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

+1959新增论文
+1550其中 AI 相关
+212已有论文修订
+59新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-17 23:35各信源进度 →

统计口径与数据新鲜度

统计 2026-11-17 04:33 — 2026-11-18 04:33(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:2026-11-18 04:28最近发现新增版本:2026-11-17 23:35库内论文最新更新时间:2026-11-17 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。

25,786 已收录论文21,491 AI 相关7,725 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 26 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
版本更新 · v4arXiv:2602.02343

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

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

原文摘要

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

cs.CL · cs.AI · cs.CV · cs.IR · cs.LG

版本更新 · v3arXiv:2603.02578

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

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

原文摘要

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

cs.CL · cs.AI · cs.HC · cs.LG

新论文 · v1arXiv:2611.20227

Learning From the Best: What Makes Popular Hugging Face Models?

Yinan Wu, Zhou Yang, Bowen Xu, Bo Wang, Bach Le, David Lo

原文摘要

Hugging Face has become a central platform for sharing and reusing pre-trained models, yet it remains unclear which practices are associated with higher visibility and adoption. Prior studies describe model reuse and model-card quality but do not systematically compare popular and unpopular models across documentation, repository structure, platform metadata, and technical metadata. We studied 66,909 Hugging Face models collected on 2026-02-28 and extracted 31 features across four dimensions: Documentation, Models, Platforms, and Techniques. For downloads and likes, we compared the top 10% popular models with the bottom 80% using statistical tests and effect sizes, trained five classifiers to predict popularity, ranked feature importance with the best-performing model, and tested whether predictions held across model domains and contributor affiliations. Popular models consistently had longer model cards, more code examples, and more GitHub and arXiv links, while quantized models showed contrasting patterns under downloads and likes. A Random Forest reached AUC 0.861 for downloads and 0.848 for likes, with root-level file counts and model size also among the most important features. Cross-group transfer stayed mostly above random but dropped by 0.12-0.17 AUC on average, showing that prediction performance was partly group-specific. The findings remained stable after removing download outliers and under alternative popularity thresholds, and a manual review of popular model cards linked the key documentation features to concrete content. We derive recommendations for model practitioners and directions for future research on model-sharing platforms.

cs.SE

版本更新 · v2arXiv:2607.26057

Pass the Baton: Trajectory-Relayed On-Policy Distillation

Haolei Xu, Xiaowen Xu, Haiwen Hong, Zixuan Ni, Hongxing Li, Yiwen Qiu, Weiming Lu, Yongliang Shen

原文摘要

On-policy distillation (OPD) grounds token-level supervision in the student's own trajectory, yet suffers from prefix failure: once the student commits to a wrong reasoning direction, all subsequent generation builds on this deviation, producing misdirected continuations that elicit unreliable supervision and waste compute. We identify a teacher-student continuation asymmetry on failed prefixes, where the teacher tends to redirect while the student continues along the original direction, and convert it into a label-free handoff trigger in Relay On-Policy Distillation (Relay-OPD). During training, Relay-OPD constructs relay trajectories by letting the teacher briefly take over at detected trigger points to produce a teacher leg, after which the student resumes and is optimized on the resulting trajectory. A limited relay budget concentrates intervention on critical early positions while limiting departure from the student policy. With a Qwen3-4B-Instruct-2507 teacher and Qwen3-0.6B/1.7B-Non-Thinking students on eight mathematical reasoning benchmarks, Relay-OPD achieves the best or second-best results on every benchmark, outperforming standard OPD by +5.73% and the strongest baseline FastOPD by +1.49% on average for 1.7B, with consistent gains at 0.6B. Training trajectory length is reduced by over 50%.

cs.CL · cs.AI

新论文 · v1arXiv:2611.20050

Clinician Assessment of Laryngeal Disease with Auditory AI before Laryngoscopy: A Paired Reader Study

Siyuan Hou, Shuting Yang, Zengrui Jin, Liyu Cheng, Siyin Wang, Yunhan Shi, Xueyan Li, Rong Hu 等 16 位作者

