大模型对齐与安全 - arXivDaily 专题

2507.16663 2025-09-26 cs.CL cs.AI 62%

Turning Internal Gap into Self-Improvement: Promoting the Generation-Understanding Unification in MLLMs

Yujin Han, Hao Chen, Andi Han, Zhiheng Wang, Xinyu Liu, Yingya Zhang, Shiwei Zhang, Difan Zou

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.AI

Comments 31 pages, 16 figures, 12 tables

2509.19336 2025-09-25 cs.CL cs.AI 62%

Cognitive-Level Adaptive Generation via Capability-Aware Retrieval and Style Adaptation

Qingsong Wang, Tao Wu, Wang Lin, Yueying Feng, Gongsheng Yuan, Chang Yao, Jingyuan Chen

机构 * Zhejiang University（浙江大学）

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

Comments Accepted to Findings of EMNLP 2026

2509.19265 2025-09-24 cs.AI cs.CL 62%

Cross-Cultural Transfer of Commonsense Reasoning in LLMs: Evidence from the Arab World

Saeed Almheiri, Rania Hossam, Mena Attia, Chenxi Wang, Preslav Nakov, Timothy Baldwin, Fajri Koto

机构 * Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）

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

Comments EMNLP 2025 - Findings

2509.18316 2025-09-24 cs.CL cs.AI 62%

Brittleness and Promise: Knowledge Graph Based Reward Modeling for Diagnostic Reasoning

Saksham Khatwani, He Cheng, Majid Afshar, Dmitriy Dligach, Yanjun Gao

机构 * University of Colorado Boulder（科罗拉多大学博尔德分校）； University of Colorado Anschutz（科罗拉多大学安舒茨分校）； University of Wisconsin - Madison（威斯康星大学麦迪逊分校）； Loyola University Chicago（芝加哥洛约拉大学）

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

2506.04039 2025-09-23 cs.CV cs.AI cs.CL 62%

Mitigating Hallucinations in Large Vision-Language Models via Entity-Centric Multimodal Preference Optimization

Jiulong Wu, Zhengliang Shi, Shuaiqiang Wang, Jizhou Huang, Dawei Yin, Lingyong Yan, Min Cao, Min Zhang

机构 * Soochow University（苏州大学）； Baidu Inc.（百度公司）； Shandong University（山东大学）

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

Comments This paper is accepted by EMNLP2025

2509.16475 2025-09-23 cs.LG cs.CL 62%

Towards Universal Debiasing for Language Models-based Tabular Data Generation

Tianchun Li, Tianci Liu, Xingchen Wang, Rongzhe Wei, Pan Li, Lu Su, Jing Gao

机构 * Purdue University（普渡大学）； Georgia Institute of Technology（佐治亚理工学院）

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.LG

Comments EMNLP 2025 Findings

2505.14442 2025-09-22 cs.CL cs.AI 62%

Creative Preference Optimization

Mete Ismayilzada, Antonio Laverghetta, Simone A. Luchini, Reet Patel, Antoine Bosselut, Lonneke van der Plas, Roger Beaty

机构 * EPFL（苏黎世联邦理工学院）； Università della Svizzera Italiana（瑞士联邦理工学院）； Wesleyan University（韦斯利安大学）； Pennsylvania State University（宾夕法尼亚州立大学）

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

Comments Accepted to EMNLP 2025 Findings

2503.12123 2025-09-22 cs.CL cs.AI 62%

MT-RewardTree: A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling

Zhaopeng Feng, Jiahan Ren, Jiayuan Su, Jiamei Zheng, Hongwei Wang, Zuozhu Liu

机构 * Zhejiang University（浙江大学）

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

Comments EMNLP 2025 Findings. Project page:https://sabijun.github.io/MT_RewardTreePage

2412.11006 2025-09-22 cs.LG cs.CL 62%

Entropy-Regularized Process Reward Model

Hanning Zhang, Pengcheng Wang, Shizhe Diao, Yong Lin, Rui Pan, Hanze Dong, Dylan Zhang, Pavlo Molchanov, Tong Zhang

机构 * University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）； University of Toronto（多伦多大学）； NVIDIA（英伟达）； Princeton University（普林斯顿大学）； Salesforce Research（Salesforce研究）

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

Comments Upate TMLR version

2503.01658 2025-09-18 cs.LG cs.AI cs.IR 62%

CoPL: Collaborative Preference Learning for Personalizing LLMs

Youngbin Choi, Seunghyuk Cho, Minjong Lee, MoonJeong Park, Yesong Ko, Jungseul Ok, Dongwoo Kim

机构 * Graduate School of Artificial Intelligence, POSTECH（POSTECH人工智能研究生院）； Department of Computer Science and Engineering, POSTECH（POSTECH计算机科学与工程系）

