跳到正文
研踪PaperTrail

YOUR RESEARCH, IN SIGHT

让下一篇重要论文，不再错过。

覆盖计算机全学科，以 AI 为快捷入口。关注方向、作者和关键词，把发现变成持续的研究积累。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1072新增论文
+845其中 AI 相关
+94已有论文修订
+100新增 AI 导读

arXiv 最近完成扫描：2026-11-09 14:26 · 全站最近入库：2026-11-09 13:22各信源进度 →

统计口径与数据新鲜度

统计 2026-11-08 14:40 — 2026-11-09 14:40（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：2026-11-09 14:26最近发现新增版本：2026-11-09 13:22库内论文最新更新时间：2026-11-07 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。

18,172 已收录论文15,506 AI 相关4,527 有修订112 HF 社区发现

浏览论文库

研究日报 / 周报 →按信源浏览 →

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 5,189 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-09 14:25（北京时间）
版本更新 · v2arXiv:2606.11806

External Experience Serving in Production LLM Systems: A Deployment-Oriented Study of Quality-Cost Trade-offs

Lin Sun, Heming Zhang

原文摘要

Production LLM systems accumulate reusable operational experience, but the practical deployment issue is not merely whether such experience can help. It is how different serving strategies trade off quality against online cost under realistic constraints. Injecting external experience can improve task quality, yet it also increases prompt burden, latency, and serving pressure. We study \textit{external experience serving} as a deployment-oriented quality-cost trade-off problem. We evaluate this question in a real production moderation setting, with tool-use and GPQA as supporting contrast tasks that expose different output-cost regimes. We compare no-experience baselines, random experience controls, global prompt injection, and retrieval-based selective injection, and analyze both task quality and serving cost. The results show that, once experience becomes case-dependent, selective retrieval provides a stronger operating point than unconditional global injection. They further show that retrieval quality matters more than simply increasing Top-$K$, and that the same serving policy can exhibit substantially different cost-benefit profiles across short-output and decode-heavy regimes. These findings suggest that external experience is best treated as a selective, cost-aware serving decision rather than as a universal add-on. Overall, in the settings studied here, external experience pays off only when both the serving interface and the task-specific cost structure make its quality gains worth the online cost.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02368

Beyond Accuracy: How Quantization Affects Coding Agents

Areej Mehboob, Muhammad Ali Shafique, Muhammad Usman Qadeer, Hamza Farooq

原文摘要

Agents are increasingly used for coding tasks, but the inference cost of large language models (LLMs) limits where they can run. Quantization reduces model size, but its effects on coding agents that execute and revise programs remain underexplored. We compare BF16 with 4-bit AWQ and GPTQ for Qwen2.5-Coder-Instruct (7B, 14B, and 32B) and Devstral-Small-2507 (24B) on OJBench coding problems across three seeds. We detect no statistically significant accuracy loss for any of the coding agents. However, final accuracy can conceal substantial behavioral changes: AWQ reduces Qwen-7B one-shot solves from 17 to 0 while increasing loop-based solves from 10 to 25. All Qwen-32B successes involve tool use, whereas Devstral rarely makes a valid executor call. Failure profiles also differ: Qwen-32B produces far more compilation errors than the smaller Qwen models, rising slightly under quantization, while Devstral's failures shift significantly from runtime errors toward wrong answers and timeouts. These findings indicate that quantization robustness in coding agents should be assessed through solution trajectories, tool use, and failure modes, in addition to final accuracy.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02353

VisCodeBench: a new benchmark dataset for Text-to-Vis reflecting real-world practice

Hy Nguyen, Thomas James Elton, Ella Wilson, Andrew Reeson, Cecile Paris, Josiah Poon, Jonathan K. Kummerfeld

原文摘要

Generating visualisations from natural language requests (text-to-vis) enables users to explore data more efficiently. While LLMs can complete simple visualisation requests, their performance on challenging ones is not well covered by existing benchmarks. We introduce VisCodeBench, an end-to-end text-to-vis benchmark with 2,672 samples from three different domains: educational, scientific, and general audience scenarios with varying levels of complexity. We benchmark three visualisation systems---Codex, Claude Code, and LIDA---paired with a range of open and commercial LLMs across seven metrics covering execution, chart content, and visual quality. While basic execution is near-solved for strong systems, no system comes close to reproducing the reference data, with the best displayed data similarity reaching only $0.81$ and falling to $0.69$ on our hardest subset for this metric. We further find that frontier models separate on chart correctness and efficiency, with ClaudeCode-Opus-4.7 improving chart correctness by around 5 points over Codex-GPT-5 while running roughly 20 seconds faster. Separately, we find that system design matters as much as the underlying model: despite both running on GPT-5, Codex clearly outperforms LIDA, showing that the surrounding prompting, execution, and refinement pipeline can shape outcomes as much as the model itself. Finally, latency remains a practical obstacle, as even the strongest systems take 39-78 seconds to produce results. VisCodeBench is available at \href{https://github.com/SAIPLL/VisCodeBench}{https://github.com/SAIPLL/VisCodeBench} and accessed via a leaderboard; the full dataset is kept private to prevent LLM training use.

