大模型对齐与安全

2510.16167 2025-10-21 cs.LG cs.CL 86%

Alignment is Localized: A Causal Probe into Preference Layers

Archie Chaudhury

机构 * Independent（独立研究者）

专题命中偏好对齐：alignment(title,abstract)；RLHF(abstract)；safety(abstract)；分类 cs.CL、cs.LG

2502.01203 2025-10-21 cs.LG stat.ML 83%

KL-Regularized RLHF with Multiple Reference Models: Exact Solutions and Sample Complexity

Gholamali Aminian, Amir R. Asadi, Idan Shenfeld, Youssef Mroueh

机构 * The Alan Turing Institute（艾伦·图灵研究所）； Statistical Laboratory（统计实验室）； University of Cambridge（剑桥大学）； Massachusetts Institute of Technology（麻省理工学院）； IBM Research USA（IBM美国研究）

专题命中偏好对齐：RLHF(title,abstract)；alignment(abstract)；分类 cs.LG

Comments Extra experiments are added in new version

2411.17041 2025-10-21 cs.CV cs.AI cs.LG 81%

Free$^2$Guide: Training-Free Text-to-Video Alignment using Image LVLM

Jaemin Kim, Bryan Sangwoo Kim, Jong Chul Ye

机构 * Graduate School of AI, KAIST（人工智能研究生院，韩国科学技术院）

专题命中偏好对齐：alignment(title,abstract)；分类 cs.AI、cs.LG

Comments ICCV 2025 accepted

2504.05831 2025-10-21 cs.CL 79%

Leveraging Robust Optimization for LLM Alignment under Distribution Shifts

Mingye Zhu, Yi Liu, Zheren Fu, Yongdong Zhang, Zhendong Mao

机构 * University of Science and Technology of China（中国科学技术大学）； State Key Laboratory of Communication Content Cognition（通信内容认知国家重点实验室）

专题命中偏好对齐：alignment(title,abstract)；分类 cs.CL

Comments NeurIPS 2025

2405.15065 2025-10-21 cs.LG 77%

Direct Preference Optimization With Unobserved Preference Heterogeneity: The Necessity of Ternary Preferences

Keertana Chidambaram, Karthik Vinay Seetharaman, Vasilis Syrgkanis

机构 * Stanford University（斯坦福大学）

专题命中偏好对齐：alignment(abstract)；RLHF(abstract)；DPO(abstract)；分类 cs.LG

2510.14400 2025-10-21 cs.CL cs.AI cs.IR 76%

MedTrust-RAG: Evidence Verification and Trust Alignment for Biomedical Question Answering

Yingpeng Ning, Yuanyuan Sun, Ling Luo, Yanhua Wang, Yuchen Pan, Hongfei Lin

机构 * College of Computer Science and Technology, Dalian University of Technology（大连理工大学计算机科学与技术学院）； Air Force Communications NCO Academy（空军通信NCO学院）

专题命中偏好对齐：alignment(title)；分类 cs.CL、cs.AI

Comments Accepted as a short paper at BlBM2025

2502.20548 2025-10-21 cs.LG cs.AI cs.CL 75%

$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training

Jin Peng Zhou, Kaiwen Wang, Jonathan Chang, Zhaolin Gao, Nathan Kallus, Kilian Q. Weinberger, Kianté Brantley, Wen Sun

专题命中偏好对齐：alignment(abstract)；DPO(abstract)；分类 cs.CL、cs.AI、cs.LG

Comments NeurIPS 2025

2508.12081 2025-10-21 cs.CV cs.AI cs.CL 73%

VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models

Haidong Xu, Guangwei Xu, Zhedong Zheng, Xiatian Zhu, Wei Ji, Xiangtai Li, Ruijie Guo, Meishan Zhang, Min zhang, Hao Fei

机构 * Harbin Institute of Technology (Shenzhen)（哈尔滨工业大学（深圳））； University of Macau（澳门大学）； University of Surrey（Surrey大学）； Nanjing University（南京大学）； Nanyang Technological University（南洋理工大学）； National University of Singapore（新加坡国立大学）

专题命中偏好对齐：alignment(abstract)；DPO(abstract)；分类 cs.CL、cs.AI

Comments Accepted by NeurIPS 2025; Project Page: https://walkermitty.github.io/VimoRAG

2510.12943 2025-10-21 cs.CL 57%

The Curious Case of Curiosity across Human Cultures and LLMs

Angana Borah, Zhijing Jin, Rada Mihalcea

机构 * University of Michigan - Ann Arbor（密歇根大学安阿伯分校）； University of Toronto（多伦多大学）； Vector Institute（向量研究所）； MPI for Intelligent Systems, Tubingen, Germany（图宾根德国智能系统研究所）

专题命中偏好对齐：alignment(abstract)；分类 cs.CL

Comments Preprint (Paper under review)

2510.15993 2025-10-21 q-fin.PM cs.LG q-fin.ST 57%

Aligning Language Models with Investor and Market Behavior for Financial Recommendations