原文摘要

Background: Before laryngoscopy, clinicians assess the likely cause of hoarseness and whether to suspect a precancerous or malignant lesion. Whether voice-based artificial intelligence (AI) supports pre-laryngoscopic categorization is unknown. Methods: We developed an auditory large language model that generates a patient-level assessment card (provisional category, voice-quality ratings, and rationale) and a prespecified malignancy alert from voice recordings, patient information, and acoustic measures obtained before laryngoscopy at one center. We evaluated both in a held-out test set and a later cohort. In a sequential reader study, 32 otolaryngology clinicians categorized held-out cases without AI and, 1 to 2 weeks later, reviewed their initial categories with the card (without the alert). Results: Across 800 paired assessments, reader accuracy was 49.8% before and 56.8% after review (difference, 7.0 percentage points; 95% CI, 1.7 to 13.5; P < 0.001); 99 categories changed from incorrect to correct and 43 from correct to incorrect (ratio, 2.30; 95% CI, 1.22 to 5.31). Accuracy rose 14.7 points when the model was correct and fell 8.7 when incorrect. Provisional-category accuracy was 72% in held-out testing and 67% in temporal validation. The malignancy alert flagged 85.7% and 86.8% of precancerous or malignant lesions. Sensitivity for precancerous or malignant lesions increased from 53.9% to 75.0% after AI-assisted review (difference, 21.1 percentage points; 95% CI, 4.3 to 42.0). Conclusions: In this single-center study, clinicians' pre-laryngoscopic categorization accuracy was higher after reviewing the AI assessment card, although the sequential design cannot fully separate AI assistance from reassessment; gains depended on model correctness without clearly selective adoption. Randomized and external evaluations, including of the malignancy alert, are needed.

eess.AS · cs.AI

新论文 · v1arXiv:2611.18293

PrivacyPR: A Dataset of Real-World Privacy Fixes for Benchmarking Privacy-Aware Code Review

Hua Yang, Aditya Bhardwaj, Shidong Pan, Bowen Xu

原文摘要

Coding agents increasingly review code, yet whether they catch the privacy defects that laws such as the GDPR require software to avoid remains largely untested, partly because few datasets connect real privacy defects to their accepted fixes and review context. We present PrivacyPR, a dataset of 4,131 merged GitHub pull requests (PRs) from 1,469 repositories, created between June 2018 and January 2024. Each PR includes its commits, diffs, review discussion, and LLM-generated labels for privacy relevance, fix type, and GDPR principles. PrivacyPR provides 1,169 high-confidence privacy fixes and 1,643 non-fix PRs, together with Verified, a manually checked split of 250 privacy fixes and 250 non-fix PRs, and Small, a 100 + 100 subset for evaluation. On Verified, both annotators endorsed 72% of the LLM's GDPR-principle assignments. Beyond supporting empirical study of how privacy defects arise and are fixed, PrivacyPR provides a benchmark for privacy-aware code review on real, developer-accepted fixes. Its evaluation harness reverts each fix to reintroduce the defect. On Small, two coding agents detect only 42-51% of defects under a generic review prompt, but 88-91% when explicitly asked to check privacy, while flagging 11-15% of non-fix PRs. These results show that privacy defects can be detected by coding agents, but that detection depends strongly on review intent and comes with false alarms. Data, code, reviews, judge verdicts, and an extensible evaluation harness are publicly available at https://doi.org/10.5281/zenodo.23744765.

cs.CR · cs.AI · cs.SE

新论文 · v1arXiv:2611.17123

保留漏洞、消除告警:用自然重写测试 CodeQL 污点追踪的鲁棒性

Preserve the Vulnerability, Break the Alert: Testing the Robustness of CodeQL Taint Tracking with Natural Rewrites

Zhezhen Cao, Weiyuan Ding, Bowen Xu

AI 速读 基于摘要

作者报告,用大语言模型对真实 npm 漏洞做保持漏洞语义的自然重写,可使 CodeQL 的 JavaScript 污点追踪告警消失,说明告警消失并不等于漏洞已修复。