cs.CL · cs.HC

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02348

Which Words a Model Fears: From a Refusal Direction to a Vocabulary-Wide Fear Lexicon

Boyuan Zhang, Mohammad Zandsalimy, Shanu Sushmita

原文摘要

Whether an aligned language model refuses depends in part on which words a prompt contains, regardless of whether those words carry the request's meaning. Harm words shift the hidden state at the generation boundary along a component shared across prompts and words, which crossed differencing extracts as a single direction. In eight models, writing along this direction raises and lowers refusal monotonically with dose, ahead of random and KL-matched directions. Used as a ruler over the full vocabulary, the direction gives every word a shift, from "feared" words at one end to "courage" words at the other; a word's shift is predictable from its own representation, and appending a feared word to a benign prompt is enough to make the model refuse it. Two held-out tests validate the ranking in all eight models: refusal after inserting an unseen word rises with its table shift, and replacing a prompt's most feared word by its most courageous meaning-preserving synonym significantly raises the answer rate on benign and harmful prompts alike, by an amount that follows the shift the replacement actually causes.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02338

MemShift: Separating Detection from Repair in LLM-Derived Personal Memory

Zichuan Xu, Haochen Zhang, Yuyang Dong, Makoto Onizuka, Chuan Xiao

原文摘要

AI assistants store values that a language model derived from a person's records under the person's own definitions, such as their important contacts. Keeping such a value correct takes two steps: detection, noticing that an input changed, and repair, producing the correct new value. Benchmarks that score only the final answer cannot tell which step failed, and they state every change as new text. Seeing this limitation, we present MemShift, a benchmark of five shifts: source revision, deletion, redefinition, writer switch, and export. The first three fork a synthetic world into branches that differ only in the shift. Gold values and dependencies are fixed before any text, and a deterministic scorer replaces an LLM judge. For a memory that reports its invalidations, MemShift shows for each changed item whether the memory noticed the change and whether its repair was right. Moreover, we build a reference memory with known detection and repair to validate the measure. By evaluating four open-weight LLMs, we find that detection helps only as much as repair allows: the LLMs re-derive few values computed by the person's definitions correctly, and programs that the LLM writes over typed records close most of the gap to exact recomputation. Code and data are available at https://github.com/zichuanxu/memshift.

cs.CL · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02333

When Does Adaptive Halting Win? Selection Rules Matter in Looped Language Models

Mahesh Reddy Pagadala

原文摘要

Looped language models can stop iterating early, and recent work shows that post-hoc readouts of the loop trajectory often rival the learned exit gate of the looped model Ouro. Such comparisons reduce each method to one operating point chosen by a selection rule. We ask whether that rule decides the outcome. On a frozen Ouro-1.4B with four loops, we compared a trained stability probe ($π_1$-x) with Ouro's native gate, adapted to a per-question stopping decision, in pre-registered studies; the confirmatory study was designed after an exploratory finding on development data and evaluated once on a sealed test set of 8,517 questions. In the pre-registered confirmatory comparison, which selected operating points with a bootstrap-screened 1 pp accuracy-loss rule, the probe saved more of the four-loop budget than the native gate (+0.053 gross macro saving over three benchmarks, 95% CI [+0.049, +0.057]), mostly from CommonsenseQA, where the probe's test upper bound on accuracy loss exceeded 1 pp. Under a point-only rule, a pre-specified secondary analysis on the same test set reversed the preference (-0.022 [-0.026, -0.018]), with operating points that exceeded the budget. In pre-specified secondary comparisons, the probe showed no demonstrated advantage over an untrained KL divergence between successive loops (+0.00125 [-0.0041, +0.0065]); we emphasise this result post hoc and claim no equivalence. The probe also broke its budget on one of two out-of-distribution benchmarks (3.26 pp loss on Social IQa). Savings are estimated from stored full-depth traces, not measured inference time, and our screen carries no formal risk-control guarantee. We conclude that halting methods should be compared as signals together with selection procedures, with strong untrained baselines and out-of-distribution checks.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2608.25654

When Answer Evaluation Misleads Confidence Selection: Fact-Grounded Form Grading for Open-Ended QA

Zhexi Feng, Bingrui Zhang, Wuxi Chen

原文摘要

Many benchmarks now prefer open-ended questions for testing LLM and NLP methods. We find that the choice of a calibration rule can flip when only the correctness labels change: on 141 pilot answers, rules chosen with exact-match labels look 0.239 Brier points better than raw confidence but are 0.103 worse under human labels. How such flips arise remains unclear. We look deeply into their mechanism: reference matching rejects many correct answers (exact match misses 42%), and an exact Brier identity shows these errors favour rules giving such answers lower confidence, which abstain on them at common thresholds. Guided by this, we propose FactForm, a grader that trusts reference matches and otherwise consults evidence retrieved from the question alone. One LLM call per answer fills a fixed-order closed form that a program scores. On 998 fresh EVOUNA answers with 400 fact-based human labels, FactForm reaches a balanced error of 0.077 and an AUROC of 0.985, against 0.112 and 0.930 for the best of eight traditional graders. Its same-model stability is no worse than that of LLM judges (mean absolute run-to-run difference 0.0136 against 0.0159 to 0.0206). We hope this helps benchmarks grade open-ended answers more accurately, with repeatable and discriminative scores.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2608.29455

