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

+197新增论文
+151其中 AI 相关
+16已有论文修订
+89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 2026-11-21 00:57 — 2026-11-22 00:57(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 4,321 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:2602.04417

EMA策略梯度:用EMA锚点和Top-k KL驯服LLM强化学习

EMA Policy Gradient: Taming Reinforcement Learning for LLMs with EMA Anchor and Top-k KL

Lunjun Zhang, Jimmy Ba

AI 速读 基于摘要

作者提出用指数移动平均(EMA)锚点替代固定锚点策略,并引入Top-k KL估计器,与GRPO结合后在数学推理和智能体RL任务上取得性能提升。

cs.LG · cs.AI

查看原文摘要

Reinforcement Learning (RL) has enabled Large Language Models (LLMs) to acquire increasingly complex reasoning and agentic behaviors. In this work, we propose two simple techniques to improve policy gradient algorithms for LLMs. First, we replace the fixed anchor policy during RL with an Exponential Moving Average (EMA), similar to a target network in deep Q-learning. Second, we introduce Top-$k$ KL estimator, which allows for flexible interpolation between exact KL and sampled KL. We derive the stability conditions for using EMA anchor; moreover, we show that our Top-$k$ KL estimator yields both unbiased KL values and unbiased gradients at any $k$, while bringing the benefits of exact KL. When combined with GRPO, the two techniques (EMA-PG) lead to a significant performance boost. On math reasoning, it allows R1-distilled Qwen-1.5B to reach 53.9% on OlympiadBench compared to 50.8% by GRPO. On agentic RL domains, with Qwen-3B base, EMA-PG improves GRPO by an average of 35.6% across 7 datasets of Q&A with search engines, including 29.7% $\rightarrow$ 45.8% on HotpotQA, 27.4% $\rightarrow$ 40.5% on 2WikiMultiHopQA. Overall, we show that EMA-PG is a simple, principled, and powerful approach to scaling RL for LLMs. Code: https://github.com/LunjunZhang/ema-pg

版本更新 · v2arXiv:2611.16843

捷径还是记忆?重新审视大语言模型在自然语言理解中的数据损坏实验

Shortcuts or Memorisation? Revisiting Data Corruption in Natural Language Understanding with Large Language Models

Aarne Talman, Stergios Chatzikyriakidis

AI 速读 基于摘要

作者通过扰动实验和污染指标分析,发现大语言模型在熟悉的NLI基准上的鲁棒性反映的是数据集捷径而非记忆,且与预训练暴露无关。

cs.CL

查看原文摘要

Pre-trained encoders fine-tuned on natural language inference (NLI) and other language understanding tasks keep much of their accuracy after the input is perturbed, e.g. by removing all nouns or shuffling the words. Does the same hold for large language models (LLMs), and if so, is it because they have seen the benchmarks during pre-training? In this paper we revisit these perturbation experiments with eight LLMs alongside fine-tuned encoders on GLUE and the three rounds of ANLI. Regardless of model size, every system that answers directly retains 36-72% of its above-chance performance without nouns on MNLI compared to only 14-29% on ANLI. Only a model that reasons before answering departs from this pattern. We also check contamination indicators (including exact $n$-gram counts in the public pre-training data of two of the models) and find no link between exposure and robustness. Our strongest evidence comes from 368 new, human-validated items that none of the models can have seen. All eight LLMs keep more of their performance without nouns on items written in MNLI's style than on items written without its cues. When the perturbation instead removes the function words carrying such cues, this advantage reverses. In light of these results, we argue that robustness to perturbation on familiar NLI benchmarks reflects dataset shortcuts rather than memorisation, and tells us more about the benchmark than about the model.

版本更新 · v2arXiv:2607.15277

划分、提示、聚合:语言模型中的统计自洽性

Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models

Patrik Wolf, Thomas Kleine Buening, Andreas Krause, Celestine Mendler-Dünner

AI 速读 基于摘要

作者通过二叉树递归划分总体并提示LLM估计子群体条件分布,检验其聚合后是否满足全概率公式,发现广泛违反自洽性,并揭示“宏观谬误”现象。

cs.CL

查看原文摘要

Large language models (LLMs) are often used to perform a form of conditional inference. In-context learning is one implicit way to do this. The prompt specifies a context and the model's output is treated as an estimate of the corresponding conditional distribution. If this interpretation holds, then LLM estimates should satisfy basic probabilistic identities. In particular, the law of total probability asserts that prior-weighted conditional distributions aggregate into population-level marginals over any valid partition of the population. In this work, we investigate to what extent LLM estimates adhere to this self-consistency principle. We use binary trees as an evaluation scaffold to recursively partition a population into increasingly fine-grained subpopulations. We then prompt LLMs with verbalized subpopulation descriptions in context, aggregate the resulting estimates back into population-level estimates, and compare them across partitions of varying granularity. Translating these comparisons into a benchmark score and evaluating frontier models across problem domains, we demonstrate widespread violations of basic consistency properties. An in-depth study of persona prompting further reveals a pattern we call the macro fallacy: estimates reconstructed from more fine-grained subpopulation responses are often better aligned with human reference data than direct population-level estimates. The effect persists across variations in partitioning, verbalization, and estimation task, suggesting that models possess relevant subpopulation knowledge but do not reliably propagate it into aggregate estimates. This gap establishes statistical self-consistency as an unsaturated, reference-free criterion for evaluating LLMs.

版本更新 · v4arXiv:2605.03799

自然语言处理:从经典文本处理到多模态与智能体系统的应用研究实践

Natural Language Processing: An Applied Research Practicum From Classical Text Processing to Multimodal and Agentic Systems

Mullosharaf K. Arabov

AI 速读 基于摘要

这是一份以研究为导向的NLP实践指南,覆盖从分词、文本规范化到Transformer、大语言模型、检索增强生成、偏好对齐与多模态系统的完整流程,并特别关注低资源与形态丰富语言。

cs.CL

查看原文摘要

This preprint presents a systematic, research-oriented guide to modern natural language processing, covering the full NLP pipeline from tokenisation, text normalisation, vector representations, and classical machine learning to transformer architectures, large language models, retrieval-augmented generation, preference-based alignment, and multimodal systems. The material comprises eighteen practical works combining theoretical foundations, reproducible implementation examples, detailed experimental procedures, evaluation metrics, and research-oriented assignments. Each work includes three structured core tasks and optional advanced investigations. Common requirements for reporting, data provenance, experimental evaluation, and reproducibility are established across the guide. Particular attention is devoted to low-resource and morphologically rich languages, including Tajik, Tatar, and Bashkir. The guide addresses subword tokenisation, cross-lingual transfer, machine translation, language model adaptation, and the evaluation of NLP systems under limited data and computational resources. Advanced topics include prompt engineering, reasoning strategies, multi-agent collaboration, model compression, efficient inference, AI safety, and LLMOps. A capstone project integrates methods from different chapters into the development and evaluation of domain-specific NLP systems. The guide is intended for senior undergraduate students, graduate students, researchers, and developers interested in implementing and critically evaluating NLP methods.

新论文 · v1arXiv:2611.26553

智能体安全事件的复现:审计与调查(ROGUE-AI)

Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)

Mengxiao Wang, Mahbub Alam, Kalyan Nakka, Nitesh Saxena

AI 速读 基于摘要

提出多智能体框架 ROGUE-AI,将媒体报道的 LLM 智能体安全事故离线复现,并设置仅缺少被归因原因的控制组,以检验事故成因与防御效果。

cs.CR · cs.AI · cs.MA

查看原文摘要

Agents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e.g., Replit's coding agent deleting a production database), but a report often only says what went wrong and rarely (in)validates the incident or finds its root cause. Incident catalogs collect these reports, and benchmarks record whether agents fail, but neither replicates an incident nor tests the cause its report might identify. In this paper, we present ROGUE-AI (Replication of Agentic Security Events: Audit and Investigation), a multi-agent framework that closes the above gap and rebuilds each reported incident offline together with a matched control that lacks only the blamed cause, and runs agents on both. Our results suggest that ROGUE-AI is able to reproduce the reported harm in 89% of its 108 attempted replications. Agents built on five current GPT, Claude, Gemini and DeepSeek models made the reported mistake in 18% to 33% of runs, against at most 1% without the blamed cause. Of the 72 incidents with real-world harm, 62 had no underlying attacker, and in these, agents did what their task asked for but missed an ordinary hazard, such as a database shared with the live site. Replications without an attacker caused more harm than replications of prompt injection for all five models. Defenses against prompt injection look for injected instructions, but these replications contain none. Further, a warning after each tool output that names the hazard and its consequence lowered harm for all five models, whereas a note of the same length that does not mention the hazard left most of the harm in place. Built from real incidents, the ROGUE-AI replications form a reusable audit framework that vendors can rerun before each new release.

新论文 · v1arXiv:2611.26545

谎言的几何:用于检测大语言模型欺骗的知识与人格方向

The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models

Md. Abdullah Al Zami, Suraiya Shanaz, Mahdi Hasan, Shoaib Ahmed Dipu, Farig Yousuf Sadeque, Md. Sabbir Ahmed

AI 速读 基于摘要

该论文探索基于人格方向和知识方向的无监督线性探针,以检测大语言模型在不同类型谎言上的表现,并报告了跨模型和跨谎言类型的迁移效果。

cs.AI

查看原文摘要

Linear probes are the cheapest LLM lie detectors but often fail to transfer across kinds of lie. Label-free directions from honest and deceptive role-play are untested at 70B scale and on lies about knowledge, where a lie and an honest denial can be identical. We widen a published persona bank with roles that deny knowledge they have, plus honest counterparts. At each model's middle layer the resulting persona direction ranks lies across Liars' Bench better than single-dataset probes on three of its four models; it fails on Qwen, even when Qwen stays in role far more often. On Llama-3.3-70B-Instruct it beats directions from a random persona bank, but its gain over the published direction is within the range of random widenings. On lies about knowledge, persona directions mainly track a reply's wording: where lies and honest denials share their words, our three main contrasts stay inside the random-direction band on Llama and Qwen and reach only its edge on Gemma, though on Mistral one clears it. A knowledge direction contrasts questions the model can answer with ones nobody can, each followed by the same scripted denial. It separates lies from identical honest denials on two models and narrowly on a third, but not on the fourth; its other results partly depend on the benchmark's wording. On the Gender-Secret fine-tune it clears the band by a few hundredths on Llama and fails on Mistral. The sum of the two directions ranks lies better than a supervised probe at the same depth on all four models (on one, both near chance) but is level with the persona direction on Llama's average and behind that probe on held-out lies. More varied role-play questions do not help the persona direction on Liars' Bench.

新论文 · v1arXiv:2611.26492

胸部X射线报告中的多模态错误检测:大语言模型读取图像证据

Multimodal Error Detection in Chest X-ray Reports: A Large Language Model Reading Image Evidence

Yanda Cheng, Bo Wen, Yuhao Lu

AI 速读 基于摘要

作者提出Rad-Linter,利用图像分类器分数和微调语言模型检测胸片报告与图像证据的不一致,在受控配对评估中优于规则基线。

cs.CL

查看原文摘要

Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores. Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy. On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-tuning remains near 50% accuracy. On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%. These controlled results show that, given the same image-derived scores, a trained language model detects report-score inconsistencies more accurately than the tested rule baseline, with the advantage concentrated in interpreting how findings are worded.