cs.CR · cs.SE

查看原文摘要

Static application security testing (SAST) tools such as CodeQL serve as security gates in continuous integration and code review, and a gate is also read in reverse. When a previously reported alert disappears after a change, developers and the tools around them often take the vulnerability to be gone. That inference holds only if the analyzer is robust to changes that keep the vulnerability. We test this robustness for CodeQL's JavaScript taint tracking with natural, vulnerability-preserving rewrites. A large language model, given only the alert CodeQL exposes and no knowledge of its implementation, rewrites one function along a reported flow as a plausible maintenance change, and we verify by manual audit and controlled experiments that the flaw survives and the alert does not. On 172 flows that CodeQL detects in real npm vulnerabilities, three rewrites per flow silence 26.2% of them, at about USD 0.51 per silenced flow. The same generator, asked only to refactor the same functions with no knowledge of the alert, silences none of them. Every silencing rewrite traces to one of seven gaps at the library-modeling boundary, and 64 of the 70 rewrites cross a gap inside an API CodeQL already models. The largest is the WHATWG URL constructor, which propagates taint to a URL's query components but not to its pathname. The boundary is not specific to CodeQL. 74% of the silencing rewrites also silence an independently built taint analyzer, and 53% escape three analyzers combined. A disappearing alert after an ordinary-looking change is therefore weak evidence that a vulnerability has been fixed, the gaps point SAST developers to narrow and concrete repair targets, and robustness to natural rewrites belongs alongside detection rate in how taint analyzers are evaluated.

版本更新 · v3arXiv:2608.25735

指路而不露终点:面向大规模实用隐私密集检索

Pointing the Way, Hiding the Destination: Practical Private Dense Retrieval at Scale

Peichun Hua, Danyang Chen, Junan Zhang, Haifeng Sun, Jingyu Wang, Diwen Xue, Mingyu Li, Yunming Xiao

AI 速读 基于摘要

利用学习式深度哈希生成随机二进制码作为隐私过滤器,在保护查询和结果选择的同时,以短候选列表实现接近全语料检索的精度和可接受的延迟。

cs.CR · cs.AI · cs.IR · cs.LG

查看原文摘要

Hosted retrieval-augmented generation (RAG) and semantic search allow users to query valuable provider-held corpora, raising two competing demands: to hide each query and chosen result, yet reveal only the documents that the user is authorized to receive. Existing cryptographic approaches either make this costly by processing the entire corpus for every query, or sacrifice quality for efficiency by scanning a few clusters. We repurpose learned deep hashing as a private filter: a randomized binary code points the provider to a short candidate list, while encrypted reranking and oblivious key transfer protect the precise query and final selection. This shortlist short-circuits full-corpus cryptographic search without sacrificing retrieval quality: with 200-500 candidates, it closely matches full-corpus retrieval across five zero-shot corpora spanning 25K to 5.4M documents. On the full 2.68M-passage NQ corpus over a 10-Gbps link, our protocol only adds 0.38 seconds, or 6.5%, to a 128-token Qwen3-32B RAG pipeline. The released code satisfies directional metric differential privacy (DP) and substantially reduces embedding-inversion and property-inference leakage, demonstrating that a carefully learned shortlist can make private dense retrieval both accurate and practical.

新论文 · v1arXiv:2611.16442

DesignTaste: Benchmarking AI-Generator Visual Design Against Award-Wining Works

Zhou Chen, JiaXiang Li, DanYing Pan, Ye Li, TianRui Yang, EnBao Zhang, YaNan Xiao, LiXin Sun 等 17 位作者

原文摘要

Text-to-image benchmarks commonly assess text-image alignment or overall preference. Whether models can produce good designs when a brief leaves the appearance open remains less well understood. We introduce DesignTaste, a visual design benchmark that compares generated images with award-winning works. From these works and their publicly available context, we derive 403 design briefs: 100 poster tasks from 100 Beste Plakate 2025 and 303 tasks spanning 20 categories from ADC 2026. Six leading commercial text-to-image models each generate an image from the same brief. A single model judge scores each output and the corresponding award-winning work independently on five evaluation dimensions. On each dataset, only ArtArk achieves a higher mean overall score than the award-winning works. All six models score lower in Creative Originality on both datasets. These results show that producing finished, visually appealing images does not imply that a model can develop good design solutions.

cs.CV · cs.AI

版本更新 · v2arXiv:2609.01870

ArcticSwarm: Deferring Early Consensus in Long-Horizon Multi-Agent Research

Soyoung Yoon, Boyi Liu, Yite Wang, Ruofan Wu, Canwen Xu, Nikki Lijing Kuang, Seung-won Hwang, Yuxiong He 等 9 位作者