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

Comments 19pages, 13 figures, 11 tables

2503.12301 2025-09-16 cs.LG cs.CL 62%

One Goal, Many Challenges: Robust Preference Optimization Amid Content-Aware and Multi-Source Noise

Amirabbas Afzali, Amirhossein Afsharrad, Seyed Shahabeddin Mousavi, Sanjay Lall

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

2503.01062 2025-09-16 cs.LG cs.AI 62%

Offline RLAIF: Piloting VLM Feedback for RL via SFO

Jacob Beck

机构 * Jacob Beck 1

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

Comments Code is provided at https://github.com/jacooba/OfflineRLAIF

Journal ref Published at The RLC 2025 Workshop on Reinforcement Learning Beyond Rewards: Ingredients for Developing Generalist Agents

2411.08302 2025-09-12 cs.CL cs.AI 62%

RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution

Jiahui Li, Lin Li, Tai-wei Chang, Kun Kuang, Long Chen, Jun Zhou, Cheng Yang

机构 * Zhejiang University（浙江大学）； Ant Group（蚂蚁集团）； The Hong Kong University of Science and Technology（香港科技大学）

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

2508.19576 2025-09-09 cs.AI cs.LG 62%

ReST-RL: Achieving Accurate Code Reasoning of LLMs with Optimized Self-Training and Decoding

Sining Zhoubian, Dan Zhang, Jie Tang

机构 * The Knowledge Engineering Group (KEG), Tsinghua University（清华大学知识工程小组（KEG）、清华大学）

专题命中偏好对齐：DPO(abstract)；分类 cs.AI、cs.LG

Comments 21 pages, 4 figures

2506.07963 2025-09-09 cs.AI cs.CL cs.CV 62%

SUDER: Self-Improving Unified Large Multimodal Models for Understanding and Generation with Dual Self-Rewards

Jixiang Hong, Yiran Zhang, Guanzhong Wang, Yi Liu, Ji-Rong Wen, Rui Yan

机构 * Gaoling School of Artificial Intelligence, Renmin University of China（中国人民大学人工智能学院）； School of Computer Science（计算机科学学院）； Baidu Inc.（百度公司）； School of Computer Science, Wuhan University（武汉大学计算机学院）

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

2509.05605 2025-09-09 cs.CL cs.AI 62%

Icon$^{2}$: Aligning Large Language Models Using Self-Synthetic Preference Data via Inherent Regulation

Qiyuan Chen, Hongsen Huang, Qian Shao, Jiahe Chen, Jintai Chen, Hongxia Xu, Renjie Hua, Ren Chuan, Jian Wu

机构 * Zhejiang University（浙江大学）； Soochow Securities Co.,Ltd.（苏州证券有限公司）； HKUST(GZ)（香港科技大学（广州））； Nanjing University（南京大学）

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

Comments EMNLP 2025 Main

2508.17637 2025-08-26 cs.CL cs.AI 62%

Weights-Rotated Preference Optimization for Large Language Models

Chenxu Yang, Ruipeng Jia, Mingyu Zheng, Naibin Gu, Zheng Lin, Siyuan Chen, Weichong Yin, Hua Wu, Weiping Wang

机构 * Institute of Information Engineering, Chinese Academy of Sciences（中国科学院信息工程研究所）； School of Cyber Security, University of Chinese Academy of Sciences（中国科学院大学网络安全学院）； Baidu Inc.（百度公司）

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.AI

Comments EMNLP 2025

2508.15652 2025-08-26 cs.AI cs.IT cs.LG cs.MA math.IT 62%

Understanding Action Effects through Instrumental Empowerment in Multi-Agent Reinforcement Learning

Ardian Selmonaj, Miroslav Strupl, Oleg Szehr, Alessandro Antonucci

机构 * Istituto Dalle Molle di Studi sull’Intelligenza Artificiale (IDSIA), USI-SUPSI（日内瓦人工智能研究所（IDSIA）、USI-SUPSI）

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

Comments European Conference on Artificial Intelligence (ECAI) 2025

2508.15790 2025-08-25 cs.CL cs.AI 62%

KG-o1: Enhancing Multi-hop Question Answering in Large Language Models via Knowledge Graph Integration

Nan Wang, Yongqi Fan, yansha zhu, ZongYu Wang, Xuezhi Cao, Xinyan He, Haiyun Jiang, Tong Ruan, Jingping Liu

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.AI

2502.01819 2025-08-25 cs.LG cs.AI math.OC 62%

Score as Action: Fine-Tuning Diffusion Generative Models by Continuous-time Reinforcement Learning

