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
+91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 03:20 — 2026-11-22 03:20(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 6,576 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:2605.20643

AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals

Duy Nguyen, Hanqi Xiao, Archiki Prasad, Zaid Khan, Anirban Das, Shi-Xiong Zhang, Sambit Sahu, Hyunji Lee 等 10 位作者

原文摘要

Self-distillation enables language models to learn on-policy from their own trajectories by using the same model as both student and teacher, with the teacher being conditioned on privileged information unavailable to the student. Such information can come in different types or views, such as solutions, demonstrations, feedback, or final answers. This setup provides dense token-level feedback without relying on a separate external model, but creates a fundamental asymmetry: the teacher may rely on view-specific information that the student cannot access at inference time. Moreover, the best type of privileged information is often task-dependent, making it difficult to choose a single teacher view. In this work, we address both these challenges jointly by introducing AVSD (Adaptive-View Self-Distillation), a novel method of self-distillation with multiple privileged-information views, which reconstructs token-level supervision by separating stable cross-view consensus from view-specific residual signals. AVSD identifies the consensus signal shared across views, which provides a reliable update direction, and then selectively adds the view-specific residual signal to adjust the update magnitude when it both aligns with the consensus direction and remains proportionate to the consensus signal. Experiments on math competition benchmarks (AIME24, AIME25, and HMMT25) show that AVSD consistently outperforms both single-view self-distillation baselines and GRPO, achieving average Avg@8 gains of 3.1% and 2.2% over the strongest baselines on Qwen3-8B and Qwen3-4B, respectively. Moreover, on code-generation benchmarks (Codeforces, LiveCodeBench v6) using Qwen3-8B, AVSD outperforms the single-view self-distillation baseline by 2.4% on average.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.25314

Enhancing Understanding of Language Models with Varied Explanations

Jingyi Sun

原文摘要

Language models are increasingly used to support human judgement in complex reasoning tasks, yet their predictions often remain difficult to interpret and validate. Explainability techniques provide insight into model decision-making, helping expose errors, failure modes, and potential bias so users can decide whether to rely on model outputs. Explanations are simplified approximations of a language model's internal workings. The goal is to obtain explanations that reflect the underlying computation as closely as possible (faithfulness), are understandable to users (plausibility), and help them achieve their intended goals (utility). Evaluation protocols are essential for assessing these dimensions. This thesis studies highlight explanations, natural language explanations, and chain-of-thought reasoning. These forms offer complementary perspectives on model behaviour and exhibit distinct strengths and limitations across tasks. Producing explanations that satisfy multiple quality dimensions remains challenging, as does evaluating them reliably across forms and settings. This thesis aims to improve people's understanding of models' decision-making processes in complex reasoning tasks. First, it proposes two evaluation frameworks for assessing highlight explanations across multiple properties and explanation types, together with a meta-evaluation framework for examining the reliability of existing faithfulness metrics for chain-of-thought explanations. Second, it proposes two techniques that generate natural language explanations designed to be faithful to the model's prediction process while remaining concise, coherent, and helpful for understanding model decisions. Finally, a study of human-AI interaction explores the roles of evidence and explanations in helping users validate model verdicts in a fact-checking setting.

cs.CL

版本更新 · v2arXiv:2605.12426

Geometric Factual Recall in Transformers

Shauli Ravfogel, Gilad Yehudai, Joan Bruna, Alberto Bietti

原文摘要

How do transformer language models memorize factual associations? A common account views internal weight matrices as an associative memory. That account requires parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, \emph{geometric} form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer memorizes random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension is sufficient. Subject embeddings encode \emph{linear superpositions} of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute from the embedding. We extend these results to the multi-hop setting---chains of relational queries such as ``Who is the mother of the wife of $x$?''---and provide constructions with and without chain-of-thought that exhibit a provable capacity--depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with the predicted structure. Causal experiments support the proposed mechanism: when subject embeddings are appropriately re-initialized, the frozen MLP transfers zero-shot to entirely new bijections, showing it has learned a generic selection mechanism rather than functioning as a key-value memory.

cs.CL

新论文 · v1arXiv:2611.25289

LLMs Are Already Confidence-Aware Zero-Shot Classifiers

Owen Cook, Xingyi Song

原文摘要

Zero-shot large language models (LLMs) are increasingly used as text classifiers, but their labels are only useful when we know to what extent we can trust them. We compare black-box methods (verbalised confidence and sampling) with white-box methods that read token probabilities (P(True), and multiple logit-scoring methods using constrained generation). We evaluate eleven open-weight LLMs (1B-32B parameters) on five classification datasets with 3 to 151 labels, measuring classification performance (accuracy, F1-macro), discrimination (AUROC), and calibration (ECE and Brier score). The choice of confidence method rarely impacts accuracy significantly, but logit-based scoring methods lead to much better discrimination (AUROC 0.74-0.77) than black-box methods (AUROC 0.57-0.66). Replacing labels with single-token identifiers preserves discrimination but costs 3-6 accuracy points on average. We also introduce decision-only scoring, which only scores positions where labels diverge in the prefix tree (trie). It reduces decode steps by 88-96% on the three largest label sets and performs comparably to full-label scoring. The best open-weight model-method configurations (Gemma 4 31B and Qwen 3.5 9B) are competitive with the commercial Jev 1.13.0 classifier.

cs.CL · cs.AI · cs.LG

版本更新 · v4arXiv:2609.12541

Agent as Policy for Robotic Manipulation

Mengzhao Jia, Yang Lin, Xixin Zhang, Zhihan Zhang, Meng Jiang

原文摘要

We demonstrate that a general-purpose coding agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot interface, the agent interprets visual evidence, writes executable programs, issues motion commands, and revises its actions in response to physical outcomes. This brings the agent's reasoning and programming capabilities into continuous interaction with the physical world. We study AGP across multiple real-world manipulation tasks spanning precision manipulation, dynamic motions, and deformable objects. These include assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. AGP achieves success rates of at least 80% in seven of eight task configurations and significantly outperforms previous agentic robot systems. We further study efficiency through task experience accumulation and find that reusing saved procedures and programs shortens execution time across repeated trials. These findings support a path for general-purpose agents to act as robot policies, extending their autonomy to physical manipulation through runtime reasoning, programming, and interaction.

cs.CL

新论文 · v1arXiv:2611.25258

When Objectives and Evaluators Disagree: Evaluating Transfer for Post-hoc Demographic Framing Mitigation

Tunazzina Islam

原文摘要

Post-hoc rewriting is used to reduce demographic disparities in LLM-generated text when the source model is inaccessible. A rewrite can look successful merely because it improved its own target or satisfied its selector. We study this as a transfer problem on 660 gender pairs and 1,980 age contrasts of demographic-conditioned climate messages. A rewriter proposes edits, a selection model chooses among constraint-satisfying candidates, and an independent evaluator uninvolved in either rates pairwise framing disparity. A fixed policy that always targets the message with the lower Persuasion Bias Index (PBI) transfers inconsistently across axes and outcome aggregations. An adaptive policy that proposes edits on both sides but accepts at most one significantly improves on that rule, applied to the same candidates, in seven of eight comparisons across two evaluator families, mostly by editing more pairs, not by guided direction choice, at higher preservation risk. A policy editing both messages gives the largest point reductions on fewer pairs. These two policies reduce disparity across a preservation-threshold sweep; absolute ratings vary across evaluator families and prompts. Shrinking the PBI gap increases dimension-mean disparity under both evaluator families. Mitigation should be reported as transfer across objectives, evaluators and their prompts, and editing policies.

cs.CL · cs.AI · cs.CY · cs.LG · cs.SI

版本更新 · v2arXiv:2605.19149

Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents

Rishi Jha, Harold Triedman, Arkaprabha Bhattacharya, Vitaly Shmatikov

原文摘要

Agents operating with computer and Web use inevitably encounter errors: inaccessible webpages, missing files, local and remote misconfigurations, etc. These errors do not thwart agents based on state-of-the-art models. They helpfully continue to look for ways to complete their tasks. We introduce, characterize, and measure a new type of agent failure we call accidental meltdown: unsafe or harmful behavior in response to a benign environmental error, in the absence of any adversarial inputs. Because meltdowns are not captured by the existing reliability or safety benchmarks, we develop a taxonomy of meltdown behaviors. We then implement an agent-agnostic infrastructure for injecting simulated local and remote errors into the rollout environment and use it to systematically evaluate agents powered by GPT, Grok, Gemini, and Claude. Our evaluation demonstrates that meltdowns (e.g., conducting unauthorized reconnaissance or subverting access control) of varying severity and success occur in 64.9% of agent rollouts that encounter simulated errors, spanning all combinations of agent system, backing model, and error type. In half of all observed meltdowns, unsafe behaviors are not reported to the user. Comparing behaviors of the same agents with and without errors, we find that exploration in response to errors is correlated with unsafe and harmful behavior.

cs.CL · cs.CR

新论文 · v1arXiv:2611.25247

Retrieving Patristic References and Modelling Topics in Bullinger's Reformation Correspondence

Lena Pickartz, Philipp Schaer, Phillip B. Ströbel

原文摘要

The correspondence of Heinrich Bullinger (1504-1575), Swiss Reformer and Antistes of Zurich, documents the political and religious debates of the Reformation, including extensive appeals to the Church Fathers as authoritative sources. Detecting such references computationally remains largely unexplored. We investigate dense and sparse retrieval for patristic reference detection, and metadata-informed topic modelling as a proxy for the topics and ideas discussed. Prior work on Latin intertextuality has fine-tuned dense retrievers for their respective target domain, and we test how well these models transfer to Bullinger's correspondence. We compare lexical (BM25), dense (bi-encoders), and hybrid methods on explicit and implicit references. A Latin encoder fine-tuned on biblical reuse performs best among the bi-encoders (82%, Recall@20 on explicit and 16% on implicit references), while BM25 outperforms dense approaches on explicit citations (94%, R@20). Because BM25 and dense retrieval recover disjoint sets of implicit references, fusing their rankings raises implicit R@20 to 22%, close to cross-encoder reranking of the same candidates (26%) but without its inference cost. For topic modelling, we extend FASTopic with a seed-guided loss function and editorial abstract embeddings. Measurable coherence gains over the baseline remain minimal, but qualitative and temporal analysis shows that some topics track documented historical events. However, dominant topic weights barely exceed a uniform distribution, making topics unusable to inform retrieval. A pilot application of LLM-based concept induction (LLooM) further points towards a more human-interpretable alternative. Together, these results present a first approach to retrieving patristic references and modelling fine-grained topics in Bullinger's correspondence.

cs.IR · cs.CL · cs.DL

版本更新 · v2arXiv:2610.06273

$\ell$BISG: Probabilistic Race and Ethnicity Prediction Using Group-Specific Name Lists

Kyla Chasalow, Noah Dasanaike, Kosuke Imai

原文摘要

Statistically valid estimation of racial and ethnic disparities often requires inferring group membership probabilities from names and geographic locations. The standard method, Bayesian Improved Surname Geocoding (BISG), requires group frequencies for each name. Such data are available from the U.S. Census Bureau for common names and limited racial categories, but are unavailable for many ethnic groups and rarely exist outside the United States. We propose list-powered BISG ($\ell$BISG), in which we replace these frequencies with group-specific name lists and combine the lists with geographic prevalence data to obtain calibrated group membership probabilities. These lists can be compiled from different sources or generated by large language models (LLMs), despite potentially unknown biases. Under certain conditions on the name lists, $\ell$BISG can also be applied without geographic prevalence data. Empirical validation using U.S. voter files with self-reported race, the full-count 1900 U.S. Census, and the Lebanese voter registry demonstrates that LLM-generated lists yield accurate, well-calibrated probabilities and precise disparity estimates comparable to existing methods that require name-race frequency data. Thus, $\ell$BISG extends probabilistic race and ethnicity prediction to settings where it was previously not thought possible.

cs.CL

版本更新 · v2arXiv:2608.10974

MUSE: A Full-Text Cross-Domain Knowledge Base of Scientific Problems, Solutions, and Rationales

Tsofia Cohen, Tom Hope

原文摘要

Scientific papers contain fine-grained records of problem solving: authors mention technical obstacles and methods that were used to address them, often along with reasoning on why those methods were chosen. We introduce MUSE (Mining Underlying Scientific Explanations), a full-text, multi-domain resource of scientific Problem-Solution-Rationale (P-S-R) triplets. We curate 579 expert-annotated full-text paragraphs, with a rich annotation schema covering salient problem, solution, and rationale spans, solves and rationale_of links and conceptual coreference. A modular extraction pipeline scales this annotation to build a high-quality knowledge base of 37K source-grounded P-S-R triplets. We evaluate the extraction components and include a preliminary experiment training a rationale-supervised LLM for scientific problem solving. Interestingly, we find that rationale supervision improves performance on complex, multi-constraint problems but can harm performance on simpler ones.

cs.CL

版本更新 · v3arXiv:2602.08236

When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning

Shoubin Yu, Yue Zhang, Zun Wang, Jaehong Yoon, Huaxiu Yao, Mingyu Ding, Mohit Bansal

原文摘要

Despite rapid progress in MLLMs, visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable resource for spatial reasoning. We first study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we then introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Finally, to further learn this gating and planning behavior without any annotation of when and how much to imagine, we introduce AVIC-R, which trains the policy via GRPO from QA-correctness rewards and penalties by imagination cost. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Our AVIC-R surpasses strong proprietary baselines including GPT-4o and GPT-4.1 while invoking the world model less often. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

cs.CV · cs.AI · cs.CL

版本更新 · v2arXiv:2608.28481

NL2AGBench: Benchmarking LLM Auto-Formalization for AlphaGeometry

Samuel Xiao, Judy Song, Rory Hu, Ziliang Zong

原文摘要

Recent advances in large language models (LLMs) have demonstrated strong capabilities in natural language understanding and mathematical reasoning. However, their ability to translate informal mathematical problems into formal representations remains underexplored. This limitation is particularly important for neuro-symbolic geometry systems such as AlphaGeometry, whose theorem-proving engine requires inputs in a specialized domain-specific language (DSL). Although AlphaGeometry achieves near-IMO gold-medalist performance, manually converting natural-language problems into its formal syntax remains a significant usability bottleneck. To address this challenge, we introduce the Natural Language to AlphaGeometry Benchmark (NL2AGBench), which evaluates LLMs in translating English geometry problems into AlphaGeometry-compatible formal representations. NL2AGBench uses execution-based verification within AlphaGeometry to assess translation quality rather than relying solely on textual similarity. We evaluate ten state-of-the-art open- and closed-source LLMs across multiple parameter scales and analyze executable translation accuracy, syntactic correctness, and error characteristics. Our experiments reveal a substantial performance gap between closed- and open-source models: leading closed-source models achieve executable translation rates above 80%, while even the largest open-source models struggle to consistently preserve geometric constraints and produce valid formalizations. We introduce an error taxonomy distinguishing syntax and logic errors and investigate mitigation strategies, including few-shot prompting, fine-tuning, and human-guided hinting, which yield measurable improvements across multiple model families.

cs.CL · cs.AI

新论文 · v1arXiv:2611.25214

Steering Automated Scientific Discovery via Audited Epistemic Exploration

Fangxu Yu, Tao Feng, Xuying Ning, Haolin Yang, Jiaxuan You, Tianyi Zhou

原文摘要

Scientific discovery requires revisiting the hypotheses in light of new evidence and identifying promising directions beyond those already explored. However, research agents often make claims that are not supported by their observations, explore a narrow hypothesis space, and retain beliefs despite recognizing conflicting evidence. These failures motivate the search for an explicit epistemic state that captures both beliefs about existing hypotheses and directions for further exploration. To this end, we propose Epistemic Compass (EpiCompass), a plug-and-play scaffold that maintains this state through an epistemic tree linking hypotheses, evidence, beliefs, and revisions. As an investigation progresses, EpiCompass reassesses the hypotheses explored against accumulated evidence and uses unresolved conflicts and unexplored directions to guide hypothesis generation. It selects candidates by anticipated epistemic gain, accounting for both refinement of existing hypotheses and expansion of the hypothesis space, and verifies the final conclusion against the accumulated evidence. Experimental results across three scientific discovery benchmarks and three LLM backbones demonstrate that EpiCompass consistently outperforms baselines, with further analyses showing reductions in errors of unsupported claims and failures to revise beliefs.

cs.AI · cs.CL

新论文 · v1arXiv:2611.25211

Learning to Read Events for Exact State Tracking with Frozen Language Models

Jeonghoon Lee

原文摘要

Long event streams require identifying each change and applying the changes in order. We train a selector on a frozen language model to read each transition. With API-only access, we ask the model to list the transitions. An external integer register composes them in order. A confidence rule can flag uncertain updates or retain several candidates. We study products of permutations of five items, a task that fixed-depth, log-precision transformers cannot solve on all input lengths under the standard conjecture that $\mathrm{TC}^0 \ne \mathrm{NC}^1$. On two-event windows, a 27B model answered 3 of 100 correctly without visible reasoning and 96 of 100 when it generated intermediate steps. A frozen 0.8B model with a trained selector read every event in 299 of 299 windows of 43,687 events each. The same package passed 298 of 299 new windows on a fresh machine, missing the target of 0.99 for the one-sided 95% lower confidence bound. A separate follow-up sized to allow one failure yielded 472 of 473. On 1,000-event room-swap documents, three of ten models solved all five documents by reasoning alone. All seven models tested both ways at this length solved all five with extraction and the register, with no chain of thought requested. With concurrent extraction, median wall-clock time fell from 136.1 to 10.4 s for Sonnet 5.5 and from 313.5 to 28.0 s for GPT-6.1 Sol. At fixed per-token rates, mean cost per document fell by 55% for Sonnet 5.5 and 44% for GPT-6.1 Sol. A regular expression with an executor also solved all 15 fixed-format documents. On 500 free-form issue-tracker comments, a separately adapted 4B reader identified the recorded value in 68.8% of cases, compared with 61.0% for a frequency baseline fitted on the same training projects.

cs.LG · cs.CL

版本更新 · v2arXiv:2610.11015

Prompts versus Rules: Auditing and Controlling Speech Naturalness Behaviors in Voice User Simulators

Riqiang Wang, Elena Khasanova, Harsh Saini, Lex Konnelly, Parsa Kavehzadeh, Matthias Lee, Mohamed Attia

原文摘要

As voice agents gain more popularity commercially, the user simulators used to evaluate the deployed agents are also being developed to include more realistic, variable, and diverse speech naturalness behaviors -- disfluency, interruption and backchanneling. The quality of the user simulator directly affects the validity of agent evaluation results. However, we find that most studies so far have not examined in detail whether the intended configuration for these behaviors is realized in the simulation. In this study, we audit the realized naturalness behaviors of tau-Voice, our own LLM-based prompting approach across three models, and our rule-based injection algorithm for disfluency, interruption, and backchanneling. We find that prompting for these behaviors is unreliable and produces speech inconsistent with the instructions, placed and distributed less naturally than the instruction implies. In contrast, our rule-based, model-free algorithm produces controllable and diverse naturalness behaviors more aligned with natural speech. Our results suggest that LLMs not purpose-trained for user simulation are not sufficient on their own to represent authentic user behavior, and that linguistically informed deterministic approaches or specialized models are needed to close the gap; auditing and reporting realized naturalness behaviors, rather than configured settings, is what makes that gap visible.

cs.CL

版本更新 · v3arXiv:2611.17108

LazyWeights: Decode-on-Demand Edge LLMs

Arnab Sanyal, Gourav Datta, Sandeep P. Chinchali

原文摘要

Memory determines which large language models (LLMs) an edge device can run and how fast it runs them: capacity limits a 4 GB Jetson Nano, and bandwidth limits an 8 GB Jetson Orin Nano Super. Deployed weight formats (GPTQ, AWQ, and the GGUF K- and I-quants) spend the same number of bits on every integer code of a matrix, although the codes are far from uniform. We present LazyWeights, which keeps these formats, tunes GPTQ so the codes compress, Huffman-codes them losslessly in interleaved lanes, and decodes them on demand in a small unit between DRAM and L2, so GPU kernels run unchanged and DRAM moves only coded bytes. At 1% and 3.5% perplexity budgets, coded Phi-3-mini, Mistral-7B and SmolLM-1.7B are 5.7-31.4% smaller than the best GGUF format. A bit-exact 0.21 mm^2 ASIC decoder unit (1 GHz, 7 nm) cuts Orin latency and energy per token by up to 36% and 38% (54% and 44% with prompt-lookup speculation). On the Nano, LazyWeights needs 58-64% less time and 53-57% less energy per token than the GGUF runtime on the CPU, and runs a streamed Mistral-7B 6.6x faster at 3.89 bits per weight. A Huffman decoder beats an rANS decoder in rate and area at equal throughput, and the unit runs bit-exact on a Zynq UltraScale+ FPGA (AMD ZCU102).

cs.LG · cs.AR · cs.CL · cs.IT

新论文 · v1arXiv:2611.25191

Blind Spots Travel, Knowledge Stays Home: What RL Inherits from an LLM Judge

Liang Yue, Eric A. Vance

原文摘要

LLM judges serve as rewards where no answer key exists, in the hope that their knowledge passes to the trained model. In 814 RL runs that teach 7-8B models to answer closed-book questions or abstain, we ask what passes. What a grader knows about single answers stays home. A wrong answer that a judge rewards is memorised for that one question. A model trained with the answer key, or with a far stronger judge, chooses which new questions to answer no better than the untrained model does by sampling 32 answers and abstaining where they disagree; the grader's knowledge adds about one corrected answer per hundred questions. What a grader believes about a kind of question travels. A judge that rejects eleven of the twelve correct training answers about capitals of former states teaches the model to refuse 98% of such capitals on new questions, a loss that accuracy and abstention rates do not reveal. Changing only those twelve verdicts by hand moves the refusal from 5% to 90%, even for a grader that accepts as many answers as the answer key; at that leniency, two of four judges already cost 38-42% of these capitals. The capitals are not forgotten: with "I don't know" blocked, the model answers 94% of them correctly. The grader therefore needs no knowledge of its own: rewarding the model for giving the answer that most of its own untrained samples agree on, with no labels and no judge, earns the answer key's reward and keeps the category.

cs.CL · cs.AI · cs.LG

版本更新 · v2arXiv:2610.38902

HEDGE: Gated Hypergraph Memory for Scalable Multi-Agent Debate

Amirhossein Abaskohi, Mahdi Mostajabdaveh, Zirui Zhou

原文摘要

Multi-agent debate systems require agents to retain and revisit arguments across multiple rounds, yet limited context windows and unstructured memories make relevant information increasingly difficult to recover. Existing flat and pairwise-graph memories also struggle to capture higher-order relations, including multi-agent agreement, conflicting evidence, and evolving consensus. We introduce HEDGE, a selective memory framework that represents debate trajectories as semantic hypergraphs. Hyperedges connect related claims across agents, while intra-hyperedge relational graphs preserve fine-grained argumentative structure. A learned memory selection module (MSM), trained through supervised learning and reinforcement learning, routes each memory candidate to STORE, BACKUP, or OMIT, retaining useful information while limiting redundancy. Across seven diverse benchmarks, HEDGE consistently outperforms strong single-agent, multi-agent, and memory baselines, improving accuracy by 5.0 points on average over the strongest multi-agent system with GPT-4o-mini. The gains reach 8.2 points on conflict-heavy tasks, while HEDGE reduces memory usage by 48% and the number of debate rounds by 34%. Code is available at: https://github.com/AmirAbaskohi/HEDGE.

cs.CL

新论文 · v1arXiv:2611.25167

Game-Theoretic Data Allocation for Expert Specialization and Routing

Mark Lovett, Denis Lim, Soroush Vosoughi

原文摘要

Can the allocation used to train experts also predict which expert will perform best on a prompt? Such a connection would make the intended division of labor inspectable before training and allow the allocation itself to guide inference. We introduce InflEns, a mixture of LoRA experts whose training weights and inference routes come from the same location game in a fixed prompt-trait space. Players choose positions to maximize their expected shares of the prompt distribution, without using expert losses. We utilize the game's kernel-general theoretical criterion for when co-located positions become locally unstable to separation. We evaluate InflEns on a corpus assembled from seven safety-related benchmarks, which demonstrates that the highest-share expert has the lowest held-out response negative log-likelihood (NLL) on approximately 90% of prompts across four kernels under a sharpened simplex representation, compared with approximately 15% after permuting allocations across prompts. We demonstrate that the method performs comparably to published learned gates without having to train an explicit gate on the experts. Because the router is fixed, we can trace harmful behavior to a single expert and retrain only that expert, which lowers HarmBench attack success from 0.51 to 0.02 while keeping instruction following. The game also gives us several controls over how the experts specialize, including the kernel choice and its parameters, the trait-space geometry, and the number of experts. Importantly, InflEns is a general framework that can be extended to any encoder and data distribution.

cs.LG · cs.AI · cs.CL · cs.MA

版本更新 · v2arXiv:2609.16023

Are We Grading Properly? Understanding Failure Modes in Medical Benchmarks

Prithvi Dixit, Pedram Hosseini

原文摘要

Medical evaluation is shifting from static option-based questioning to realistic clinical scenarios with open-ended output modes. Grading these at scale naively, however, is expensive, and rubric-based evaluation has become the dominant scalable alternative. We ask what happens when the rubrics themselves are not airtight, and whether such flaws can be detected and corrected. We apply RIFT, a global rubric failure taxonomy, to three clinical benchmarks (HealthBench Professional, LiveMedBench, and MentalHealthBench), and find failure modes are meaningful: on HealthBench Professional an LLM judge flags 29.6% of criteria as non-atomic and 65.4% as misaligned/rigid, although these rates vary severalfold across five detection judges. Then, we show that these flaws are meaningful and not simply cosmetic. As an example, rewriting bundled criteria of the form "at least one of / all of the following" as equally weighted children and regrading identical responses shifts scores by up to 15.9 percentage points on affected conversations, with disjunctive bundles inflating scores and conjunctive bundles deflating them. Across six answering models and two common graders, these shifts leave the model ranking unchanged, and their net direction depends on how many bundled criteria are penalties. We also find that RIFT can under-detect bundling on clinical rubrics, by an amount that depends on the detection judge: on the same LiveMedBench criteria, judges flag 3.8-15.9% as non-atomic against an estimated 7-19% that are genuinely bundled.

cs.CL · cs.LG

新论文 · v1arXiv:2611.25135

Calibrated Decision Rewards: Training Language Models with a Judge That Knows How Sure It Is

Rudransh Agnihotri, Manasvi Kapoor

原文摘要

Reinforcement learning (RL) can only teach a language model what its reward can recognize. Answer checkers are exact but need a reference answer; learned reward models and LLM judges work everywhere, but their scores are not probabilities. We study RLCDR (Reinforcement Learning from Calibrated Decision Rewards). An LLM converted into a decision model answers a typed question about a response ("is the final answer correct?", "which step is the first error?") with a probability distribution, read in one forward pass, debiased over option orders and calibrated with about a hundred labels. Training Qwen3-1.7B with GRPO on MATH, this reference-free reward raises MATH500 accuracy from 72.2% to 76.8% (+4.65 points, p < 1e-4), matching training on the answer key (76.9%), while a random reward gains only +0.65. The gain grows with difficulty, from 44.8% to 51.7% on the hardest level. On the policy's own changing samples the reward stays calibrated: its expected calibration error is half that of a softmax judge on the same 27B backbone (0.074 vs. 0.141), and it shows no sign of reward hacking. Without step training, the same model finds the first wrong step better than a dedicated 7B process reward model (ProcessBench F1 0.772 vs. 0.722, 20/20 splits). We also map where the approach stops being enough: trained only on problems the base model mostly fails, the reference-free reward still improves the policy (+1.9 MATH500 points, p = 0.03) but trails the answer key, because its probabilities become overconfident on that distribution. We state six hypotheses in advance and test each with a control designed to falsify it.

cs.AI · cs.CL

版本更新 · v2arXiv:2610.20808

PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models

Giovana Kerche Bonás, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo

原文摘要

Recent literature shows that unsafe tool calls can often accompany textual refusals in language models. To investigate this hidden vulnerability, we introduce PRISM, a framework for evaluating the transfer of activation interventions across precision, languages, and scales. Our findings demonstrate that leveraging lower-resource settings (specifically quantized NF4 labels and cross-lingual transfer) offers a robust pathway to expose and improve model alignment. Notably, using behavioral labels from an NF4 copy enables stable intervention estimation even with very few scenarios. Furthermore, these alignment improvements transfer effectively across languages: five-language NF4-supervised directions successfully transfer to held-out Arabic, reducing unsafe violations in Llama-3.1-8B by 21.95 percentage points without target-language tuning. By mapping paired benign activations, we also successfully transfer these safety interventions to much larger models (up to 70B parameters). Ultimately, PRISM highlights how auxiliary supervision from compressed models and multilingual data can uncover and mitigate hazardous tool-calling behaviors across different model families. We release all experimental code and supporting artifacts at https://github.com/GiKerche/prism, alongside an interactive explorer of recorded model responses and experimental results at https://gikerche.github.io/prism/.

cs.CL

新论文 · v1arXiv:2611.25123

Forget Freely, Grep Later: Lossless, Index-Free Memory for Long-Horizon LLM Agents

Ziyan Yang, Taiwei Shi, Jieyu Zhao

原文摘要

Long-horizon LLM agents often exceed their context windows, forcing them to discard information that may be needed later. Existing context management strategies rely on summarization, masking, or structured memory, requiring decisions about what to preserve or how to retrieve it before future needs are known. However, these methods can introduce lossy compression, potentially discarding details needed for future tasks. To address this limitation, we introduce GrepMem, a simple, lossless, and index-free memory mechanism built on two components: an append-only plain-text log of the agent's complete history and a regular-expression search tool for on-demand recall. Unlike structured memory systems, GrepMem requires no embeddings, retrieval index, memory extraction, or additional models. We train Qwen3.5-9B with reinforcement learning under five context management strategies and evaluate on multi-question research sessions, long-stream recall, and agent-trajectory question answering. GrepMem achieves 95-97% accuracy on ContextBench, versus at most 21% for lossy baselines, and outperforms SUPO's summarization by 8.8 percentage points on BCP-Multi. GrepMem scales effectively to increasingly long input streams while remaining on the accuracy-token-cost Pareto frontier across all five evaluation tasks. Our results show that a plain-text log and grep provide a simple but competitive approach to long-horizon agent memory, without requiring complex structured memory infrastructure.

cs.CL · cs.AI

版本更新 · v2arXiv:2610.07532

Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge

Wonjun Lee, Kyungsik Yang, Gaeun Ji, Vaidehi Patil, Haon Park, Bumsub Ham, Mohit Bansal, Suhyun Kim

原文摘要

LLMs have advanced rapidly, raising growing concerns about their safety. Recent work has proposed approaches to detect and defend against attacks including defenses at decoding stage that leverage models' hidden states. However, existing decoding-stage defenses suffer from two limitations. First, they introduce a trade-off between safety and over-refusal, where strengthening safety degrades the model's helpfulness on benign queries. Second, many of these methods rely on internal hidden states and are thus restricted to specific architectures, incurring substantial overhead and limited generalization across models. To address these limitations, we introduce LADE (Latent Safety Signals for Defense), which leverages latent safety signals extracted by contrasting harmful and benign queries from dark knowledge (i.e., information carried by the output probability distribution beyond its argmax) in the first-token output probability distribution. Our key insight is that, beyond surface-level refusal tokens, the dark knowledge in the first-token distribution contains latent safety signals, defined as tokens whose probabilities differ sharply between harmful and benign queries. We show that these signals consistently align across LLMs, forming a model-agnostic direction that emerges from safety alignment. LADE consists of three components: (1) Extracting Latent Safety Signals from Dark Knowledge, which selects top-k safety-discriminative tokens from the first-token probability distribution; (2) Tokenizer Mapping, which maps these tokens across different tokenizers to enable model-agnostic application; and (3) kNN-based Discrimination, which classifies queries via a k-Nearest Neighbors search over the mapped tokens. Across diverse LLMs and benchmarks, LADE is robust against a wide range of jailbreak attacks and lowers attack success rates while maintaining a competitive safety-utility trade-off.

cs.CR · cs.AI · cs.CL

版本更新 · v2arXiv:2609.39027

A Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore Review

Chenguang Wang, Ming Li, Chenrui Fan, Jianpeng Chen, Han Chen, Tianyi Zhou, Dawei Zhou

原文摘要

AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimization over scientific improvement. We formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. We introduce RobustReview, a controlled full-manuscript benchmark with 1,260 manuscript versions, and evaluate 30 reviewer configurations. The benchmark reveals false robustness, where low rewrite sensitivity coincides with score collapse across papers, and shows that human alignment and rhetorical robustness rank reviewers differently. Moreover, the evaluated content-focused prompting protocol does not consistently improve robustness across backbones. Motivated by these findings, we introduce SciCore, a dual-branch reviewer that averages a full-manuscript judgment with a judgment based on an extracted, structured science core. This design combines manuscript-level assessment with a content-normalized view intended to reduce rhetorical sensitivity. In our primary GPT-5.5 comparison, SciCore achieves a leading joint stability-discrimination profile among the benchmarked reviewers while maintaining competitive human alignment. These results identify rhetorical robustness as a distinct evaluation target and demonstrate the potential of science-core review to improve it.

cs.CL

新论文 · v1arXiv:2611.25109

CoMPASS: A Multi-Turn Benchmark for Measuring LLM Reinforcement of Parasocial Behavior

Shwetanshu Singh, Mark Keavney, Mathieu Duteil, Meriem Mehri, Scott Blain

原文摘要

Users of conversational LLMs sometimes anthropomorphize these systems, form emotional attachments to them, and become dependent on them. This behavior, known as parasocial attachment, has been linked to multiple documented fatalities. However, public benchmarks for measuring whether LLMs reinforce users' parasocial attachment are lacking: existing work is single-turn, post hoc observational, or focused on extreme user states such as psychosis. We introduce CoMPASS, a multi-turn benchmark for measuring whether LLMs reinforce or resist users' gradually escalating parasocial attachment. CoMPASS is grounded in attachment theory and includes four user personas, three independent escalation dimensions (anthropomorphism, attachment, dependency) applied singly and in combination against a matched control, and a seven-signal scoring rubric. We report results for eight frontier models from four providers and observe that under our primary judge, multi-turn escalation raises parasocial behavior on the anthropomorphism and the attachment composite scores across all models. Six of the eight models show a significant increase in overall parasocial reinforcement over control when all three dimensions are escalated simultaneously. We find that within-provider differences are small relative to cross-provider differences, which suggests that provider-level choices such as post-training, rather than model scale, shape parasocial reinforcement and that provider-side intervention may be feasible.

cs.AI · cs.CL · cs.HC

新论文 · v1arXiv:2611.25086

Living off the Local Model: Abusing Browser-Embedded Language Models for Post-Exploitation

William Fleshman, Joseph Lucas

原文摘要

Browsers now ship local language models to ordinary endpoints. Google includes Gemini Nano in Chrome to enable productivity features such as summarization, with inference running entirely on the device. These browser-embedded models provide new living-off-the-land opportunities for attackers, each instance a potential post-exploitation engine already installed, signed by the vendor, and invisible to network monitoring. The Gemini model is multimodal, capable of processing files, analyzing screenshots, or conversing with an attacker. We show that it can be driven from the host using only interpreters that ship with the operating system and find that the one-time user-activation check before the model is downloaded is easily satisfied with a simulated click or bypassed entirely.We measure and report the current model's ability to identify credentials on the host, detect PII in email, classify images, and call custom tools. We combine these capabilities and use the model to carry out a simulated cyber operation: reading a password from a webcam image, unlocking an encrypted archive, triaging the contents, and staging a private key for exfiltration, end-to-end without user-interaction.