Fernando Spadea, Oshani Seneviratne

机构 * Rensselaer Polytechnic Institute（拉特兰理工学院）

专题命中偏好对齐：alignment(abstract)；分类 cs.LG

2502.06023 2025-10-21 cs.CV 50%

Dual Caption Preference Optimization for Diffusion Models

Amir Saeidi, Yiran Luo, Agneet Chatterjee, Shamanthak Hegde, Bimsara Pathiraja, Yezhou Yang, Chitta Baral

机构 * School of Computing and Augmented Intelligence（计算与增强智能学院）； Arizona State University（亚利桑那州立大学）

专题命中偏好对齐：DPO(abstract)

2510.17402 2025-10-21 cs.CL cs.AI cs.LG 75%

Leveraging Group Relative Policy Optimization to Advance Large Language Models in Traditional Chinese Medicine

Jiacheng Xie, Shuai Zeng, Yang Yu, Xiaoting Tang, Guanghui An, Dong Xu

专题命中安全训练：alignment(abstract)；trustworthy(abstract)；分类 cs.CL、cs.AI、cs.LG

2510.16144 2025-10-21 cs.NI cs.AI cs.MA 70%

Agentic AI for Ultra-Modern Networks: Multi-Agent Framework for RAN Autonomy and Assurance

Sukhdeep Singh, Avinash Bhat, Shweta M, Subhash K Singh, Moonki Hong, Madhan Raj K, Kandeepan Sithamparanathan, Sunder A. Khowaja, Kapal Dev

专题命中安全训练：safety(abstract)；trustworthy(abstract)；分类 cs.AI

2510.17369 2025-10-21 cs.RO cs.AI cs.LG 62%

Bridging Embodiment Gaps: Deploying Vision-Language-Action Models on Soft Robots

Haochen Su, Cristian Meo, Francesco Stella, Andrea Peirone, Kai Junge, Josie Hughes

机构 * EPFL（苏黎世联邦理工学院）； LatentWorlds AI ； TUDelft（代尔夫特理工大学）； Embodied AI SA（具身人工智能股份有限公司）

专题命中安全训练：safety(abstract)；分类 cs.AI、cs.LG

Comments Accepted by NeurIPS 2025 SpaVLE workshop. 4 pages, 2 figures(in main paper, excluding references and supplements)

2510.15896 2025-10-21 cs.HC cs.AI cs.CY 62%

From Coordination to Personalization: A Trust-Aware Simulation Framework for Emergency Department Decision Support

Zoi Lygizou, Dimitris Kalles

专题命中安全训练：safety(abstract)；分类 cs.AI、cs.CY

2510.16255 2025-10-21 cs.CR cs.AI 57%

Detecting Adversarial Fine-tuning with Auditing Agents

Sarah Egler, John Schulman, Nicholas Carlini

机构 * MATS & Anthropic Fellows Program（MATS与Anthropic Fellow项目）； Thinking Machines Lab（Thinking Machines实验室）； Anthropic

专题命中安全训练：safety(abstract)；分类 cs.AI

2510.15937 2025-10-21 q-fin.RM q-fin.TR 50%

Tail-Safe Stochastic-Control SPX-VIX Hedging: A White-Box Bridge Between AI Sensitivities and Arbitrage-Free Market Dynamics

Jian'an Zhang

专题命中安全训练：safety(abstract)

Comments 52 pages; 3 figures; PRIMEarxiv template; fully reproducible artifact (code, configs, plots)

2504.00682 2025-10-21 cs.RO 50%

Immersive Explainability: Visualizing Robot Navigation Decisions through XAI Semantic Scene Projections in Virtual Reality

Jorge de Heuvel, Sebastian Müller, Marlene Wessels, Aftab Akhtar, Christian Bauckhage, Maren Bennewitz

机构 * University of Bonn（波恩大学）； Lamarr Institute for Machine Learning and Artificial Intelligence（拉马尔人工智能与机器学习研究所）； Center for Robotics（机器人中心）； University of Mainz（美因茨大学）； Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS（弗劳恩霍夫智能分析与信息系统研究所）

专题命中安全训练：trustworthy(abstract)

2312.12861 2025-10-21 cs.RO 50%

Safe Multi-Agent Reinforcement Learning for Behavior-Based Cooperative Navigation

Murad Dawood, Sicong Pan, Nils Dengler, Siqi Zhou, Angela P. Schoellig, Maren Bennewitz

机构 * Humanoid Robots Lab, University of Bonn（波恩大学人形机器人实验室）

专题命中安全训练：safety(abstract)

2510.15430 2025-10-21 cs.CV cs.AI 85%

Learning to Detect Unknown Jailbreak Attacks in Large Vision-Language Models

Shuang Liang, Zhihao Xu, Jialing Tao, Hui Xue, Xiting Wang

专题命中越狱攻击：jailbreak(title,abstract)；alignment(abstract)；safety(abstract)；分类 cs.AI

Comments Withdrawn due to an accidental duplicate submission. This paper (arXiv:2510.15430) was unintentionally submitted as a new entry instead of a new version of our previous work (arXiv:2508.09201)