Hanyang Zhao, Haoxian Chen, Ji Zhang, David D. Yao, Wenpin Tang

机构 * StonyBrook（石溪大学）

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

Comments arXiv admin note: text overlap with arXiv:2409.08400

2508.15449 2025-08-22 cs.LG cs.AI 62%

Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection

Chengcan Wu, Zeming Wei, Huanran Chen, Yinpeng Dong, Meng Sun

机构 * Peking University（北京大学）； Tsinghua University（清华大学）

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

Comments 10 pages, 9 figures, Under review as a full paper at AAAI 2026. A preliminary version is under review at the NeurIPS 2025 Workshop on Reliable ML from Unreliable Data

2504.16438 2025-08-20 cs.LG cs.AI cs.CR cs.DC 62%

POPri: Private Federated Learning using Preference-Optimized Synthetic Data

Charlie Hou, Mei-Yu Wang, Yige Zhu, Daniel Lazar, Giulia Fanti

机构 * Pittsburgh Supercomputing Center, Pittsburgh, USA（匹兹堡超级计算中心）； Department of ECE, Carnegie Mellon University, Pittsburgh, PA（电子工程系，卡内基梅隆大学）

专题命中偏好对齐：DPO(abstract)；分类 cs.AI、cs.LG

Comments ICML 2025 camera-ready

2508.13250 2025-08-20 cs.AI cs.CL cs.IR 62%

Explicit v.s. Implicit Memory: Exploring Multi-hop Complex Reasoning Over Personalized Information

Zeyu Zhang, Yang Zhang, Haoran Tan, Rui Li, Xu Chen

机构 * Gaoling School of Artificial Intelligence, Renmin University of China（中国人民大学人工智能学院）； National University of Singapore（新加坡国立大学）

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

Comments 15 pages, 13 figures, 3 tables

2508.13189 2025-08-20 stat.ML cs.AI cs.LG 62%

Preference Models assume Proportional Hazards of Utilities

Chirag Nagpal

机构 * Meta Superintelligence Labs (MSL)（Meta超智能实验室）

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

2508.07970 2025-08-19 cs.LG cs.AI 62%

WeChat-YATT: A Scalable, Simple, Efficient, and Production Ready Training Library

Junyu Wu, Weiming Chang, Xiaotao Liu, Guanyou He, Tingfeng Xian, Haoqiang Hong, Boqi Chen, Hongtao Tian, Tao Yang, Yunsheng Shi, Feng Lin, Ting Yao, Jiatao Xu

机构 * Tencent（腾讯）

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

Comments arXiv admin note: substantial text overlap with arXiv:2507.22789

2503.22402 2025-08-18 cs.DB cs.AI cs.CL 62%

EllieSQL: Cost-Efficient Text-to-SQL with Complexity-Aware Routing

Yizhang Zhu, Runzhi Jiang, Boyan Li, Nan Tang, Yuyu Luo

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.AI

Comments COLM 2025

2508.10701 2025-08-15 cs.LG cs.AI 62%

REFN: A Reinforcement-Learning-From-Network Framework against 1-day/n-day Exploitations

Tianlong Yu, Lihong Liu, Ziyi Zhou, Fudu Xing, Kailong Wang, Yang Yang

机构 * School of Artificial Intelligence, Hubei University（湖北大学人工智能学院）； Huazhong University of Science and Technology（华中科技大学）； University of Southern California（美国南加州大学）

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

2504.14493 2025-08-15 cs.IR cs.AI cs.LG 62%

FinSage: A Multi-aspect RAG System for Financial Filings Question Answering

Xinyu Wang, Jijun Chi, Zhenghan Tai, Tung Sum Thomas Kwok, Muzhi Li, Zhuhong Li, Hailin He, Yuchen Hua, Peng Lu, Suyuchen Wang, Yihong Wu, Jerry Huang, Jingrui Tian, Fengran Mo, Yufei Cui, Ling Zhou

机构 * 1SimpleWay.AI 2McGill University 3University of Toronto 4University of California, Los Angeles 5The Chinese University of Hong Kong 6Duke University 7Universit\'e de Montr\'eal 8Mila - Quebec AI Institute 9Noah's Ark Lab 10CG Matrix Technology Limited ； 1SimpleWay.AI 2McGill University 3University of Toronto 4University of California, Los Angeles ； 5The Chinese University of Hong Kong 6Duke University 7Universit\'e de Montr\'eal 8Mila - Quebec AI Institute ； 9Noah's Ark Lab 10CG Matrix Technology Limited

专题命中偏好对齐：DPO(abstract)；分类 cs.AI、cs.LG

Comments Accepted at the 34th ACM International Conference on Information and Knowledge Management (CIKM2025)

2508.06963 2025-08-12 cs.AI cs.LG 62%

MASteer: Multi-Agent Adaptive Steer Strategy for End-to-End LLM Trustworthiness Repair

Changqing Li, Tianlin Li, Xiaohan Zhang, Aishan Liu, