprehension and syntactic awareness of LALMs. Because CAF-Score scores an audio-text pair symmetrically, one configuration applies to both alignment directions: audio captioning evaluation (the BRACE benchmark) and text-to-audio (TTA) generation evaluation (the RELATE and PAM benchmarks), with a single fusion weight selected once on BRACE and transferred without re-tuning. Across all three human-annotated benchmarks, CAF-Score improves agreement with human judgments over standalone CLAP in the large majority of CLAP-LALM combinations (and in all combinations with the strongest LALM backbone), and outperforms standalone LALM metrics on the main captioning benchmark. It also remains competitive with reference-based metrics in challenging scenarios.

cs.SD · cs.AI · cs.CL

版本更新 · v2arXiv:2609.32083

When Should a Human Take Back Control? Optimal Delegation under Turbulent AI Risk

Haoze Yan, Julien Roze, Ved Upadhyay, Unal Tatar, Thibaut Mastrolia

原文摘要

When should a human take control from an autonomous AI agent, how much checking is worthwhile, and when should autonomy resume? These decisions become harder when incidents cluster and recent errors signal elevated risk. Exploratory analyses of AgentLogs coding-agent traces and ODCV-Bench task runs show higher held-out log likelihood for Hawkes incident-history models than matched no-history baselines, with positive gains whose conditional $90\%$ intervals exclude zero. AgentLogs evaluates continuous-time incident timing and type; ODCV evaluates failure occurrence at discrete decisions. Motivated by this predictive evidence, we formulate a continuous-time delegation problem that jointly chooses productive action, verification effort, and switching rates between human-directed, AI-assisted operation and autonomous AI operation. The model assumes that verification weakens incident propagation and balances project benefits against operating, supervision, risk, and handoff costs. Our Hawkes-PPO method summarizes observed incidents with exponential memory filters and learns these decisions from rewards. In synthetic experiments, it achieves a higher evaluated objective than either fixed regime and PPO without history filters, and approaches an approximate full-information benchmark. Further experiments show how incident excitation and supervision costs shape delegation.

cs.LG · cs.AI · math.OC

新论文 · v1arXiv:2611.20551

GranularGPT: Natural-Language-Driven End-to-End Discrete Element Method Simulations

Erlei Li, Sirui Tang, Ai Wang, Mingwei Hu, Yuchen Dai, Geoff Wang

原文摘要

Discrete element method (DEM) simulations are widely used to investigate granular and particulate systems, but their application often requires specialized expertise in particle physics, numerical modelling, and simulation post-processing. Large language models (LLMs) offer new opportunities to simplify these workflows by enabling users to interact with simulation tools through natural language. This study offers GranularGPT, an LLM-assisted agentic framework for end-to-end DEM simulation workflows, from natural-language problem descriptions to simulation execution and automated post-processing. The framework enables users to generate DEM simulation cases, execute and monitor simulations, and automatically extract and report physical and process-level metrics. Evaluation across 60 normal-condition cases covering 12 DEM scenarios and parameter variations achieve a simulation success rate of 91.7%. In the error-handling evaluation, all 10 repairable errors are detected and 9 are successfully resolved, while all 5 non-repairable cases are correctly identified and reported. The framework automatically extracts a range of DEM-relevant metrics, including particle size distributions, packing and density measures, velocity profiles, discharge rates, mixing indices, contact force characteristics, and thermal quantities. GranularGPT provides a foundation for improving the accessibility and efficiency of DEM simulation workflows while reducing the specialized expertise required for their operation.

physics.comp-ph · cs.AI

版本更新 · v4arXiv:2606.10380

CRADLE-Dialogue: Expert-Level Crisis Detection in Mental Health Conversations

Grace Byun, Abigail Lott, Rebecca Lipschutz, Sean T. Minton, Elizabeth A. Stinson, Jinho D. Choi