Item-Mean Surrogates: Why Richer Personas Fail to Deliver Individual Fidelity in LLMs as Human Surrogates

Daehwan Ahn, Chengfeng Mao, Dokyun Lee

原文摘要

LLMs are increasingly used to simulate and study human behavior. Such studies assume richer personas improve predictions of specific people's answers. We test this premise across four public datasets. First, we split each person's answer to an item into the item mean (the average human answer to that item) and the person's deviation from the item mean. Much of the accuracy that surrogate studies report comes from predicting item means. LLM predictions explain little of the deviations. We call this pattern item-mean surrogacy. Second, we split each deviation into stable signal and noise using a dataset in which the same participants answered the same questions twice. Stable signal is the part of a deviation that repeated answers share. Noise is the part that changes. In this dataset, more than half of the variance in deviations persists across occasions. Most of what the LLM misses is stable signal rather than noise. Third, we split stable signal into the person signal, a person's average deviation across questions, and the person-by-item signal, their remaining question-specific deviations. The person-by-item signal accounts for about 8.5 times as much variance as the person signal. The LLM recovers about a tenth of it. Richer personas, different models, and fine-tuning on human answers still leave most of the person-by-item signal unrecovered. Both fine-tuned models recover less of this signal than their untuned versions. These findings call for closer empirical scrutiny of LLMs as human surrogates.

cs.CL · cs.CY · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02332

LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents

Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang 等 20 位作者

原文摘要

Tool-using language-model agents often execute actions that their domain policies forbid. We present LeanGuard, a runtime that checks every proposed tool call against a policy written in Lean4 before the call reaches the environment. Policies combine state predicates with scoped past-time temporal operators over the execution history, evaluation is fail-closed, and the decision procedure is proved equivalent to a LeanLTL semantics. A denied call can be corrected in two ways. With feedback, the agent receives the violated rule and a hint; with proof-guided repair, Lean metaprogramming searches for a nearby call that preserves the user's intent and proves that the policy admits it, without the model. On Tau-Bench with five Qwen models (0.8B-27B), unguarded agents execute at least 33 calls that the Lean policy forbids, while LeanGuard executes none; with feedback, task success does not change significantly. On DTap, LeanGuard reduces direct prompt-injection success from 57.9% to 5.9%, and in telecom, where benign tasks need no writes, benign success is unchanged. Feedback reduces repeated blocked calls and raises task success for four of five models over a bare denial. On replayed denials, repair proves an admissible edit for 88 of 252 denials caused by argument values, in a median of 2,ms.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.29207

norregel: a preregistered benchmark of dated numeric values in Norwegian law

Eirik Botten Nicolaysen

原文摘要

norregel asks language models for numeric values in Norwegian statutes, regulations and parliamentary decisions (NOK amounts, percentages, multiples of the National Insurance basic amount (G) and age limits), each at a stated date. The 916 items were extracted mechanically from Lovdata's open data, and each carries the dated history of its provision, so an answer can be classed as correct, outdated, wrong or uninterpretable. We report a first and a second preregistered run on the same items. In the first run (five models, listed publicly only after local analysis), H1, that outdated answers are more frequent for values that took effect after a model's cutoff, could not be tested in substance, and H2 and H3 were inconclusive; Opus 5.5 and Sonnet 4.6 answered 46.7% and 40.2% of the 880 primary items correctly. The second run was preregistered and published before any of its calls, and eight models were analysed. Its pooled primary test supported H1 under the preregistered decision rule (d = 17.8 percentage points, 95% interval [7.7, 28.9], p = 0.0001). In the preregistered sensitivity analysis S8, which counts as outdated only the most recent value that differs from the key, d was 3.9 [-3.6, 12.7] (p = 0.1754), an interval that contains 0. H1 was supported in every other preregistered sensitivity analysis, but S8 is the only one that gives both groups one target value, so we do not regard the support for H1 as robust. The share of outdated answers equal to that most recent value (H2) was 56.1% [46.6, 65.0], an interval that contains 0.5, and the pooled test of a lower share correct in Nynorsk than in Bokmål (H3) was inconclusive. Several models, among them Haiku 4.5, gave mostly uninterpretable answers. Items and keys are withheld; item IDs, per-answer classes, code and results are on Zenodo.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02319

Accessed, never disclosed: Silent access to protected health information by medical artificial intelligence agents

Jae Wang Lee, Abhirup Banerjee, Yeong Chan Lee

原文摘要

Medical artificial intelligence agents should retrieve patient identifiers only when a task requires them, the data-minimization principle behind the HIPAA minimum necessary standard for protected health information (PHI). An agent can retrieve identifiers without revealing them. Such silent access remains invisible to evaluations that inspect only the output. We embedded canaries in synthetic patient records and linked each tool invocation to canary exposure in the final answers across five language models, four boundary levels, and seven attack frames for identifier-free clinical tasks. Three models invoked the PHI tools unnecessarily, whereas Claude Opus 4.8 and GPT-5.5 did not. Stronger boundaries reduced this access, and the residual access under the strictest policy was confined to frames in which the user overrode the policy or asserted approval. A pre-execution gate given the task-necessity criterion nearly eliminated this access. Output-only evaluation cannot verify minimum-necessary access; boundaries, gates, and access-log audits are required together.