原文摘要

Multi-agent systems have shown strong performance in domains with reliable verifiers such as coding, where a verifier can select among candidates generated in parallel. Long-horizon web search has no such verifier. Majority voting over independent runs cannot add an answer that none of the runs reaches. Agents that instead read each other's findings adopt an early candidate before alternatives are tested, a failure we call early consensus. We present ArcticSwarm, a multi-agent research architecture that separates evidence gathering from evidence integration. A gathering gate lets selected search tasks post findings to a shared board without reading it. An integration gate checks a candidate at three steps, and a failed check reopens search instead of only rejecting the candidate. On all 830 BrowseComp-Plus questions with the open-weight Qwen 3.5-27B, ArcticSwarm reaches 82.6%. MiroFlow reaches 70.6% under the same model, tools, and limits, and majority voting reaches 62.5% at the same token budget. Without the gathering gate, teams keep a wrong first candidate in 32.2% of runs instead of 22.3%, and accuracy drops by 3.8 points. Removing the integration gate lowers accuracy by 5.9 points. With GPT-5 on live-web BrowseComp, ArcticSwarm reaches 73.6%, above the published scores of GPT-5 (54.9%) and MiroFlow (63.4%). These results show that multi-agent search without a verifier benefits from deferring consensus until candidates are checked, instead of voting after search ends.

cs.MA

新论文 · v1arXiv:2611.15019

obfsgen: Generative Tunnel Obfuscation

Kaidong Wu, Wentao Du, Shinan Liu, Diwen Xue

原文摘要

Encrypted tunnels are widely used to circumvent censorship, which has drawn censors' efforts to detect and block them. As the circumvention tunnels in widest use have largely converged on TLS, detection has shifted from the content of their traffic to its shape, where recent research and real-world blocking events have revealed exploitable fingerprints of tunneling. Existing traffic-shaping obfuscators such as obfs4, however, are fundamentally limited in hiding these fingerprints, as their application-driven design can only add to the shape that the application creates. A schedule-driven obfuscator, which decides when and how much to send independently of the application, can give tunnels any shape but raises new challenges, first and foremost what a good shape is. This paper presents obfsgen, a schedule-driven obfuscator for tunneling traffic. obfsgen samples its schedule from a generative sequence model of normal traffic, which predicts the future shape of a flow from its whole history, to capture the long-range dependencies that make a flow's shape normal. Built around this model, obfsgen addresses system problems that such an obfuscator faces in practice, such as synchronization across a noisy network. We evaluate the (in)distinguishability that obfsgen offers, the performance overhead that it adds and the computational resources that it requires. We see the schedule-driven architecture as the next generation of tunnel obfuscators, and we hope that obfsgen and the rationale behind its design help make such obfuscators practical and deployable.

cs.CR · cs.NI

版本更新 · v4arXiv:2611.01666

Stand-In Effects and Pooling Amplification: Lessons from Biological World Models for Testing Action Effects

Bowen Xu, Boyu Chen, Jiaying Li

原文摘要

Whether an action-conditioned world model responds to actions is usually checked by replacing or removing the actions and measuring the loss in accuracy. Taking biological experiments as the guiding case, we show that such diagnostics report action effects where there are none: under autocorrelated designs, future actions stand in for history the model does not see. When the design law is known, a conditional randomization test written as an e-value is valid in finite samples whatever the model. A second failure arises whenever many sites share one action sequence, as in world models: the pooled score turns small model errors along the action contrast into an e-value near zero, so a valid test loses its power although the effect is real. Tempering the score changes the bet, not the validity. The log e-value is concave in the tempering factor and zero at zero, so too small a factor costs little and too large a one can cost everything; every single factor we tried, estimated or tuned, failed in some setting, and mixtures over a fixed grid never did. Our own preregistered analysis of optogenetic ERK experiments failed by pooling; a tempered statistic, registered before the data were obtained, then rejected the null on a never-opened field (E = 1.1e10). In three registered tests on models trained by others, the pooled test rejected in 0 of 96 units of a diffusion model, and reached the power bar in 10 of 26 Atari games under IRIS's token likelihood and 1 of 6 control tasks under TD-MPC2; mixtures fixed in advance reached it in 96 units, 24 games and 6 tasks. On recorded robot video with iVideoGPT the mixture rejected in 155 of 159 trajectories; the pooled test did not collapse there (143). Under exact nulls the statistics made 14 rejections in 26,664 tests.