原文摘要

Real-world crisis intervention is inherently conversational, yet existing research largely focuses on static texts. When applied to multi-turn dialogues, current models exhibit significant performance degradation, struggling to track risk signals that emerge as context evolves. To address this gap, we introduce CRADLE-Dialogue, a clinician-annotated benchmark for turn-level crisis detection in conversational settings. The dataset features 600 dialogues with multi-label annotations across clinically grounded risks, including suicide ideation, self-harm, and child abuse, distinguishing past from ongoing risk. We further propose an Alert-Confirm evaluation protocol that distinguishes early warning signals (Alert) from turns where a specific crisis becomes explicitly identifiable (Confirm), reflecting the clinical need to intervene before risk becomes explicit. Experiments show that identifying when risk emerges is much harder than recognizing that it exists: models achieve only mid-40% to high-60% Micro F1. Additionally, we release a synthetic training corpus and a 32B-parameter model that substantially outperforms existing open-source models and achieves competitive or superior results against proprietary models across turn-level, dialogue-level, and confirm-only evaluation settings. We further validate our model on real patient interviews, demonstrating its ability to generalize beyond synthetic conversations.

cs.CL · cs.AI

新论文 · v1arXiv:2611.20535

VAHBench: A Harness Benchmark for Agentic Long-Video Generation

Hongbo Zhang, Liuyang Song, Hang Zhou, Xinchen Liu, Wei He, Kai Han, Yunhe Wang

原文摘要

AI agents are commonly used for everyday information retrieval and coding. We ask whether they can also interpret human requests and turn them into complete videos. We introduce VAHBench, a benchmark that tests this ability through five short-drama briefs specifying story, characters, dialogue and duration. Agents plan and carry out production using shared generation tools. Evaluation combines video completeness and basic quality with checks of production reliability and reporting accuracy. Across 240 runs covering 48 configurations of general-purpose and video-specific agents, higher quality scores do not necessarily mean more complete delivery. Missing episodes, repeated footage and static filler reveal gaps between producing a video file and fulfilling the brief. Added composition tools help some agents but do not consistently improve performance. VAHBench measures how well agents translate user intent into finished video content and helps identify where that process breaks down.

cs.AI

新论文 · v1arXiv:2611.20521

LINK: Lightweight and Inductive Node-to-Node Propagation Rule Learning for Traffic Kriging

You Shang, Yu-Kwan Ching, Weifeng Su, Guanyao Li, Ruiyuan Li, Gary Shueng-Han Chan

原文摘要

Spatio-temporal traffic kriging aims to reconstruct traffic time series at unobserved locations with no historical traffic observations by leveraging spatial and temporal dependencies across the road network. Existing Graph Neural Network (GNN)-based approaches primarily rely on neighborhood aggregation to capture spatial dependencies, thereby modeling traffic interactions implicitly and making it difficult to accurately recover traffic magnitudes at locations without historical observations. To address this limitation, we propose LINK, a Lightweight and Inductive Node-to-node propagation rule learning paradigm for traffic Kriging. LINK introduces three key designs: 1) a propagation-centric traffic kriging paradigm that explicitly learns a shared node-to-node transition rule; 2) a context-aware Spectral Propagation Network that performs traffic-state propagation in both the spectral and time domains; and 3) TraffiKrig, a realistic traffic kriging benchmark that preserves heterogeneous road-segment characteristics and road-specific metadata while enabling systematic comparison of representative kriging methods under a unified metadata-enhanced setting. Extensive experiments demonstrate that LINK achieves strong reconstruction accuracy and computational efficiency. Compared with the strongest baseline on each subset, LINK achieves an average MAE reduction of 54.7% over original baselines. LINK also achieves substantial efficiency improvements, reducing training time by 66.5% and inference time by 86.6% compared with the fastest competing methods. Zero-shot experiments further demonstrate the transferability of the learned propagation rule across regions.

cs.LG · cs.AI

版本更新 · v2arXiv:2609.22308

GameReplica: A Benchmark for Black-Box Visual Game Replication by Vision-Language Agents

Boyu Qiao, Zixin Tang, Kaige Li, Xiaoshuai Hao, Dongyu She, Xiaochun Cao

原文摘要

Benchmarks for coding agents typically assess implementation from explicit specifications, leaving open whether agents can discover and reproduce systems whose rules are undocumented and accessible only through interaction. We study this capability in web games, which offer controlled environments with observable actions and outcomes, and introduce GameReplica, a benchmark for end-to-end black-box game replication. It comprises 125 instances spanning 25 web games and five core mechanism families, with each game instantiated at five difficulty levels. Under container isolation, an agent accesses the target game only through an action--screenshot interface, may iterate between probing the game, implementing a replica, and checking it, and delivers a self-contained runnable replica together with a rule checklist. A reference-grounded evaluation then scores the deliverables separately on rule consistency, implementation consistency, and visual fidelity. Experiments across six models and five agent runtimes show that black-box replication remains challenging: Claude Opus 4.8 achieves the highest Overall score, 71.6%, while the other models obtain 4.0%--43.0%. Every model scores highest on visual fidelity and lowest on rule consistency, suggesting that reproducing appearance is easier than recovering mechanics. Difficulty widens the gap: from L1 to L5, the four lower-scoring models show relative Overall score declines of 33%--63%, whereas the strongest model declines by less than 5%. In the evaluated configurations, outcomes vary more across models than across agent runtimes, and most attributed failures of weaker models arise during exploration, before replica code is written. Resources and interactive demonstrations are available at https://qqqqqqby.github.io/GameReplica/.

cs.CV · cs.AI

版本更新 · v2arXiv:2608.11229

ToM-2 Understanding Statements for Teacher-Guided Preference Learning

Jack Mirenzi, Henny Admoni

原文摘要

Preference learning is a powerful tool for reward learning, but treats the teacher as a passive oracle that answers queries driven by the learner. This learner-guided query selection ignores the teacher's knowledge of the solution. We reverse the typical preference query selection method by having the teacher create ordered preference pairs, since a teacher who knows the target can choose more informative pairs than a learner choosing its own queries. Exploiting this advantage requires the teacher to model what the learner knows, and that model goes stale when the learner's belief changes in ways the teacher cannot see, as when several teachers each teach without knowing what the others taught. We call this teacher-model drift. In this work, we use second-order Theory-of-Mind (ToM-2) to have the learner create understanding statements that correct the teacher's model of the learner's knowledge. In simulation, we show that a learner with a perfect teacher, one with perfect knowledge of the learner, needs about half as many rounds as learner-guided querying to learn the reward. Yet when teacher-model drift is induced by alternating between four co-teachers, the advantage erodes. We tested two types of understanding statement generation: the learner's mean belief and ToM-2 drift minimization. ToM-2 statements reduced the drift and recovered more of the lost advantage than mean statements. In a within-subjects study (n = 101) with a simulated ideal co-teacher, ToM-2 statements raised the learner's final alignment with the true reward above both uninformed and mean-informed teaching; on the regret participants removed in their own rounds, they outperformed uninformed teaching but not mean statements.

cs.AI · cs.RO

版本更新 · v4arXiv:2602.02343

Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics

Ziwen Xu, Chenyan Wu, Hengyu Sun, Haiwen Hong, Mengru Wang, Yunzhi Yao, Longtao Huang, Hui Xue 等 12 位作者

原文摘要

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

cs.CL · cs.AI · cs.CV · cs.IR · cs.LG

版本更新 · v3arXiv:2603.02578

How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities

Ziwen Xu, Kewei Xu, Haoming Xu, Haiwen Hong, Longtao Huang, Hui Xue, Ningyu Zhang, Yongliang Shen 等 11 位作者

原文摘要

Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.

cs.CL · cs.AI · cs.HC · cs.LG

版本更新 · v2arXiv:2611.18464

Typing Matters: Muscle, Language and Typing Behaviour in Live Keystroke Decoding from sEMG

Abhilash Durgam, Nyle Siddiqui, Mubarak Shah, Qiushi Fu

原文摘要

Surface electromyography (sEMG) from the wrists can recover typed text without a keyboard. It is usually posed as a signal-inference problem: characters are read out of the muscle signal and character error rate is reported offline. We argue that this framing removes most of the problem. The decoder is constrained simultaneously by the muscles, by the language being typed, and by typing behaviour such as typos, corrections, backspaces and line breaks, and a usable system must predict keys live while the hand keeps moving. Existing systems fall short on two counts: (1) they process sEMG in fixed windows and carry nothing from one window to the next, discarding context accumulated over a session, and (2) they lack any understanding of the linguistic structure that drives typing, so every keystroke is inferred from the muscle signal alone. We propose a fully causal, zero-lookahead sEMG typing decoder built on a state space model (SSM), in which a single model decodes a whole session without windowed partitions, conditioned on what has already been typed. We achieve this by fusing a traditional typing decoder with an efficient autoregressive language model, instilling both sEMG and linguistic understanding in one model. To train such a model for per-character typing rather than whole-word text, we introduce a retokenization method that rewrites ordinary prose as human-like typing, with keystrokes, typos, backspaces and line breaks included, converting WikiText into a large typing corpus where none existed. Our model, having never seen a single sEMG-to-character transcript in typing, achieves a state-of-the-art character error rate of 25.13% on emg2qwerty's cross-user benchmark, a 5.4% relative improvement over the previous best with 6.7 times fewer parameters, while running strictly causally with no future context.

cs.AI

版本更新 · v3arXiv:2604.14655

AgentGA: Evolving Code Solutions in Agent-Seed Space

David Y. Y. Tan, Kellie Chin, Jingxian Zhang

原文摘要

We present AgentGA, a framework that evolves autonomous code-generation runs by optimizing the agent seed: the task prompt plus optional parent archives that initialize a fresh workspace. The outer loop searches over reusable starting conditions rather than editing code directly. Each generation launches a fresh autonomous run in an isolated workspace, while selected parent archives provide inherited artifacts that descendants can inspect and reuse. AgentGA couples a population-level genetic algorithm with long-horizon agents, deterministic 1:1 elite tournaments, and online operator allocation through a modified Hedge controller. We evaluate the approach for tabular AutoML on the 16-competition Weco-Kaggle Lite benchmark, including model-matched AIDE and AutoGluon comparisons. AgentGA with DeepSeek V4 Flash (0731) achieves 71.13% mean Exceeds % of Human, compared with 57.61% for model-matched AIDE and 57.89% for AutoGluon, outranking AIDE on 14/16 tasks. A control that removes inheritance, in which every agent starts from scratch, reaches 65.76% under AgentGA's stopping rule and 67.25% at the full budget, above a single long-lived agent (61.20%) and below AgentGA. These descriptive results support agent-seed optimization as a practical design choice for autonomous code-search systems, without isolating the causal contribution of inheritance.

cs.AI · cs.LG

版本更新 · v4arXiv:2505.17760

But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors

Leon Eshuijs, Archie Chaudhury, Alan McBeth, Ethan Nguyen

原文摘要

LLM-as-a-judge is widely used as a scalable substitute for human evaluation, yet judges are least reliable on subtle dishonesty such as sycophancy and manipulation, where a judge scoring a response in isolation lacks a reference point. We introduce Judge Using Safety-Steered Alternatives (JUSSA), a framework that reduces this evaluator uncertainty using the assessed model's own internal representations. An honesty-promoting steering vector, optimized from a single training example, generates a contrastive alternative against which each response is judged. On a new manipulation dataset with human-validated response pairs, JUSSA improves AUROC for both GPT-4.1 (0.893$\rightarrow$0.946) and Claude Haiku (0.859$\rightarrow$0.929) judges. The gains concentrate almost entirely on the pairs a judge would otherwise misrank. They reverse once a judge has saturated a task and for the weakest judge we test, so contrastive evaluation helps most when a judge is capable enough to use the reference but has not yet saturated the task. A layer-wise analysis further locates steering effectiveness in the middle layers, where model representations of honest and dishonest prompt processing diverge. Our work demonstrates that steering vectors can serve as tools for evaluation rather than for improving model outputs at inference, opening a new direction for white-box auditing.

cs.LG · cs.AI

版本更新 · v2arXiv:2610.05162

Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy

Ruqing Ning, Haibo Meng, Zhishang Xiang, Zerui Chen, Jinsong Su, Xin Wang, Qinggang Zhang