cs.AI · cs.CL · cs.CR · cs.CY

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02318

A Capitulation Direction Emerges With Scale and Is Causally Load-Bearing Where Validation Is Strong: Eight Models, Two Families, and Five Measurement Hazards

Saad Aamir, Muhammad Awais Bin Adil

原文摘要

Sycophancy in language models is well documented as a behaviour but poorly localised as a mechanism. We ask whether capitulation under user pushback, the abandonment of a correct answer when a user objects, corresponds to a linear direction in the residual stream, and whether such a direction emerges with scale. Across eight instruction-tuned models spanning 0.5B to 14B parameters in two families, under a pre-registered protocol with leave-one-question-out cross-validation, within-question permutation nulls, and a positive control, we find no usable direction below 7B and a direction that clears our validation gate at 7B, 8B and 14B. Ablating that direction in Qwen2.5-7B reduces the rate at which the model reverses a correct answer by 4.8 percentage points, roughly a quarter in relative terms, while the identical intervention on a gate-failing direction and on a marginally passing one produces no effect. Probe quality, not statistical significance, predicts causal relevance. Behaviourally, pushback destroys three to five times more correct answers than it repairs at 1B and reaches parity by 7B, falsifying our own pre-registered prediction, and a sign-reversing crossover between pushback types survives an order of magnitude of scale in Qwen and holds at 8B in Llama, with marginalisation over five paraphrases per type. We additionally quantify five measurement hazards, including a numerical-precision effect that changes roughly one episode label in ten.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02314

GRIPS: An Empirical Study of Pre-SQL Grounding Information for SQL Generation

Ekata Mitra, Yufei Tao, Kamalamalini Nagasundaram, Ameeta Agrawal

原文摘要

Text-to-SQL systems answer natural language questions over structured data, but can fail before SQL generation by linking question mentions to the wrong table, column, value, or relation. Even valid SQL can then return an incorrect answer. This raises a central question: What pre-SQL information helps a SQL generator, and when does additional information cease to help? We investigate this question on TempTabQA-C and BIRD. Among the audited SQL Static failures, grounding errors form the largest category on both benchmarks. We introduce GRIPS (Grounding Representation and Input Processing for SQL generation), a three-stage method that derives structured pre-SQL information to make these decisions explicit before query generation. Across GPT-5-nano and Llama-70B, GRIPS-Adaptive improves over SQL Static in 18 of 20 settings, with an average accuracy gain of 11.8 percentage points, while reducing model-call token usage by 47.8%. Our controlled analysis shows which grounding information can improve SQL generation, whereas adding more pre-SQL information does not consistently help. By separating question-input grounding from SQL generation, GRIPS makes text-to-SQL more accurate, efficient, and easier to diagnose.

cs.CL

查看论文
arXiv ↗PDF ↗
首次观测 · 记录 #1软件学报:jos:7771软件学报

中文生成式引擎优化

李彤, 吴荣榜, 李清昊, 贾贵, 高翔宇, 范举, 徐恪

原文摘要

大语言模型(large language model, LLM)和生成式引擎正在改变用户获取网络信息的方式. 传统搜索引擎优化主要关注网页能否进入搜索结果并获得较高排名, 而生成式引擎优化则关注网页内容能否被大语言模型采用并呈现在回答中. 然而, 现有研究主要面向英文场景, 面向中文网页和国内生成式平台的系统性生成式引擎优化 (generative engine optimization, GEO)研究仍然缺失. 为此, 构建面向中文场景的生成式引擎优化评测框架, 围绕中文数据集构建、生成式引擎优化策略设计和来源可见度评测开展研究, 并从策略、平台、领域和来源位置等方面分析中文生成式引擎优化的效果. 在此基础上, 进一步比较中文生成式引擎优化与英文生成式引擎优化的差异, 并总结面向中文场景的实践启示.

cs.* · cs.CL

软件学报 · 2026 · 正式发表

查看论文
软件学报 ↗PDF ↗
版本更新 · v2arXiv:2604.16656

Defragmenting Language Models: An Interpretability-based Approach for Vocabulary Expansion

Maitrey Mehta, Nishant Subramani, Zhichao Xu, Ashim Gupta, Vivek Srikumar

原文摘要

All languages are equal; when it comes to tokenization, some are more equal than others. Tokens are the hidden currency that dictates the cost and latency of contemporary LLM access. However, many languages written in non-Latin scripts observe a poor exchange rate: LLMs take several multiples of tokens to encode the same information in many languages as they do for English. We observe that this issue, known as token overfragmentation, persists in modern open-weight LLMs. The standard remedy is vocabulary expansion that adds target language items missing from the model's vocabulary. We focus on two core decisions in the vocabulary expansion process: What items should we add? and How should we initialize their corresponding input and output embeddings? We comprehensively study and advance the new research direction of interpretability-based vocabulary expansion along these two core decisions, making four key contributions: First, we question the conventional use of frequency-based methods to choose candidate vocabulary items to add (a decision long treated as settled), and show that interpretability-based methods offer a superior performance-token efficiency tradeoff. Second, we strengthen the case for interpretability-based embedding initialization by showing large gains (~20 pts) over baseline initialization methods for several languages written in non-Latin scripts. Third, we identify the phenomenon of "subword detokenization," where models progressively merge fragmented subword tokens into larger subwords across layers. Fourth, we propose FragMend, a method that leverages this phenomenon to push the efficiency ceiling of interpretability-based expansion, and validate its effectiveness.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2602.05842

Reinforcement World Model Learning for LLM-based Agents

Xiao Yu, Baolin Peng, Ruize Xu, Yelong Shen, Pengcheng He, Suman Nath, Nikhil Singh, Jiangfeng Gao 等 9 位作者

原文摘要

Large language models (LLMs) have achieved strong performance in language-centric tasks. However, in agentic settings, LLMs often struggle to anticipate action consequences and adapt to environment dynamics, highlighting the need for world-modeling capabilities in LLM-based agents. We propose Reinforcement World Model Learning (RWML), a self-supervised method that learns action-conditioned world models for LLM-based agents on textual states using sim-to-real gap rewards. Our method aligns simulated next states produced by the model with realized next states observed from the environment, encouraging consistency between internal world simulations and actual environment dynamics in a pre-trained embedding space. Unlike next-state token prediction, which prioritizes token-level fidelity (i.e., reproducing exact wording) over semantic equivalence and can lead to model collapse, our method provides a more robust training signal and is empirically less susceptible to reward hacking than LLM-as-a-judge. We evaluate our method on ALFWorld and $τ^2$ Bench and observe significant gains over the base model, despite using no expert data, task-success rewards, or strong-LLM annotations - only an off-the-shelf embedding model as a distance metric. When combined with task-success rewards, our method outperforms direct task-success reward RL by 6.9 and 5.7 points on ALFWorld and $τ^2$ Bench respectively; it further outperforms expert-data training on ALFWorld and is comparable to it on $τ^2$ Bench.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v4arXiv:2609.37788

A Proposed Rubric for Evaluating Expressed Clinical Reasoning in Large Language Model Responses

Zhangshu Joshua Jiang, Zina Ibrahim, James T. Teo

原文摘要

We propose a rubric for assessing expressed clinical reasoning in model responses, drawing on three bodies of work: medical education assessment frameworks (ART, SCT, Key Feature Problems and OSCE); clinical LLM benchmarks (MedR-Bench, HealthBench, TIMER-Bench, DR. BENCH, PrIME-LLM and PatientSafeBench); and general LLM reasoning evaluation research, including the Factuality-Validity-Coherence-Utility taxonomy, FaithCoT-Bench and C2-Faith. We use groundedness as a clinically oriented adaptation of the taxonomy's factuality category. The rubric brings these concepts together in a multidimensional framework for scoring free-text responses to gold-standard clinical vignettes. It includes provisional behavioural anchors, applicability rules and a separate flag for case-specific safety-critical errors. General-domain frameworks inform its design but are not treated as validated clinical instruments. The rubric does not replace case-specific reference criteria or the task-specific metrics of existing benchmarks. It has not yet been tested for inter-rater reliability, construct validity or clinical utility. Its immediate purpose is to make evaluation decisions explicit and open to scrutiny before empirical testing.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02278

Concise-9B: Removing the Slack from a Reasoning Model's Thinking

Alexander Boesgaard Lorup

原文摘要

Reasoning models answer better when they think longer, but a thinking trace keeps going well after its answer has formed. We measure this slack in Qwen3.5-9B and remove most of it with concise self-training: two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, giving Concise-9B. On 8,983 held-out problems (competition and hard mathematics, SuperGPQA, a five-area confirmation set and MMLU-Pro), Concise-9B generates 0.66x the tokens of Qwen3.5-9B at a 32K thinking budget, with accuracy equivalent within +/-1 point (-0.20 [-0.77, +0.36] points). Forcing each trace on the hard and confirmation sets to answer every 2,048 thinking tokens, we find that 54-64% of Qwen3.5-9B's thinking tokens come after the point from which every forced answer equals the final one (41-48% for Concise-9B). Stopping there would keep every answer at 0.35-0.44x the base model's tokens, but finding that point takes hindsight: stopping when three consecutive forced answers agree costs the base model 2.4-9.2 points. Concise self-training removes slack without that cost: 86% [83, 90] of the thinking tokens it saves come after the answer has formed. It has a limit: with traces continued to 80K tokens, Concise-9B solves 25 fewer of 1,276 hard problems (-25 [-49, -1]), mostly in competition mathematics. The measurement also yields a cheap commit signal for adaptive sampling that matched two-sample escalation within +/-1 point at 0.67x its tokens in a pre-registered test. Results on Qwen3.5-4B, on which the recipe was developed, and a replication on Olmo-3-7B-Think suggest that the saving depends on how verbose the base model is.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2605.06914