新论文 · v1arXiv:2611.26422

基于大语言模型推理轨迹分析的可解释Q矩阵构建

Reasoning-Trace Analysis Using Large Language Models for Interpretable Q-Matrix Construction

Jihong Zhang, Zijing Jin, Mingfeng Xue

AI 速读 基于摘要

提出推理轨迹分析(RTA)框架,通过链式思维提示让大语言模型生成属性标注的推理步骤,以重建、比较和检查Q矩阵的每个项目-属性分配,并报告了模拟与实证研究结果。

cs.AI

查看原文摘要

Large language models (LLMs) can generate Q-matrices for diagnostic classification models that approximate expert specifications, but evaluations have focused on the final matrices, leaving the basis of individual item-attribute assignments unexamined. This opacity limits the responsible use of AI in measurement. We propose Reasoning-Trace Analysis (RTA), a framework in which an LLM generates attribute-labeled reasoning steps through chain-of-thought (CoT) prompting and these steps are analyzed to reconstruct, compare, and inspect each assignment. RTA is intended to supplement expert review and empirical validation. In a simulation study with perturbed fraction subtraction items, four LLMs recovered the required attributes and withheld attributes the items did not require. In an empirical study with 11 fraction subtraction items, responses from 536 examinees, four LLMs, up to three temperatures, and 100 replications per condition, CoT-derived Q-matrices overlapped substantially with the reference Q-matrix (mean overlap .809) but less than directly generated Q-matrices (.877), and they varied more across replications. Inspection of the traces linked disagreements to item-type coding, labeling rules, and how operations were expressed as steps. Q-matrices with similar overlap differed in fit to examinee responses. RTA provides inspectable evidence that matrix-level agreement and fit do not.

版本更新 · v3arXiv:2610.09186

模式识别与逐步推理的二分法

The Dichotomy Between Pattern Recognition and Step-by-Step Reasoning

Amrut Nadgir, Vijay Balasubramanian, Pratik Chaudhari

AI 速读 基于摘要

论文提出模式识别与逐步推理是同一谱系的两端,其区分取决于下一个词元所依赖的上下文长度,并证明逐步推理具有样本效率,且状态频率可平衡准确性与鲁棒性。

cs.LG · cond-mat.dis-nn

查看原文摘要

We argue that pattern recognition and step-by-step reasoning are two ends of a spectrum. A large language model (LLM) learns to reason step-by-step when data is structured such that the next token depends on a small amount of preceding context. Inference in LLMs resembles pattern recognition when the next token depends on a large amount of preceding context. If the next token depends on only the $c$ most recent tokens, reasoning traces are paths on a de Bruijn graph whose nodes are $c$-length contexts and edges are next-token transitions between contexts. The set of reasoning traces of a task forms a directed acyclic subgraph of the de Bruijn graph. An LLM that has learned all edges of this subgraph can compose them to solve longer, unseen tasks, i.e., it reasons step-by-step. We prove that the number of edges is vanishingly small compared to the number of reasoning traces. We also prove and show empirically that the number of training samples a transformer needs is a linear in the number of edges, so learning to reason step-by-step is sample efficient. We can induce de Bruijn structure in any task by maintaining a ``state'' that makes future reasoning independent of the past. The frequency of states in the reasoning trace determines $c$. We show, by fine-tuning Qwen2.5-1.5B-Instruct to solve equations and answer questions about stories, that frequent states (small $c$) result in higher accuracy but greater fragility to perturbations at test time. LLMs trained with a large $c$ are only as good as models that perform pattern recognition without reasoning. A moderate density of states balances accuracy and robustness. Real-world data has de Bruijn structure: Qwen3-14B and Qwen3-32B retain over 75\% of their accuracy on GSM8K, MATH-500 and GPQA-Diamond when attention is restricted to a sliding window less than 15\% as long as the full reasoning trace.

新论文 · v1arXiv:2611.26408

为LLM驱动的权重空间编程训练视觉模型

Training Vision Models for LLM-Driven Weight-Space Programming

Timo Ropinski, Sebastian Kranz

AI 速读 基于摘要

作者构建了一个LLM驱动权重空间编程的视觉模型基准,并提出Edit-Ready Training(ERT),通过预演LLM的直接闭式权重编辑来提升编辑成功率。

cs.CV

查看原文摘要

It has been demonstrated that, in certain settings, large language models can perform weight-space programming to directly modify the weights of neural networks. However, it is unknown how dedicated training of these networks can contribute to the successes of weight-space programming. We present a benchmark of LLM-driven weight-space programming on vision models, in which five LLMs edit convolutional networks and vision transformers on relabeling and error-correction tasks on Waterbirds, CIFAR-100, PACS and Entity-30. We report more than 2,600 such edit processes, whereby each edit is scored on held-out images for success, side effects and accuracy loss. We compare standard training with label smoothing, gradient-based editable training, and our Edit-Ready Training (ERT), a novel method that rehearses the direct, closed-form weight edits LLMs perform. We show that preparing networks for editing raises the success of LLM edits substantially where standard networks are hard to edit. ERT further achieves the highest mean score for GPT-6 Astra and Claude Opus 5.5 across all evaluated baselines. We will release our LLM testing framework alongside all tasks, editing trajectories and scores. Project page: https://viscom.uni-ulm.de/wsp