原文摘要

Long-term memory enables LLM-based agents to retain and reuse information across tasks and sessions, supporting personalization and long-horizon interactions. However, persistent memories can also induce sycophancy, causing agents to over-align with users' historical beliefs even when they are inaccurate, outdated, or inconsistent with objective evidence. Existing mitigation methods assume that memory-induced sycophancy originates from biased or incorrect memories and attempt to reduce this risk by filtering such memories at different stages of the memory pipeline. However, in the real world, objective and correct memories can still induce sycophancy, and the same memory can warrant different influence across different contexts. To this end, we propose MemAdapter, a novel framework that adaptively integrates retrieved memories to support objective and reliable reasoning. Specifically, MemAdapter consists of three components: (i) Counterfactual Induction, which leverages counterfactual reasoning to uncover the potential risk of retrieved memories; (ii) Context-Aware Reflection, which calibrates the inferential influence of each retrieved memory in light of the current task via self-reflection; and (iii) Evidence-Based Reasoning, which grounds the final response in appropriate evidence while preserving the legitimate influence of memory. Extensive experiments on three benchmarks demonstrate that MemAdapter consistently improves memory reliability across diverse scenarios. Our code is available at https://github.com/DEEP-JLU/MemAdapter.

cs.AI

版本更新 · v4arXiv:2610.21390

PriorityFEC: Deadline-Aware Streaming FEC for KV-Cache Handoff in Edge Multi-Agent Systems

Han Shen, Qifu Tyler Sun, Fanyang Meng, Ye Wang, Yongsheng Liang

原文摘要

Collaborative edge large language model (LLM) systems often require repeated prefill computation when task contexts are transferred between agents. Reusing previously computed key--value (KV) caches can eliminate this overhead, but requires KV-cache delivery before the receiving agent begins decoding. Under strict deadlines and unreliable edge links, complete KV cache transfer may be infeasible, making the timely delivery of useful KV information critical. This paper proposes PriorityFEC, a utility-aware KV-cache handoff scheme that combines prioritized source transmission with adaptive loss recovery. PriorityFEC first organizes KV blocks into a utility-aware source queue to prioritize important information. It then employs ACK-sliding Streaming FEC and a lightweight Double Deep Q-Network (Double-DQN) scheduler to dynamically balance source delivery and repair transmission under deadline constraints. We implement PriorityFEC in a distributed edge LLM-serving prototype and evaluate it using Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct across multiple workloads. Experimental results demonstrate that PriorityFEC preserves more useful KV information and achieves substantially higher task accuracy than competing transport schemes, particularly when the deadline is insufficient to deliver the complete KV cache.

cs.NI · cs.AI

版本更新 · v3arXiv:2607.26643

SkillBoost: A Constrained Exploration-Exploitation Process for Mitigating Skill Overfitting

Hongqiang Lin, Chao Liu, Xiaofan Bai, Xuan Jin, Yuhong Li, Nenggan Zheng, Xipeng Cao

原文摘要

Enabling large language model (LLM) agents to accumulate and reuse experience from past interactions remains a central challenge in real-world applications. A promising solution is to treat skills as trainable states and optimize them in the same way as model parameters in neural network training. However, data-driven skill optimization is prone to overfitting to the limited trajectories collected from real environments. Overexploiting these trajectories overfits the current batch, while unconstrained exploration causes regression on previously solved cases. This tension motivates a constrained search view of skill self-evolution, governed by an exploration--exploitation trade-off. We propose SkillBoost, a three-stage framework that mitigates both risks: structured exploitation localizes observed failures to editable skill components, prior-guided exploration draws on prior knowledge in the LLM to generate diverse repair candidates, and verified acceptance commits a candidate only when it improves performance within a regression bound. Experiments across 23 model--benchmark configurations show that SkillBoost achieves state-of-the-art performance while mitigating overfitting, outperforming both human-crafted and LLM-generated skills. Transfer experiments further show that optimized skills can be reused by other agents on similar tasks. The code is available at https://github.com/HQ-Lin/SkillBoost

cs.AI · cs.LG

新论文 · v1arXiv:2611.20488

EVIDENT: Evaluating Evidence Grounding in GUI Agents

Yerin Nam, Minkyung Cho, MinKyu Kim, SeungWoo Song, Hangyeol Yoo, Jihyo Kim, KyungTae Lim

原文摘要

GUI agents are increasingly evaluated by whether they complete web tasks, but surface success can hide a critical failure: the final answer may not be supported by evidence observed during navigation. We introduce EVIDENT, an evidence-driven benchmark for evaluating whether GUI agents ground their decisions in observable web evidence. EVIDENT contains 220 real-world English and Korean tasks spanning semantic grounding, contextual inference, and visual grounding, each paired with human-curated evidence metadata specifying what must be verified. We further separate surface task completion from evidence-grounded success using three metrics: TSR, ESR, and GSR, which respectively measure outcome success, required evidence observation, and claim-level evidence support. Experiments on commercial and open-source GUI agents reveal substantial gaps between surface success and evidence-grounded success, especially in contextual and visual tasks. These results show that current evaluations can overestimate agent reliability and highlight the need to assess not only whether agents answer, but whether they verify.

cs.AI · cs.HC

新论文 · v1arXiv:2611.20484

Agent-Guided Configuration Tuning for SMT Solvers

Cong Wang, Qiaoyi Zheng, Zhaoguo Wang, Fan Long

原文摘要

SMT solvers expose many configuration choices that determine how well their existing algorithms perform on a workload. We present, to our knowledge, the first workflow that uses general-purpose LLM agents to tune SMT solver configurations. Agents read solver documentation, source code, and persistent experiment records to propose option changes, and a deterministic harness measures each proposal, validates the returned models, and selects changes separately for each competition Division. Over roughly nine weeks, our tuning campaign recorded 4,545 proposals for cvc5 across 13 Divisions and 36 logics. We evaluate the tuned configurations on the 2026 and 2025 SMT-COMP ModelValidation benchmarks under a 24-second cutoff. On problems not used for training, they validate more problems than official cvc5 in ten of the 13 Divisions in 2026 and in eight in 2025, with geometric-mean gains of 7.8% and 6.8% across Divisions. On the complete official selections, which include training problems, the tuned floating-point counts exceed those of every individual competition participant in both years. Further studies show how configuration changes make existing algorithms effective by changing access to strategies, scheduling, and option compatibility, and how benchmark families trade off against one another. The agents' experiment traces contain hypothesis-based proposals in every tuning group, and 64.3% of the extracted hypotheses draw on prior measurements. These results show that agent-guided tuning can specialize a mature SMT solver and that accumulated experimental evidence informs sustained search.

cs.AI

版本更新 · v2arXiv:2609.33822

Vestrum: Learning the Verification, Retrieval and Structure of Agent Harnesses from a Few Long Episodes

Jayant Parashar, Eugene F. Douglass, William C. Bastian, Suchendra M. Bhandarkar

原文摘要

Agent harnesses decide what the model retrieves, decomposes, remembers and verifies, and most people use harnesses that are pre-built and fixed, such as Claude Code or a published memory system. Harness optimizers rewrite their own harness code rather than editing pre-built systems, and they learn from tens to hundreds of tasks, while long-horizon work offers a handful of episodes, each hundreds of tool calls long. We present Vestrum, which improves existing harnesses component by component. It reads full raw traces, states each failure as a recognizable class, and changes the one component where the error arose (a self-verification, a retrieval view, a decomposition step or a memory operation). Each change targets a failure that recurs across tasks of the same kind. Before anything runs, a generalization filter refuses changes that copy facts from the development tasks. Vestrum learns from one to ten long development episodes and improves tasks it has never seen in all five settings. On Claude Code, Vestrum raises Terminal-Bench 4 Hard from 63.7% to 70.3% of test checks at 1.03x the test cost. It raises expert cell-type agreement from 67.5% to 77.8% on held-out sections of one slide. On the GAM memory system, it raises UltraHorizon from 47.6 to 59.8, AMA-Bench from 64.2% to 74.2% and LoCoMo from 71.5% to 74.9%. It beats the GEPA prompt search we ran on the same data in every paired evaluation, and these gains hold under a test model different from the one Vestrum's search used. Across the searches, every self-verification the loop kept acts early: it checks an intermediate result against recorded evidence before the result is committed, or, on Terminal-Bench, requires an invariant and a probe before the first edit. In development, auditors that re-examined every finished answer broke more correct answers than they fixed, and the loop did not keep them.

cs.AI

新论文 · v1arXiv:2611.20478

Beyond Seeing: Fine-Grained Preference Validation and Severity-Aware Alignment for Mitigating Visual Hallucinations in Vision-Language Models

Abdulaziz Alqahtani, Alwaleed Alharthi, Muzammil Behzad

原文摘要

Large vision-language models can produce fluent responses containing objects, attributes, or relations unsupported by the image. We study an implemented alignment pipeline that normalizes released fine-grained hallucination supervision, validates chosen-rejected preference pairs with vision-language judges, repairs rejected chosen responses, and trains LLaVA adapters with standard or severity-aware direct preference optimization. The study uses 16,143 normalized feedback records and 8,386 released preference pairs, comparing direct training with zero-shot and two-shot judged severity-margin variants. The zero-shot variant reduces Object HalBench CHAIRs from 53.00 to 36.00 and CHAIRi from 15.72 to 10.41. Across aligned models, AMBER-Hal decreases from 35.90 to 25.10-26.90, accompanied by reduced visual coverage. In contrast, the base model retains the highest valid POPE adversarial F1, 84.18 versus 83.87-83.91 after alignment. These results support benchmark-dependent reductions in generative hallucination rather than uniform superiority of severity-aware alignment. We distinguish implemented pipeline components from the preference paths used in the reported experiments and discuss the limits of attributing gains to judging, repair, or the training objective individually.

cs.CV · cs.AI

新论文 · v1arXiv:2611.20476

Improving X-Ray Baggage Instance Segmentation with Shared Point-Based Mask Refinement

Abdulaziz Alqahtani, Qasim Umer

原文摘要

Overlapping projections and complex object contours make prohibited items difficult to delineate in X-ray baggage images. This study investigates shared point-based mask refinement to improve instance segmentation within Cascade Mask R-CNN. The framework combines a ResNet-101 backbone with an attention-based feature pyramid and progressively updates uncertain mask locations, increasing mask resolution from $28 \times 28$ to $112 \times 112$. A single refinement module is shared across cascade stages and refinement levels, with transformer and multilayer perceptron (MLP) variants used to examine the role of interactions among selected points. Ten configurations were evaluated on the official PIXray split across three training seeds using the original target-instance annotations for evaluation. The transformer and MLP refiners achieved mean mask average precision (AP) of 71.79 and 72.47, respectively, compared with 70.50 for the unrefined detector, yielding gains of 1.29 and 1.97 AP points. The MLP achieved the highest mean mask AP among the evaluated configurations. Both variants also improved boundary agreement on commonly detected objects, supporting improved mask delineation, while detection diagnostics did not establish a corresponding positive gain. These findings demonstrate that shared point-based refinement can improve segmentation quality in the tested cascade architecture using existing target-mask annotations, without additional occluder supervision. Evaluation on independent X-ray datasets is needed to determine how well these improvements generalize.

cs.CV · cs.AI

新论文 · v1arXiv:2611.20472

Model Immunization Beyond Condition Numbers: The Importance of Plateau Regions

Quoc Nguyen, Trung Le, Dinh Phung, Mehrtash Harandi

原文摘要

Harmful fine-tuning can rapidly compromise safety alignment in Large Language Models, underscoring the vulnerability of deployed models to post-release modification. While prior work motivates model immunization by inducing ill-conditioned curvature to slow harmful optimization, this approach is not scalable due to the high computational cost of Hessian estimation. We refine this paradigm by extending the theoretical framework to prioritize plateau regions alongside curvature. We further introduce an immunization algorithm based on efficient Hessian estimation and bounds on the extremal singular values of the Hessian. We demonstrate the proposed immunization framework in three applications: (i) Fine-tuning-as-a-service settings where benign user data may be mixed with poisoned harmful data, (ii) open-weight settings where an adversary can fine-tune on harmful-only data with varied training dynamics, and (iii) robust unlearning, where immunized models better resist relearning attacks after unlearning. Extensive experiments confirm that our approach achieves robust defense against adversarial fine-tuning while largely maintaining general utility and downstream trainability. Our code is available at https://github.com/minhquoc0712/BECON.

cs.LG · cs.AI