cs.LG · q-bio.QM · stat.ME

版本更新 · v4arXiv:2610.05533

What Does a Harness Repair? Four Preregistered Studies Around a Longevity-Tuned Language Model

Bowen Xu, Boyu Chen

原文摘要

A harness search around a frozen longevity-tuned 1.7B model reported a rescue. An audit traced its gain to rows the start left unparsed and found that the caller sent reference answers as assistant turns; replication through the same pipeline missed this defect. We preregistered four studies and report every endpoint. In the first two, thinking off raised test accuracy over capped thinking in 7 of 12 cells, mainly by parsing rows the cap left unparsed, and was not meaningfully worse than a rescue in 7 of 9 small-model test cells. Replication reproduced an injected defect's distortion in 9 of 20 evaluable cases. At 14B, two GEPA harnesses beat thinking off by 34 test points but were not detectably above the strongest constant label. In the third study, GEPA beat the constant on FPB (confirmatory), thinking off was not meaningfully worse than any BoolQ harness, and 3 of 4 MMLU-Pro searches kept their seed; the fourth comparison was inconclusive on test. On exploratory Autocast, harnesses were within 0.7 points of the constant and gave one label in 93.1% to 96.8% of parsed answers on test. On the prediction tasks (14B mortality, Autocast) the harnesses ended near the constant label. The fourth study varied search-set skew and objective on CREAK and resolved Manifold questions. All its results are descriptive because render hashes were not committed before the first replication call (D001). Minority-recall estimates were lower at 80:20 than at 50:50 in all four test cells. The balanced objective and re-selection from each search's pool had higher minority- and lower majority-recall estimates in all four cells. Five of their eight minority-recall intervals included zero, as did all eight balanced-accuracy contrast intervals. This does not show that skew explains the earlier collapse.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.10058

From Delegation to Action Approval: An Exploratory Study of Developers' Practices in Collaborating with AI Coding Agents

Yinan Wu, Dominik Wermke, Ze Shi Li, Kathryn Thomasset Stolee, Bowen Xu

原文摘要

Developers now collaborate with AI coding agents that write code and take actions on their behalf. Beyond reviewing this code, developers decide what to delegate and which agent actions to approve. Some of these actions, such as running shell commands or reading files, can have effects that never appear in the resulting code. How developers make these decisions and how much effort this requires are not well understood. To better understand this collaboration, we surveyed 203 developers who use AI coding agents. We found that careful code review does not necessarily extend to action approval: 61% of developers who carefully reviewed most or all generated code still often approved at least one type of agent action without close review. Developers aware of security or privacy incidents or concerning practices did so more often than those who were not. We also found that autonomy varied more between developers than between development activities. Although developers reported a lighter workload, supervising agents remained mentally demanding. When asked what would make them grant more autonomy, developers wanted more than reliability: more than half of their responses called for agents that are bounded, observable, or reversible. Our findings suggest that as agents take on more of the coding, the challenge shifts toward supporting the developers who supervise them.

cs.SE

新论文 · v1arXiv:2611.04066

Graph-Localized Coverage for Conformal Prediction on Inductive Graphs

Jinwen Xu, Qin Lu, Georgios B. Giannakis

原文摘要

Conformal prediction (CP) provides marginal coverage guarantees, but marginal coverage does not determine where prediction errors occur on a graph. Queries from different graph neighborhoods can exhibit substantially different levels of uncertainty even when aggregate coverage remains near its target. We introduce GLocalCP, an online CP framework for inductive node classification and edge prediction that combines marginal and graph-local calibration with uncertainty-adaptive prediction sets. Specifically, GLocalCP puts forth a novel uncertainty-adaptive nonconformality score that relies on the mutual information obtained from Monte-Carlo (MC) dropout based prediction. Then, it augments a global online threshold with a kernel-based local correction that propagates coverage feedback among similar graph queries. On the theoretical horizon, we establish pathwise $O(N^{-1/2})$ marginal calibration error without exchangeability or stationarity and, for the exact-kernel formulation, an explicit $O(N^{-1/2})$ graph-localized calibration guarantee. Experiments on seven node-classification and seven edge-prediction benchmarks show that similar marginal coverage can conceal substantially different local reliability. Our proposed GLocalCP maintains coverage near the target while consistently reducing graph-local calibration error. Uncertainty scaling in GLocalCP further reduces prediction-set size across all evaluated node and edge datasets while leaving marginal coverage essentially unchanged.

cs.LG · stat.ML

新论文 · v1arXiv:2611.04062

StreamLTT: Risk Certification for Policy Selection under Stream Dependence

Jinwen Xu, Qin Lu

原文摘要

Learn-then-Test (LTT) provides a general framework for certifying deployment policies against a prescribed risk target, but its standard formulation targets a fixed population risk estimated from calibration losses. In sequential deployment, losses may instead be dependent and nonstationary, so the risk of a fixed policy can evolve with the observed history. We introduce StreamLTT, which replaces this fixed population risk with average predictable risk, defined by conditioning each decision's loss on the information available before its outcome is revealed. Conditional centering turns empirical deviations from this target into a martingale sum, enabling both a test that adapts to predictable variance and a finite-sample Hoeffding alternative. Combined with family-wise error rate control, StreamLTT yields a simultaneously certified policy region that supports subsequent deployment selection. Across seven forecasting datasets and seven deployment risks, StreamLTT places the certified frontier substantially closer to the prescribed risk target than IID, robust-variance, block-bootstrap, and concentration-based alternatives, while empirical family-wise error closely follows the prescribed level. A finite-sample ablation further shows that the predictable-risk formulation drives the tighter risk frontier, while adaptation to predictable variance makes more effective use of the available family-wise error budget. StreamLTT therefore extends risk-certified policy selection to dependent deployment streams by addressing both what risk should be certified and how it should be tested.

cs.LG · math.ST · stat.ML

新论文 · v1arXiv:2610.39419

Diagnosing the Reference Gap: What Visual Scaffolds Can and Cannot Fix in Multimodal Reasoning

Guizhen Chen, Weiwen Xu, Wotao Yin, Anh Tuan Luu, Hou Pong Chan

原文摘要

Multimodal large language models can recognize visual content yet struggle to reason reliably over it across multiple steps. In counting, a model may see all objects but lose track of which ones it has counted. We argue that such failures, commonly attributed to perception, reveal a reference gap: the failure to bind language to the relevant visual entities, regions, or locations. Motivated by the theory of visual routines, we test whether external references such as numbering can narrow this gap. We introduce ReGap-Bench, a paired benchmark of five visual reasoning tasks in which the scene, question, and answer are fixed and only the reference varies. Results show that task-matched references improve accuracy by up to 25\%, while mismatched references provide little benefit or even hurt performance. These findings identify the reference gap as a bottleneck distinct from visual perception and show that constructing task-appropriate visual references can serve as a training-free visual skill for improving multimodal reasoning.

cs.CV · cs.CL

版本更新 · v2arXiv:2510.13992

Optimizing Spectral Signature in Code Backdoor Detection

Quoc Hung Le, Thanh Le-Cong, Bach Le, Bowen Xu

原文摘要

As Large Language Models (LLMs) are increasingly integrated into software development workflows, they also become prime targets for adversarial attacks. Among these, backdoor attacks are a significant threat, allowing attackers to manipulate model outputs through hidden triggers embedded in training data. Detecting such backdoors remains a challenge, and one promising approach is the use of Spectral Signature defense methods that identify poisoned data by analyzing feature representations through eigenvectors. While some prior works have explored Spectral Signatures for backdoor detection in neural networks, recent studies suggest that these methods may not be optimally effective for code models. In this paper, we revisit the applicability of Spectral Signature-based defenses in the context of backdoor attacks on code models. We systematically evaluate their effectiveness under various attack scenarios and defense configurations, analyzing their strengths and limitations. We found that the widely used setting of Spectral Signature in code backdoor detection is often suboptimal. Hence, we explored the impact of different settings of the key factors. We discovered a new proxy metric that can more accurately estimate the actual performance of Spectral Signature without model retraining after the defense. Our findings contribute to a deeper understanding of the configurations of Spectral Signature defenses in securing code models and offer insights into potential improvements for future research.

cs.SE · cs.LG

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1362ACL Anthology · ACL 2026

Modal Dependency Parsing as Structured Prediction over Source-Cue Scope

Jayeol Chun Brandeis University, Nianwen Xue Brandeis University

原文摘要

Modal dependency parsing-the task of identifying a semantic graph that represents who is responsible for an event-centered claim and with what degree of certainty-relies on recognizing source-introducing cues and correctly linking them to their associated content. However, prior work has largely focused on identifying sources only, treating cue expressions and their modal coverage as auxiliary signals. In this work, we propose a structured prediction framework that leverages large language models (LLMs) to explicitly identify source-cue pairs as well as their respective scope, which together define the modal contexts governing downstream source attribution for events. By concentrating learning at the source-cue level and constraining event-level decisions to a small, scope-defined candidate set, our top-down approach enables more efficient inference in long, event-rich documents. Experiments show this approach surpasses prior state-of-the-art results by 3 and 4% for English and Chinese datasets, respectively.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1443ACL Anthology · ACL 2026

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

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

原文摘要

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

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1463ACL Anthology · ACL 2026

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

Ziwen Xu Zhejiang University, Chenyan WU Zhejiang University of Technology, Hengyu Sun, Haiwen Hong Alibaba Group, Mengru Wang, Yunzhi Yao Zhejiang University, Longtao Huang Alibaba Group, Hui Xue 等 12 位作者

原文摘要

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

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1614ACL Anthology · ACL 2026

TPS-Bench: Evaluating AI Agents’ Tool Planning & Scheduling Abilities in Compounding Tasks

Hanwen Xu Shanghai Jiaotong University, Xuyao Huang, Yuzhe Liu, Zhijie Deng Shanghai Jiaotong University

原文摘要

Large language model (LLM) agents have exhibited strong problem-solving competence across domains like research and coding. Yet, it remains underexplored whether LLM agents can tackle compounding real-world problems that require a diverse set of tools to complete. Given a broad, heterogeneous tool repository, LLM agents must not only select appropriate tools based on task planning analysis but also strategically schedule the execution order to ensure efficiency. This paper introduces TPS-Bench to benchmark the ability of LLM agents in solving such problems that demand Tool Planning and Scheduling. TPS-Bench collects 200 compounding tasks of two difficulty levels, based on a tool repository containing hundreds of model context protocol (MCP) tools. In particular, each task is composed of multiple subtasks, such as web search, map navigation, calendar checking, etc., and each subtask can be completed by a basic tool. Our evaluation emphasizes both task completion rate and efficiency. The empirical studies on popular closed-source and open-source LLMs indicate that most models can perform reasonable tool planning, but differ in scheduling. For example, GLM-4.5 achieves an outperforming task completion rate of 64.72% with extensive sequential tool calls, hence suffering from significantly long execution time. By contrast, GPT-4o prioritizes parallel tool calls but achieves only a 45.08% completion rate. Considering reinforcement learning (RL) can be a viable way to improve the scheduling efficiency without compromising performance, we perform an initial study on Qwen3-1.7B and witness a 14% reduction in execution time alongside a 6% gain in task completion rate based on only 597 RL training samples.

cs.CL · cs.AI

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1754ACL Anthology · ACL 2026

Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization

Jia Liu South China University of Technology, Pazhou Laboratory and Huazhong University of Science and Technology, Jiaxin Luo, Weiwen Xu Alibaba Group, Jonathan M. Garibaldi The University of Nottingham Ningbo China and University of Nottingham, Xiao-Kun Wu Renmin University of China, Yixue Hao, Min Chen South China University of Technology and Pazhou Laboratory

原文摘要

Chain-of-Thought (CoT) prompting enables large language models to produce multi-step reasoning, yet how such reasoning-related structure is expressed during training remains poorly understood. We present Gradient-based Structural Developer (GSD), an unsupervised framework with a principled gradient aggregation view that tracks span-level gradient during fine-tuning on reasoning benchmarks to understand how models develop structured, step-by-step reasoning capabilities. Our analysis shows that while gradients at the level of individual tokens are often noisy, aggregating gradients over contiguous reasoning-related spans reveals stable and recurring directional alignment across samples. We refer to these directionally aligned patterns as aligned sequential stresses, reflecting consistent gradient organization associated with similar reasoning procedures. Beyond capturing semantically similar reasoning instances, such gradient alignment also reveals structurally similar but semantically diverse cases that share common procedural organization. These findings position GSD as a diagnostic framework for analyzing how procedural reasoning structures emerge during training, with downstream selection results serving as auxiliary evidence correlating gradient alignment with adaptation efficiency.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.2173ACL Anthology · ACL 2026