Pacing Branch Parallelism in LLM Serving

Swapnil Gandhi, Siva Kumar Sastry Hari, William J. Dally, Christos Kozyrakis

原文摘要

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the \emph{branch externality} and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce PACE, a per-step admission controller that treats extra branches as opportunistic width, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, PACE improves goodput by $1.77\times$ over serving without branch parallelism and by $1.48\times$ over eager admission while maintaining 99\% SLO attainment. Comparable gains ($1.72$--$2.15\times$) hold on five additional dense and mixture-of-experts models with up to 671B parameters. Code is available at https://github.com/Swapnil-Gandhi/PACE

cs.DC · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02273

When the Text Is Silent: Memory Shortcuts and the Validity of LLM Annotation under Evidence Insufficiency

Wenchang Wang, Yue Song

原文摘要

Large language models (LLMs) are increasingly used to annotate text for social science research, content moderation, and training data. Their broad knowledge of real-world entities raises a concern: when labeling a text, do they rely on its evidence or substitute what they already know? We call this behavior a memory shortcut and examine when it emerges, what amplifies it, and how to mitigate it. Using 995 widely recognized public figures and four matched control names per figure, we construct counterfactual texts that separate textual evidence from prior knowledge, and evaluate three commercial LLMs on 121,594 prompts each across five stages. We find a striking asymmetry. Although all models recognize nearly every figure, none follows memory over the text when the two conflict. Shortcuts instead emerge when evidence is absent: for one model, famous names raise the rate of remembered occupations, ages, and countries by 17.0, 14.9, and 6.8 percentage points relative to synthetic names. These effects occur only for recognized figures, and 98.6% of the resulting labels are factually correct about the namesake while none is supported by the text. Evaluation threats and answer-contingent rewards amplify shortcuts by 8.7 and 12.4 points. Famous names also raise leadership ratings of sparsely documented applicants by 0.22 and 0.28 points in two models. A namesake reminder and name masking each eliminate the shortcuts, with trade-offs in gender annotation. LLM annotations can thus be factually correct yet methodologically invalid. Trustworthy LLM annotation requires audits that test evidence absence, account for incentives in annotation instructions, and evaluate the attribute-specific trade-offs of mitigation.

cs.HC · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02271

Current state of LLMs for Arabic dialectal machine translation

Josef Jon, Rawan Bondok, Ondřej Bojar

原文摘要

This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity, although the results may be influenced by possible training data contamination, which is always a concern with LLMs.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02270

Hinted Scaling Laws

Suze van Adrichem, Luke Bailey, Rohith Kuditipudi, Christine Ye, Neil Band, Ahmed Ahmed, John Duchi, Percy Liang 等 9 位作者

原文摘要

Scaling laws predict the outcome of large, expensive training runs by extrapolating trends from smaller, cheaper runs. On downstream accuracy evaluations, these predictions fail when the weaker models produced by the cheaper runs score near zero accuracy, because these scores give no trend to extrapolate to larger scales. We introduce hinted scaling laws, which use privileged information to turn a benchmark into a continuum of easier versions so that even the small models score above zero. We find that accuracy rises smoothly and predictably with the hint level, enabling new types of scaling laws that operate below the noise threshold. On the Delphi pretraining ladder, a hinted scaling law fit on models trained with at most $10^{20}$ FLOPs forecasts the GSM8K accuracy of a $10^{23}$ FLOP model (25.9%) to within 3 percentage points; a standard scaling law under-forecasts it by 13 percentage points when fit on the same models and by 12 even with models up to $10^{22}$ FLOPs in the fit. Across both this pretraining ladder and a collection of post-trained model families, hinted scaling laws reduce the median forecasting error on GSM8K, AIME and HLE from 21 percentage points to 4.3. Hinting also reduces the number of test examples needed to tell whether one model is better than another by widening the accuracy gaps between models while approximately preserving their order: a single hinted rollout per problem ranks nearly indistinguishable models as accurately as five unhinted rollouts. Our results suggest that hinting is a general primitive for improving the accuracy and cost of downstream scaling laws and other prediction tasks in model development.

cs.LG · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2608.15428

Gated Against One Model, Open to the Next: Option-Only Solvability in Legal Multiple-Choice Benchmarks