版本更新 · v3arXiv:2512.23889

大型语言模型如何歪曲美国气候舆论

How Large Language Models Misrepresent American Climate Opinions

Sola Kim, Jieshu Wang, Marco A. Janssen, John M. Anderies

AI 速读 基于摘要

作者用978名美国气候舆论调查受访者的个人资料提示六个大语言模型,发现模型生成的回答向人群均值压缩,且性别假设在种族交叉处方向出错。

cs.CY · cs.AI

查看原文摘要

Large language models are increasingly used to simulate public opinion, standing in for survey respondents as "silicon samples." Whether they faithfully represent U.S. climate opinion, particularly at the intersections of race, gender, and ideology, remains largely untested. We prompted six large language models with the profiles of 978 respondents to a nationally representative U.S. climate opinion survey and compared the generated responses with those respondents' own answers across 20 questions. The models compress the diversity of American climate opinion, pulling predictions toward the population mean: less-concerned groups are predicted as more concerned, more-concerned groups as less, and the variance of predicted responses falls, on average, 29% below that of human responses. The compression is intersectional. The models apply one gender assumption, predicting women as more pro-climate than men in every racial group. The assumption matches the direction of the human pattern for White, Hispanic, and Other Americans and runs the wrong way for Black Americans. We claim the direction of that difference rather than its size, since it rests on small cells at the ideological extreme. An audit of gender alone would report that the models understate the gender gap, and would miss that the error changes direction by race, which is what single-variable audits cannot detect. What is at stake is not accuracy alone but representation: whose views a synthetic sample renders legible.

新论文 · v1arXiv:2611.26380

EgoCaptioner: Towards Comprehensive and Fine-Grained Egocentric Video Captioning

Yuqi Tang, Yang Shi, Yuran Wang, Tengfei Liu, Rundong Wang, Zongyang Qiu, Qixun Wang, Zhuoran Zhang 等 17 位作者

原文摘要

Egocentric video captioning requires fine-grained detail and accurate timing, yet current multimodal large language models (MLLMs) often miss brief actions, confuse their order, or misinterpret the wearer's motion. Existing caption formats also offer limited support for combining detailed narration with precise timing. We present EgoCaptioner, which pairs a video-level overview of the setting, participants, and the wearer's overall activity with timed steps whose narration adapts to varying perceptual demands and marks key moments at specific timestamps. We develop a coarse-to-fine annotation pipeline that establishes event structure before expanding step descriptions, yielding 50,000 training examples from eight egocentric video sources. Complementary rewards assess factual accuracy and coverage, measure step-level temporal alignment, and penalize fragmented or redundant narration, guiding caption refinement through Group Relative Policy Optimization (GRPO). To improve the reliability of factual supervision, we supplement caption references with pose-derived motion facts when sensor data are available. Across five egocentric benchmarks, EgoCaptioner achieves leading performance among evaluated open-source models in caption quality, caption-conditioned question answering, and temporal alignment. It also surpasses Gemini 3.1 Pro on PhysCapBench and all evaluated proprietary baselines on Minerva-Ego under the caption-to-QA protocol. Ablations support the complementary roles of factual and temporal rewards and show that coarse-to-fine annotation provides more effective supervision than single-pass annotation.

cs.CV

新论文 · v1arXiv:2611.26363

HiSA: Sparsifying the Hidden Dimension of RoPE Attention in LLMs

Ambroise Borbely, Mathieu Salzmann, Olivier Verscheure

原文摘要

Attention in large language models (LLMs) scores every query against every key with a full dot product. We investigate whether this dot product must be computed completely for attention to be effective. We propose Hidden Sparse Attention (HiSA), which keeps the query and key projections intact but only computes a partial dot product, following a static, parameter-free rule set by the token positions: attention becomes sparse in the hidden (feature) dimension, an axis much less explored than token sparsity, rather than over tokens as in almost all sparse attention methods. Based on the structure of rotary position embeddings (RoPE), we decompose keys and queries into parts which can be omitted when computing attention. By training LLMs from scratch with HiSA, we find that this partial dot product does not hurt but improves performance compared with the complete product while computing only about a third of the query-key products. During inference HiSA improves decoding speed and shrinks memory footprint. We also extend HiSA to FlashAttention-3 kernels for training and inference. Finally, we motivate our approach by showing that dense attention learns to suppress the attention score parts that HiSA drops. Code is available at https://github.com/ambroise-b/hisa.

cs.LG · cs.CL

新论文 · v1arXiv:2611.26349

Perfumer: A Large Language Model-based Agentic Framework Accelerating Structural Relaxation and Electronic Band Prediction for First-Principle Calculation

Zhiwei Li, Yuhao Zhou, Yikai Yang, Haonan Wang, Qing Ye, Jiancheng Lv, Kazunari Matsuda

原文摘要

First-principles structural relaxation and band calculations require substantial computation and expertise. Here we present Perfumer, a large language model-based agentic framework that couples Quantum ESPRESSO with two predictive models. Across tested structures, pre-relaxation with Perfumer-relax model followed by QE refinement yielded a 2.75-fold overall speedup over QE alone. Perfumer-band model combines a periodic graph neural network with a latent Hermitian spectral head returns 12 band energies at any requested wavevector, preserving reciprocal lattice periodicity and time-reversal symmetry without constructing Hamiltonian. Trained on 125,153 self-curated band structures, Perfumer-band model reaches a median per-material mean absolute error of 0.157 eV on 2,607 test materials, including lanthanide, actinide and magnetic compounds, from held-out composition families spanning 85 elements, taking about 97 ms per material on two consumer CPU cores. The agent turns natural-language requests into relaxation-to-band workflows, from a single uploaded structure to automatically constructed new candidates and returns predicted properties.