2510.17006 2025-10-21 cs.CL 83%

Online Learning Defense against Iterative Jailbreak Attacks via Prompt Optimization

Masahiro Kaneko, Zeerak Talat, Timothy Baldwin

机构 * MBZUAI（马克斯·普朗克人工智能研究所）； University of Edinburgh（爱丁堡大学）

专题命中越狱攻击：jailbreak(title,abstract)；safety(abstract)；分类 cs.CL

2510.17000 2025-10-21 cs.CR cs.CL cs.LG 73%

Bits Leaked per Query: Information-Theoretic Bounds on Adversarial Attacks against LLMs

Masahiro Kaneko, Timothy Baldwin

机构 * MBZUAI Abu Dhabi, UAE（阿布扎赫德MBZUAI）

专题命中越狱攻击：safety(abstract)；jailbreak(abstract)；分类 cs.CL、cs.LG

Comments NeurIPS 2025 (spotlight)

2509.19360 2025-10-21 cs.CL cs.AI 62%

Semantic Representation Attack against Aligned Large Language Models

Jiawei Lian, Jianhong Pan, Lefan Wang, Yi Wang, Shaohui Mei, Lap-Pui Chau

机构 * Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University（香港理工大学电子与电气工程系）； School of Electronics and Information, Northwestern Polytechnical University（西北工业大学电子与信息学院）

专题命中越狱攻击：alignment(abstract)；分类 cs.CL、cs.AI

2508.19304 2025-10-21 cs.CY cs.AI cs.CE 62%

Epistemic Trade-Off: An Analysis of the Operational Breakdown and Ontological Limits of "Certainty-Scope" in AI

Generoso Immediato

专题命中越狱攻击：safety(abstract)；分类 cs.AI、cs.CY

Comments Preprint V3 (October 2025)

2510.16794 2025-10-21 cs.CR cs.LG 57%

Black-box Optimization of LLM Outputs by Asking for Directions

Jie Zhang, Meng Ding, Yang Liu, Jue Hong, Florian Tramèr

机构 * ETH Zurich（苏黎世联邦理工学院）； University at Buffalo（布法罗大学）； Bytedance, Security Research（字节跳动安全研究）

专题命中越狱攻击：prompt injection(abstract)；分类 cs.LG

2510.16128 2025-10-21 cs.CR cs.CY 79%

Prompt injections as a tool for preserving identity in GAI image descriptions

Kate Glazko, Jennifer Mankoff

专题命中提示注入：prompt injection(title,abstract)；分类 cs.CY

Comments Accepted as a poster to Soups 2025

Journal ref The Twenty-First Symposium on Usable Privacy and Security (SOUPS 2025) Poster

2505.06493 2025-10-21 cs.CR cs.AI 70%

System Prompt Poisoning: Persistent Attacks on Large Language Models Beyond User Injection

Zongze Li, Jiawei Guo, Haipeng Cai

机构 * University at Buffalo（布法罗大学）

专题命中提示注入：jailbreak(abstract)；prompt injection(abstract)；分类 cs.AI

2508.02736 2025-10-21 cs.OS cs.SE 50%

AgentSight: System-Level Observability for AI Agents Using eBPF

Yusheng Zheng, Yanpeng Hu, Tong Yu, Andi Quinn

专题命中提示注入：prompt injection(abstract)

Journal ref PACMI'2025

2510.16257 2025-10-21 cs.CL 79%

Towards Low-Resource Alignment to Diverse Perspectives with Sparse Feedback

Chu Fei Luo, Samuel Dahan, Xiaodan Zhu

机构 * Department of Electrical and Computer Engineering & Ingenuity Labs Research Institute（电气与计算机工程系及创新实验室研究机构）； Conflict Analytics Lab, Queen’s University（冲突分析实验室，女王大学）； Cornell Law School（康奈尔法学院）

专题命中幻觉与事实性：alignment(title,abstract)；分类 cs.CL

Comments Findings of EMNLP 2025, 5 pages

2502.17598 2025-10-21 cs.LG cs.AI cs.CL 67%

Hallucination Detection in LLMs Using Spectral Features of Attention Maps

Jakub Binkowski, Denis Janiak, Albert Sawczyn, Bogdan Gabrys, Tomasz Kajdanowicz

机构 * Wroclaw University of Science and Technology（沃拉日-克拉夫大学科学与技术学院）； University of Technology Sydney（悉尼技术大学）

专题命中幻觉与事实性：safety(abstract)；分类 cs.CL、cs.AI、cs.LG

Comments Accepted to EMNLP 2025. Code available at https://github.com/graphml-lab-pwr/lapeigvals

1. 偏好对齐 11 篇