Volodymyr Ovcharov

原文摘要

Multiple-choice benchmarks are graded on whether a model picks the right option, not on whether it needed the question. Measuring that gap takes care: a model answering A to most items scores above chance wherever the key sits at A, and reads as recognition when it is not. We measure it on UA-JudgeExam: 11,990 four-option items with official keys, published by Ukraine's Higher Qualification Commission of Judges. Shown the options and no question, Claude Haiku 4.5 scores 0.383 against chance, and the leak is concentrated: 11.8% of items are answered blind on all eight option orders, against 0.2 items expected by chance. It is not quotation: search over 280,059 editions of Ukrainian legislation recovers 0.128. Gating those out retains 8,128 items, on which the gating model itself now scores 0.204, while GPT-5.6, which took no part in the selection, still answers 0.515 of them with the question hidden. Scoring eighteen held-out models from eight vendors on the whole set and subtracting each one's position habit, four keep a sizeable excess: GPT-5.6 (+0.265), Gemini 3.1 Pro (+0.240), GPT-5.4 (+0.105), Sonnet 4.6 (+0.081); the rest stay within 0.031, and probes find no recall of the public bank behind it. Without the correction Llama 3.1 8B (0.292 blind, A on 92% of items) would read as a leaker. The gate does select something real: on the items it rejected, eleven of twelve models score 0.518-0.789, every interval clear of what the same model scores on the items it kept. But that signal is one model's, and filtering on it does not transfer upward. Rewriting distractors instead overshoots to 0.168, below chance and as exploitable. The same probe on LEXam returns chance: every option there points into the stem, none longer than 33 characters. Item format decides whether the problem can arise; capability decides how much is extracted. We release the corpus, predictions and harness.

cs.CL · cs.AI · cs.CY

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2511.08274

VeriCypher: Named-Entity Verification for GraphRAG via Text-to-Cypher Generation

Anton Gusarov, Anastasia Volkova, Valentin Khrulkov, Andrey Kuznetsov, Evgenii Maslov, Ivan Oseledets

原文摘要

Knowledge graph question answering has traditionally focused on Resource Description Framework (RDF) graphs, SPARQL queries, and triple-based representations. Labeled Property Graphs (LPGs), however, pose a different challenge for LLM-based query generation: executable Cypher queries must align with schema structure, traversal direction, property keys, and literal entity values. This motivated us to introduce VeriCypher, an LLM-based text-to-Cypher framework for question answering over LPGs. It combines query generation conditioned on Cypher-like graph schema summaries with runtime execution, LLM-based query evaluation, and iterative correction where the central mechanism is named-entity verification: the system extracts node labels, relationship patterns, property keys, and literal values from generated queries, checks them against the target graph and schema, and turns mismatches into correction instructions for the next generation step. We evaluate VeriCypher on the CypherBench benchmark across multiple domains and LLM backbones. Compared with an LLM baseline, this method consistently improves answer correctness and query robustness. Ablation results further show that named-entity verification helps correct hallucinated schema elements and entity values, supporting more reliable LLM-based question answering over heterogeneous property graphs.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02248

Metamorphic Testing for Evaluating Robustness of Sentiment Analysis Models

Sowjanya Karanam, Jayanth Bhargav

原文摘要

Machine learning systems are challenging to test using traditional software testing methods because their behavior is learned from data rather than explicitly programmed. This challenge is particularly significant in natural language processing (NLP), where small input variations can lead to unexpected model behavior. Conventional evaluation metrics, such as accuracy and F1-score, often fail to capture such inconsistencies. Metamorphic testing addresses this limitation by evaluating expected relationships between original and transformed inputs. With the increasing prevalence of generative artificial intelligence (GenAI) and language-based AI systems, the need for systematic robustness testing has become even more pronounced. Motivated by this growing need, this paper presents a lightweight, batch-based metamorphic testing approach for evaluating the robustness of such systems. In particular, we consider sentiment analysis as a representative NLP task and evaluate a pretrained classifier under four transformations: synonym replacement, typo injection, word-order variation, and negation. Using consistency score, flip accuracy, and confidence change, we identify strengths in handling semantic variations and vulnerabilities to textual noise and logical transformations. These findings highlight the value of metamorphic testing in identifying behavioral inconsistencies beyond conventional performance metrics and demonstrate its potential as a systematic framework for evaluating the robustness of language-based AI systems and guiding performance improvements before deployment.

cs.LG · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02244

Can AI Reviewers See Your Efforts to Improve Your Paper? A Large-Scale Study of Paper Revision Trajectories

Yaochen Wang, Chenguang Wang, Ming Li, Haomin Zhuang, Xinyue Zeng, Han Chen, Kwesi Adu Cobbina, Dang Nguyen 等 11 位作者

原文摘要

AI reviewers are increasingly used to assess scientific manuscripts, yet it remains unclear whether their judgments reflect the improvements authors make through revision. We investigate this question using PaperTrajectory-2026, the 2026 release of PaperTrajectory, a dataset of 6,959 paper trajectories across five AI conferences, with automatic annotations of changes along 14 dimensions and separate judgments of overall quality. The annotations show that major improvements emphasize empirical evidence and framing, while later revisions tend toward incremental improvement. Using independent assessments of manuscript versions from eight AI reviewers, we examine whether annotated major improvements receive higher scores, which revision dimensions are associated with score changes, and whether user opinions can outweigh observed revision gains. Every reviewer leaves scores unchanged on most major improvements, and agreement on the direction of score changes is weak. Revision dimensions show some reviewer-specific associations with score increases, but provide limited predictive discrimination and no consistent guide across reviewers. Even when a revision earns a higher neutral score, a brief negative user opinion can erase or reverse that advantage, with substantial variation across reviewers. These findings show that substantive revisions are not consistently reflected in review scores and that observed scoring gains can be vulnerable to user framing, motivating assessments that connect manuscript evidence to overall judgments and remain stable when no scientific evidence changes.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02230

Who Should Score the Cache? Mid-Depth Layers Are the Best Judges of KV-Cache Eviction

Ekaterina Seregina

原文摘要

Key-value (KV) cache eviction methods rank tokens by attention scores, and each layer scores its own cache. To our knowledge, no eviction method asks which layer should score. We find, across twenty-one models, that the depth of the scoring layer governs eviction quality. On the eighteen models with per-layer caches, spanning 0.5B to 109B, long-context retrieval accuracy collapses at shallow layers, peaks between 0.4 and 0.75 of depth, and degrades at the last layer. The curve arises because a layer judges tokens well only once it has used the question to find the answer, which takes several layers. Modeling eviction as a top-$k$ cut, we compute answer survival probability from prefill attention and reproduce the measured curves. This approach selects the scoring layer using calibration prompts with known answers, and the selected layer is the best layer measured on most models and never significantly below it. It improves on the own-layer default on sixteen of the eighteen models, with significant gains on seven, all at 14B or below, while on larger models the default is already close to the best mid-depth layer. Scoring every cache from the selected layer while keeping one set of tokens for all heads also beats per-head SnapKV by 18 to 43 points on five models up to 14B under tight budgets. This gain carries over to financial filings at 32K context, all at the same memory budget, with no extra forward pass. The same curve explains a penalty in shared-cache models, where several layers read one cache and the layer that writes it scores by default. On Gemma-4-E2B the writer is shallow and loses 28 points to a mid-depth reader at a 10% budget. We also find that in open-book question answering (LongBench) no mid-depth layer beats the own-layer default.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2601.16724

Mitigating Bias in Automated Essay Scoring for ESL Learners via Contrastive Learning

Kevin Fan, Eric Yun

原文摘要

Automated Essay Scoring systems disproportionately penalize high-proficiency English as a Second Language (ESL) learners. We propose Contrastive Learning with Matched Essay Pairs (CL-MEP), a bi-directional alignment strategy. CL-MEP reduces this scoring bias by 39.9% while improving overall accuracy, successfully disentangling valid syntactic complexity from surface-level grammatical errors.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.02208

Test-Time Scaling for Open-Ended Reasoning at Small Model Scale

Adam Y. Ahmed, Abdullah G. Elafifi

原文摘要

A vote over a model's samples can only help when the samples disagree. For open-ended generation the binding question is therefore not how to combine samples but whether a model produces samples worth combining, a question test-time scaling has answered for math but not for free-form text. We study it at small scale with semantic cluster voting, a judge-free aggregator that groups a small instruct model's samples by meaning and returns the largest cluster's representative, with no reward model or self-judging. On a 250-prompt benchmark with three tiers of answer checkability, the same pattern holds across three model families (Qwen2.5-3B, Llama-3.2-3B, Gemma-2-2B) and two generation seeds: voting raises accuracy on the math control, where samples vary but stay checkable (a gain of +10 to +29 points across the three models), yet adds at most +0.30 judge points on open-ended text. On a larger 200-problem MATH-500 run with the primary model, the math gain is +8.5 points (95% CI [+4.0, +13.0]). Almost every open-ended sample falls into one meaning cluster, so there is little to select over. The collapse is consistent with instruction-tuning rather than model size, and raising decoding temperature does not recover usable diversity, because it costs more quality than it adds (net -0.64 judge points, 95% CI [-0.93, -0.37]).

cs.LG · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.21270

Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures

Zihao Zheng, Junyi Yao, Baichuan Li, Jiayu Long

原文摘要

Trustworthy biomedical AI requires reliable action selection when evidence is missing, stale, conflicting, or late, not only accurate answers on static inputs. We instantiate an episode-level evaluation protocol as 400 synthetic healthcare workflow episodes: 40 hand-authored semantic seeds across four task classes, two surface parameterizations, and five controlled variants, scored on ANSWER/ASK CLARIFICATION/ABSTAIN/ESCALATE under a pre-declared 5:1 missed-versus-unnecessary escalation cost. Reference actions are single-author study-policy labels, not clinician-adjudicated ground truth. Across five open-weight configurations (1.7B--9.7B), clean decision accuracy reaches 0.97, yet uncertainty-prompted models miss 32--97% of required escalations. The stronger 9.7B configuration narrows but does not remove the gap, falling from 0.97 clean accuracy to 0.78 under conflict. Verbalized confidence is inconsistent (ECE 0.10--0.54; failure AUROC 0.48--0.77), and a development-tuned confidence-only policy yields no utility gain. A transparent workflow-aware controller improves 5:1 utility for every model and reduces missed escalations to 0.12--0.24, at specificity 0.83--0.93. However, a 30-concept lexical stress test reduces each rule detector's recall from 1.00 on trigger phrases to 0.00 on unseen paraphrases, limiting the controller claim to this benchmark. Seed-clustered statistics, ablations, cross-task transfer, raw outputs, and manifests are packaged for audit. The results motivate explicit workflow-state evaluation while showing that apparently effective routing controls need their own generalization tests.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