cond-mat.mtrl-sci · cs.LG

版本更新 · v2arXiv:2611.18714

How to make a moral agent: sentimentalism and rationalism as lenses on AI

David Gottlieb, Jared Moore

原文摘要

Can AI systems be moral agents? That is, can they undertake actions that can be correctly assessed as morally right or wrong? There is no settled normative criterion of moral agency that applies neatly outside paradigm human cases. Because of the difficulties with a normative account, we instead propose a research program to explore descriptive moral agency. Most basically, we ask, what are the capabilities we humans possess that underlie our moral agency? To do so, we draw on historical scholarship in philosophy and moral psychology to isolate candidate capabilities. Then we suggest refining candidate capabilities so that they are testable in the context of AI systems. Finally, we offer specific empirical recommendations to test AI systems, especially large language models (LLMs), for those refined capabilities. Our approach offers several benefits. Firstly, it may demonstrate whether the behaviors of AI systems are aligned, or in a broad sense desirable. Secondly, the debate may inform policy questions (e.g., what models should be deployed in which contexts?). Thirdly, our approach may complement engineering and design efforts: when capabilities can be precisely characterized, they can inform the development of models. Fourthly, applying our moral reasoning concepts in the new context of AI systems should improve our understanding of those concepts.

cs.AI · cs.CL

新论文 · v1arXiv:2611.26342

Training LLMs to Follow Reasoning Instructions with Minimal Edits

Alexandra Yakovleva, Luat Gia Khoi Nguyen, Juho Kannala, Harri Valpola, Alexander Ilin

原文摘要

Large language models increasingly rely on extended reasoning traces, yet the training signals used to shape this behavior remain coarse. Supervised fine-tuning on full trajectories forces the model to imitate an entire demonstration, overconstraining its output distribution, while trajectory-level reinforcement learning provides only weak credit assignment, rewarding or penalizing whole completions without identifying where they succeed or fail. We propose edit-based training, a general framework for learning from a model's own trajectories via localized textual corrections. Given a sampled completion, an editor identifies the first point of failure under a specified criterion and proposes a minimal replacement span. The model is then supervised only on these edited spans. By design, this concentrates supervision where the model's reasoning first breaks down and aims to preserve its existing reasoning patterns elsewhere. The method applies both offline and online, with the latter continually adapting supervision to the model's evolving failure modes. We instantiate this approach on reasoning instruction following with Qwen3-32B on the ReasonIF benchmark. Compared to full-trajectory imitation and trajectory-level reinforcement learning baselines, edit-based training substantially improves compliance with the reasoning instruction while largely preserving the model's skills on untargeted tasks. These results suggest that localized correction is an effective complement to these approaches for shaping how LLMs reason.

cs.AI

新论文 · v1arXiv:2611.26335

ClosureBench: A Benchmark for Reasoning under Closure Assumptions

Francesco Belardinelli, Flavio Corradini, Matteo Leonesi

原文摘要

AI agents built on large language models (LLMs) often answer questions from incomplete records. Missing information might be interpreted as unknown or false, depending on the completeness of the source. Current evaluations fail to distinguish these two aspects, as they are usually interwoven with other reasoning demands. We present ClosureBench, a benchmark that tests whether LLMs follow the stated rules for dealing with missing information. Every base scenario is tested under OWA, CWA, and LCWA, and only the treatment of missing information changes between variants. ClosureBench provides a verified base benchmark as well as agent extensions for action choice, source scope, and updates. A model is considered correct on a base scenario only if it provides correct answers in all of its closure variants. Across 960 instances, overall accuracy can mask substantial closure mistakes. Models achieve overall accuracy from 71.18% to 99.86% across replicated runs, but their core switch accuracy, measured on the scenarios whose gold label actually changes with the contract, varies between 18.23% and 99.31%. Large discrepancies emerge under LCWA. For example, some agents choose unknown despite the prompt indicating that missing information should be treated as false. Our findings demonstrate that evaluations of AI agents should not only measure whether agents can reason over existing facts but also whether they use the appropriate assumption when information is absent.

cs.AI

版本更新 · v2arXiv:2604.03257

Robust LLM Evaluation with Boundedly Bad Judges: A Constrained Maximum Likelihood Approach

Minghe Shen, Ananth Balashankar, Adam Fisch, David Madras, Miguel Rodrigues

原文摘要

Evaluating language models at scale often involves reconciling costly, but reliable, labels (e.g., from human raters) with cheap, but noisy, proxy labels (e.g., from LLM-as-a-judge autoraters). This paper proposes constrained maximum likelihood estimation (CMLE), a likelihood-based approach to large language model (LLM) evaluation that explicitly models the error in the proxy labels via their true-positive and false-positive rates, subject to data-driven constraints. Functionally, these constraints act as an adaptive regularization that restricts the feasible judge error rates, particularly when the available number of reliable labels is limited. In contrast to asymptotically unbiased estimators (such as those based on prediction-powered inference), this formulation accepts a controlled bias under misspecified constraints in exchange for variance reduction relative to various baselines including unconstrained maximum likelihood estimation (UMLE); we characterize this bias--variance tradeoff analytically under point constraints. We identify the regimes in which CMLE outperforms UMLE and prediction-powered baselines in mean squared error. We further extend the framework to heterogeneous settings using a stratified variant involving stratum-specific constraints based on judge confidence. Empirically, on synthetic and real-world LLM evaluation tasks, CMLE achieves up to roughly 9\% lower mean squared error (MSE) than non-stratified baselines, and its stratified variant (SCMLE) achieves up to roughly 30\% lower MSE than the strongest baseline considered, with the largest gains in low-label regimes.

cs.CL · cs.AI

新论文 · v1arXiv:2611.26333

OceanCoder: A Verification-Guided Large Language Model for Automated Simulation Modelling in Ocean Engineering