Reframing Responsibility: Framing-Aware Event Causality Identification

Jin Zhao, Jiayi Yao University of Washington, Xinrui Hu, Nianwen Xue Brandeis University

原文摘要

Causal explanations in political narratives are often framed and contested. Different sources may explain the same event by assigning responsibility to different actors and expressing varying levels of certainty. Standard Event Causality Identification (ECI) focuses on detecting causal links and does not capture these distinctions. We introduce Framing-Aware Event Causality Identification (FrECI), a framing-aware extension of ECI that models causal explanations as structured claims including responsibility targets, evaluative framing, source type, and epistemic modality grounded in established framing theories. We construct a multilingual dataset aligned across English, Chinese, and Arabic narratives using shared event anchors. We evaluate FrECI using prompt-based large language model baselines and supervised neural models. Results show that prompt-based baselines struggle to recover complete framed causal claims, while joint supervised models perform substantially better. Finally, we demonstrate that FrECI enables quantitative analysis of divergent causal attribution across narratives.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.670ACL Anthology · ACL 2026

Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human–Agent Interaction

Zisu Huang, Muzhao Tian, Xiaohua Wang, Jingwen Xu, Zhengkang Guo, Qi Qian, Kaitao Song Apple, Jiakang Yuan Shanghai Artificial Intelligence Laboratory and Fudan University 等 10 位作者

原文摘要

As LLM-based agents are increasingly used in long-term interactions, cumulative memory is critical for enabling personalization and maintaining stylistic consistency. However, most existing systems adopt an “all-or-nothing” approach to memory usage: incorporating all relevant past information can lead to Memory Anchoring, where the agent is trapped by past interactions, while excluding memory entirely results in under-utilization and the loss of important interaction history. We show that an agent’s reliance on memory can be modeled as an explicit and user-controllable dimension. We first introduce a behavioral metric of memory dependence to quantify the influence of past interactions on current outputs. We then propose Steerable Memory Agent, SteeM, a framework that allows users to dynamically regulate memory reliance, ranging from a fresh-start mode that promotes innovation to a high-fidelity mode that closely follows interaction history. Experiments across different scenarios demonstrate that our approach consistently outperforms conventional prompting and rigid memory masking strategies, yielding a more nuanced and effective control for personalized human-agent collaboration.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.989ACL Anthology · ACL 2026

MMAC: A Multilingual, Multimodal Alignment Framework for Cultural Grounding Evaluation

Weihua Zheng, Zhengyuan Liu Agency for Science, Technology and Research (A*STAR), Singapore, Tanmoy Chakraborty Indian Institute of Technology, Delhi, Weiwen Xu Alibaba Group, Xiaoxue Gao, Bryan Chen Zhengyu Tan, Bowei Zou A*STAR, Chang Liu Shanghai University of Finance and Economics 等 35 位作者

原文摘要

The global deployment of Large Language Models (LLMs) underscores the urgent need to evaluate their cultural alignment. However, assessing genuine “cultural awareness” across modalities (text, vision, speech) and languages remains a significant challenge. To comprehensively investigate this domain, we propose MMAC, a systematic framework that encompasses a tri-modally aligned cultural benchmark creation pipeline and a five-dimensional evaluation protocol to assess cross-country awareness disparities, evaluate cross-lingual and cross-modal consistency, and verify cultural knowledge generalization and grounding validity. Given the prevailing Western cultural bias in current models, we focus on 8 Asian countries as our dataset foundation to more acutely reveal potential cultural deficiencies in LLMs. Our dataset, MMAC-bench, features 27,000 human-curated questions across 10 languages. Crucially, it is the first dataset aligned at the input level across text, image, and speech, enabling direct cross-modal transfer tests. Each question consists of multiple-choice options accompanied by open-ended generated explanations, where 79% require multi-step reasoning grounded in cultural context, moving beyond simple memorization. We probe the causes of modal divergence, offering insights into fostering culturally robust MLLMs.

cs.CL · cs.CV

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表