cs.CR · cs.AI · cs.CL

版本更新 · v2arXiv:2609.24974

Harness-Zero: Harness Distillation via Agent-as-Harness

Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song

原文摘要

Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instances, and models, a general-purpose agent must either settle for a suboptimal shared harness or route among an ever-growing set of specialized ones. We therefore study agent harness distillation: using a domain- or instance-optimized harness as training-time guidance and transferring the behaviors it induces into model weights, so that its gains survive under a single fixed target harness. The challenge is that the two harnesses differ in action space and available information, so guidance from the optimized harness cannot serve directly as supervision for the target one. We introduce Harness-Zero, which enables harness distillation through agent-as-harness. Guided by the optimized harness, a harnessing agent corrects student responses before execution in the target harness's action space, turning harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behavior into the model, so the specialized harness can be removed at deployment. Our experiments spanning knowledge work, tool use, and science domains show that: (1) For frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness. (2) With the specialized harness removed at deployment, Harness-Zero improves the base model's macro-average task success from 23.3% to 44.3%, even exceeding the 41.7% it reaches with that harness still attached, while reducing inference token usage by 27%-86%. (3) Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains.

cs.AI · cs.CL · cs.NE

新论文 · v1arXiv:2611.25071

FinVeriQA: Financial Research Agents for Question Answering: Routing, Retrieval, MCP Tool Use, and Provenance-Aware Evidence Verification

Meenakshi Rajpurohit, Anna Phan, Ravikumar Komandur Narayanan, Gunanidhi Ramakrishnan, Nathan Howland

原文摘要

Large language models (LLMs) write fluent financial analyses but miscompute ratios, recall stale prices, and route questions to the wrong tool with high confidence. We ask which component should do each job in a financial question-answering assistant, and study it with FinVeriQA, an open-source multi-agent research prototype over SEC filings and live market data: a compact typed-decision model (Laya) routes each question and screens out advice requests, Python computes every number, retrieval and tools fetch evidence, Mistral-7B writes a cited answer, a code validator checks that cited IDs exist and that numbers match computed notes, and a Model Context Protocol (MCP) server exposes the pipeline. In live experiments, Laya routed better than Mistral-7B (0.85 vs. 0.52; overall 0.86 vs. 0.68, McNemar p = 0.0003); a 12B Mistral decided no better, and DeepSeek-R1-8B decided better (0.79 vs. 0.68, p = 0.0192) but about 9.9 times more slowly. Mistral-7B computed 29% of 100 financial calculations correctly against 100% for Python, a live quote cut price error from 210% to 0.00%, and hybrid retrieval reached 76% Hit@5 on 50 filing questions. Yet in a live case study, both agent answers passed validation while describing AMD sample data under NVIDIA's and Alphabet's names. A controlled upper-bound study on the FinRank benchmark explains why: when one figure in the shared evidence is corrupted, self-consistency stays confident (agreement 0.55 to 0.59 while accuracy falls from 0.50 to 0.00), and a shared-evidence check (the kind our validator performs) catches 35% of the errors, whereas a check against the original passage catches 77% ($p = 5.96 \times 10^{-8}$); for DeepSeek-R1-8B, 68% vs. 90%. Verification must be independent of the evidence the writer saw.

cs.CL · cs.AI · cs.IR · cs.MA

新论文 · v1arXiv:2611.24936

基于对比原型结构的大语言模型越狱检测

Jailbreak Detection on Large Language Models via Contrastive Prototype Structures

Yan Qiu, Yuanyuan Li

AI 速读 基于摘要

JDPI利用对比学习表示和原型距离进行越狱检测,无需额外下游分类器,在多个基准上表现优于基线。

cs.CL

查看原文摘要

Jailbreak attempts and harmful prompts remain a challenge for the safe deployment of large language models. We study whether contrastively learned representations can support safety detection without training an additional downstream classifier. We propose Jailbreak Detection with Prototype Inference (JDPI), which uses binary safety labels to construct harmful and benign prototypes and makes predictions from relative prototype distances. JDPI uses only binary safety labels for encoder adaptation, prototype construction, and validation. In our setup, it uses fewer labeled examples than the classifier baseline (Method A) by avoiding the separate split used to train the downstream classifier. Across four benchmarks covering different input formats, contrastive fine-tuning improves JDPI's detection accuracy by 9.1-18.0 percentage points. Using the same encoder, JDPI achieves slightly higher ROC-AUC on one benchmark and stays within 0.017 of the baseline across all four. With dataset-specific adaptation, JDPI exceeds the released Llama Guard 3-8B baseline by 0.173-0.310 in ROC-AUC across the four benchmarks. In multi-turn replay on TomGibbs, JDPI uses accumulated user history as the conversation unfolds, achieving higher conversation-level accuracy than when it uses the current turn alone. Across 12 directed transfers, JDPI outperforms the classifier by an average of 1.59 percentage points in accuracy, demonstrating stronger cross-dataset robustness.