Can Xue, Changhong He, Ziyi Sun, Anliang Ge, Shimin Yu, Bingchen Liang

原文摘要

Numerical simulation is important in ocean engineering, but model construction requires expert knowledge. Large language models support automated modelling, but code syntax and successful execution alone do not establish whether generated models and results meet engineering requirements. We present OceanCoder, a Qwen3-8B-based large language model for automated simulation modelling in ocean engineering. To address specialised programming requirements and scarce training data, we develop an extensible data construction pipeline producing 12,000 verified supervised fine-tuning (SFT) examples across 12 engineering task categories. A three-layer verification framework checks model inputs against engineering parameters and task constraints, verifies completion of required simulation analyses, and assesses the engineering plausibility of computed outputs. Its outcomes provide graded rewards for group relative policy optimisation (GRPO), guiding the model towards engineering task completion. We also construct the first benchmark for automated modelling in ocean engineering, covering out-of-distribution tasks and real-world engineering cases. Experiments and ablation studies demonstrate the effectiveness of the data pipeline, verification framework and reward design. OceanCoder outperforms both base models and the closed-source comparison, with Pass@1 increasing from 0% for the base model to 46.77% after SFT and 51.43% after GRPO. These results support engineering verification as a training signal for automated simulation modelling.

cs.LG · cs.SE

版本更新 · v2arXiv:2610.21062

SnapJudge: Multimodal AI Detection for Math-Heavy Scientific Text

Yixuan Huang, Danish Pruthi, Bhuwan Dhingra

原文摘要

As large language models become increasingly capable of generating, and assisting with, scientific writing, there is a need for greater transparency and disclosure around the use of AI in research. Consequently, a veritable subfield of research and a number of commercial services aim to flag AI-generated content automatically. However, existing detectors which rely only on text extracted from PDFs fail to benefit from the rich visual information contained in the formatting of math-heavy scientific content. To study this, we select human-written proof and prose passages from pre-2022 arXiv LaTeX sources and re-generate them using 15 modern LLMs within the context of the same paper. This leaves us with coupled human and AI outputs in three different representations: the LaTeX source, a rendered image of the PDF, and the text extracted from the PDF. We find that the accuracy of the state-of-the-art Pangram 4 drops significantly as the mathematical content in a passage increases, especially when evaluated on text as opposed to LaTeX. However, this is not an inherent limitation of detecting AI-generated math: finetuning Qwen-3.5-27B based detectors on this domain recovers the gap, with LaTeX inputs performing significantly better than text since they preserve more formatting information. Since LaTeX sources are rarely available in practice, we also train SnapJudge: an image-based detector operating on screenshots of both prose and proofs from papers. At 0.1% false positive rate, SnapJudge achieves a detection rate of 92% on proofs and 93% on prose generated by frontier models, substantially outperforming both Pangram 4 and our text-based detector. We release our data, models and code at https://nlplab.cs.duke.edu/snapjudge.

cs.CL

版本更新 · v3arXiv:2606.00384

VESTA: Visual Exploration with Statistical Tool Agents

William Rudman, Abhishek Divekar, Kanishk Jain, Sebastian Joseph, Stella S. R. Offner, Matthew Lease, Kyle Mahowald, Greg Durrett 等 9 位作者

原文摘要

Fitting quantitative models to data is a central step in scientific workflows, yet it remains one of the least automated. Recent agent-based systems leverage language and vision-language models (VLMs) to propose statistical models, but these systems struggle on challenging modeling tasks. To address these limitations, we introduce VESTA (Visual Exploration with Statistical Tool Agents), a framework that equips VLMs with a dynamically growing toolkit to guide model refinement through data transformations, hypothesis-driven visualizations, and statistical tests. Unlike prior systems that rely on iterative critique alone, VESTA explores data before and during refinement by selecting or creating reusable diagnostic tools. We evaluate VESTA using three toolkit configurations: no tools, static expert-written tools, and dynamic model-written tools. To support this evaluation, we introduce DAWN (Dataset for Automated Workflows and Numerical Modeling), a benchmark targeting distribution fitting and time series modeling with varying difficulty tiers that culminates in real-world astronomy tasks including modeling initial mass functions and gravitational-wave chirp signals. VESTA's dynamic tool creation outperforms prior agentic pipelines for statistical analysis and visual tool creation, with the largest gains on complex and domain-specific tasks. Against strong agentic coding harnesses (Claude Code, Codex, and OpenCode), VESTA is better on all time series tasks and on complex distribution fitting tasks, though the harnesses achieve better fits on simpler distributions. Further, dynamically generated tools are substantially more sophisticated than those produced by existing visual tool-creation systems, covering more diagnostic categories per function and strongly preferring visual outputs that the VESTA VLM critic can reason over directly.

cs.AI · cs.CL · cs.CV · cs.LG · stat.CO

新论文 · v1arXiv:2611.26309

Client Exits - A Way to Increase LLM Transparency on the Client Side

Marc Jansen

原文摘要

Textual recommendations generated by large language models (LLMs) can be difficult to inspect. This paper introduces \emph{Client Exits}, a lightweight mechanism for enriching LLM responses while separating semantic intent from client-specific presentation. The model inserts predefined, machine-interpretable markers into its output, and the receiving client decides whether and how to realize the corresponding enrichments. This allows supplementary information to be presented as graphics, text, speech, or other modalities without prescribing a concrete interface. We implement a prototype combining a finetuned Qwen model with a modified LibreChat client. Using the Question--Option--Criteria (QOC) approach including three XML-based client exits. We evaluate two QOC-oriented model variants trained on essentially the same examples, with client-exit annotations added for one variant. Across 100 test questions, the client-exit model includes all three markers in every response. An LLM judge, informed about the mechanism without being instructed to favor it, prefers this model's answers in 66\% of the comparisons. These results demonstrate successful marker generation on the evaluated test set and support the feasibility of client-controlled enrichment.

cs.AI

新论文 · v1arXiv:2611.26282

What If Chain-of-Thought Is Not Enough?

Marc Jansen

原文摘要

Reasoning capabilities in large language models have evolved from prompting-based methods to internalized model behaviors. In this paper, we investigate whether the internal thinking model of an LLM can be deliberately changed and whether such a change is beneficial. Instead of relying on Chain-of-Thought, we train models to follow Question--Option--Criteria (QOC), a structured decision process that decomposes a task into explicit questions, competing options, and evaluation criteria. We present a three-generation development pipeline that ranges from a basic QOC decision core to richer answer synthesis and statistically enriched QOC reasoning. We then evaluate the resulting models with respect to both internal consistency and output quality. Across the reported experiments, QOC-based fine-tuning improves answer quality relative to non-fine-tuned base models and yields particularly strong gains once models are large enough to execute the underlying QOC computations reliably. The results suggest that the internal thinking model of an LLM is trainable and that structured alternatives to Chain-of-Thought can provide a promising path toward more reliable, interpretable, and task-adapted decision support.

cs.CL · cs.AI

新论文 · v1arXiv:2611.26264

A Statistical Framework for Training Data Auditing of Large Language Models

Shengjie Niu, Yeheng Ge, Jian Huang

原文摘要

Training-data auditing asks whether data of interest were used to train a large language model (LLM), a question central to privacy auditing, copyright protection, and benchmark-contamination detection. These concerns are particularly relevant to the use of copyrighted books and scientific papers in pre-training, as well as synthetic text and medical question-answer records in fine-tuning. In practice, training membership signals may be concentrated in only a few passages of a long document, making it important to aggregate evidence within each document without diluting these localized signals. Beyond individual documents, auditing a target collection requires an overall assessment based on the document-level evidence. In this paper, we propose a unified statistical framework for document-level membership inference, target-collection global auditing, and suspicious-subset identification. We integrate complementary chunk-level membership evidence with Cauchy combination to construct document scores for document-level membership inference. The document scores and their calibrated p-values further enable an overall assessment of training-data collection use and the identification of suspicious documents through multiple testing. Under the stated exchangeability conditions, we establish finite-sample validity of document-level p-values, type-I error control for global auditing, and false discovery rate control for subset identification. Case studies on BookTection, Pile-arXiv, Cosmopedia, and HealthCareMagic demonstrate the effectiveness and reliability of the proposed framework in auditing pre-training and fine-tuning data.

stat.ML · cs.LG · stat.AP

版本更新 · v2arXiv:2603.23875

Self-Evolving Multi-Agent Framework for Efficient Decision Making in Real-Time Strategy Scenarios

Li Ma, Hao Peng, Yiming Wang, Jie Liu, Kongjing Gu, Kangli Zi, Na Han, Hui Lin

原文摘要

Large language models (LLMs) have demonstrated exceptional potential in complex reasoning,pioneering a new paradigm for autonomous agent decision making in dynamic settings. However, in Real-Time Strategy (RTS) scenarios, LLMs suffer from a critical speed-quality trade-off. Specifically expansive state spaces and time limits render inference delays prohibitive, while stochastic planning errors undermine logical consistency. To address these challenges, we present SEMA (Self-Evolving Multi-Agent), a novel framework designed for high-performance, low-latency decision-making in RTS environments. This collaborative multi-agent framework facilitates self-evolution by adaptively calibrating model bias through in-episode assessment and cross-episode analysis. We further incorporate dynamic observation pruning based on structural entropy to model game states topologically. By distilling high dimensional data into core semantic information, this approach significantly reduces inference time. We also develop a hybrid knowledge-memory mechanism that integrates micro-trajectories, macro-experience, and hierarchical domain knowledge, thereby enhancing both strategic adaptability and decision consistency. Experiments across multiple StarCraft II maps demonstrate that SEMA achieves superior win rates while reducing average decision latency by over 50%, validating its efficiency and robustness in complex RTS scenarios.

cs.MA

版本更新 · v2arXiv:2606.28666

Why Trust Your Agent? Empirical Security Gains from TRiSM-Guided Agentic Workflows in Healthcare

Liam Kearns

原文摘要

Agent-based AI has enabled the automation of tasks by exposing application tools and resources to large language models (LLMs). However, to improve scope and accuracy, agents are often given access rights that exceed those of ordinary users, introducing significant security risks and potential regulatory breaches. This paper applies the AI Trust, Risk, and Security Management (TRiSM) framework to a medical report-generation application to demonstrate how an insecure baseline workflow can be transformed into a security-conscious agentic workflow. Both designs were evaluated with five LLMs (Claude Haiku 4.5, GPT-4.1-nano, GPT-4.1-mini, GPT-5.4-mini and Gemini 2.5 Flash) across 800 report generations and 500 attack trials covering RAG poisoning, data-field injection and client-side network injection. The redesign reduced attack success from 31% to 10% for RAG poisoning and from 42% to 25% for data-field injection (McNemar p < 0.001 for both), removed network injection by construction, and raised report accuracy from 72.5% to 86.5%. In an exploratory post-hoc grouping by attack goal, success against attacks aimed outside the report task, such as data disclosure and tool abuse, fell from 21.7% to 2.5% and no patient record was disclosed under the redesign. Attacks altering the reported Covid-19 diagnosis remained effective, accounting for 32 of the 35 successful attacks against the redesign. Least-privilege architecture limits what an injection can reach, but not whether a model follows injected content within its task, so clinical outputs need verification independent of the LLM. No model resisted all three attack vectors, so model selection cannot substitute for secure workflow design.

cs.CR · cs.AI

版本更新 · v2arXiv:2610.39621

The Real Cost of Context: Predicting Compute Time and Energy Consumption in LLM Prefill and Decode Phases

Lars Krupp, Daniel Geißler, Francisco M. Calatrava-Nicolas, Stefan Gerd Fritsch, Paul Lukowicz, Jakob Karolus

原文摘要

The integration of Large Language Models (LLMs) into agentic architectures has substantially increased token consumption, raising concerns about their energy footprint. Yet, estimating the energy consumption of proprietary API-based models remains challenging, and existing approaches focus primarily on generated output tokens while overlooking the impact of the large input contexts characteristic of agentic workflows. To address this gap, we systematically profile compute time and energy consumption across a wide range of prefill (input) and decode (output) token lengths. Using generalized linear models, we show that decode compute time and energy scale linearly with output length, whereas prefill compute time and energy exhibit both linear and quadratic scaling with input length. These findings demonstrate that input context becomes a relevant factor for large-context workloads. Furthermore, we show that compute time serves as a reliable proxy for energy consumption when direct energy measurements are unavailable. Our mapping framework enables precise energy estimation and supports more sustainable deployment decisions for agentic AI.

cs.AI

新论文 · v1arXiv:2611.26221

Sightline: Calibrated Data-Flow Summaries for LLM-Integrated Applications

Zihao Xu, Yuekang Li, Ruitao Feng, Haifeng Shen

原文摘要

Programs increasingly call large language models (LLMs), filling prompt placeholders with program values and consuming the output. Such a call crosses the boundary between natural language and programming language (NL/PL). To follow data across it, an analysis needs a data-flow summary of which placeholders the output uses, but no code describes what the model does. Analyses of these LLM-integrated applications therefore taint every output with every input, creating spurious flows, or ask a second model to label each call, which is costly. We prove, and mechanize in Lean 4, that a placeholder flows to the output of a transformer, to first order, only through the attention that reads it (no reading, no flow), and that a faint reading or sensitivity cannot certify that a placeholder does not flow. Unlike reading, the log-probability gradients and Fisher signals of existing methods vanish on confident outputs, and single-span removal fails on duplicated content. Our tool, Sightline, reads a recorded call in one forward pass, learns from labels which reads are uses, and calibrates the pruning to miss at most a chosen fraction $α$ of true flows, on average over call sites, when deployment is exchangeable with calibration. On 2,103 unseen repository families and four models, Sightline is the most accurate summary against the seven closest methods and a hidden-state probe. Calibrated at $α$ = 1%, it misses about 1% of true flows and prunes 43.9-65.6% of spurious flows, 1.6 to 2.6 times as many as the best baseline, and still 43.8-60.1% when calibrated on another collection. Through an open surrogate, it reads models treated as closed about as accurately as they read themselves, without their labels. In the information-flow control of an LLM agent, it removes about half of the labels from inputs that do not flow.

cs.PL · cs.AI

版本更新 · v3arXiv:2606.27786

SPARK: Adaptive Parametric Knowledge Participation for Faithful Retrieval-Augmented Generation

Ruochang Li, Pengcheng Huang, Zhenghao Liu, Yukun Yan, Huiyuan Xie, Yu Gu, Ge Yu, Maosong Sun

原文摘要

Retrieval-augmented generation (RAG) enhances the faithfulness of large language models (LLMs) by incorporating external knowledge into the generation process. However, conflicts between retrieved context and parametric knowledge can cause LLMs to rely on inconsistent internal knowledge, hindering their ability to faithfully leverage retrieved evidence. Existing approaches mitigate this issue by editing parametric knowledge to resolve such conflicts. However, locating and modifying knowledge encoded in model parameters is challenging and may inadvertently compromise the backbone LLM's capabilities. In this paper, we introduce SPARK, a novel framework that formulates parametric knowledge control as a latent participation policy learning problem, enabling LLMs to adaptively regulate the contribution of parametric knowledge during generation. Specifically, SPARK introduces a lightweight participation policy that dynamically modulates the contribution of FFN-mediated parametric knowledge while keeping the backbone LLM frozen. The participation policy is optimized with a faithfulness-oriented reward, enabling fine-grained control over parametric knowledge participation during generation. Extensive experiments on six datasets validate the effectiveness of SPARK, which introduces and optimizes only 0.01\% additional trainable parameters while keeping the backbone LLM fully frozen, thereby preserving its general capabilities and achieving more faithful generation than competitive baselines. All datasets and code are available at https://github.com/OpenBMB/SPARK.

cs.CL · cs.AI

新论文 · v1arXiv:2611.26171

The FrontierOR Challenge: Agentic Algorithm Design for Operations Research

Hai Wang, Bretislav Hajek, Chonghe Jiang, Minwei Kong, Ao Qu, Huatian Gong, Bo Dong, Cynthia Barnhart 等 9 位作者

原文摘要

Operations research (OR) relies on algorithms that experts tailor to each problem: a new routing, scheduling or network design problem can take weeks of specialist work before it is solved well at scale. Large language model (LLM) agents now write, test and improve code, but whether they can design algorithms that scale to realistic instances is open: in the FrontierOR benchmark's one-shot evaluation, the strongest model matches Gurobi, with an objective within 1% of Gurobi's and no slower, on only 39% of the large instances. The FrontierOR Challenge asks teams to build agentic workflows that design, test, and improve algorithms for many optimization problems at once. Stage 1 releases 60 of the benchmark's 180 audited problems, spanning six application categories and five formulation types, in weekly rounds of 15; Stage 2 introduces new problems outside the benchmark, with unseen instances. Two tracks are ranked separately: in the Main track we run the submitted algorithms in a network-isolated sandbox with open-source dependencies only; in the Self-hosted track teams solve the instances on their own machines, with any software, within a timed window. Every solution is checked for feasibility, and a feasible solution scores the ratio of its objective to a frozen known reference, averaged over hidden instances.

cs.AI · math.OC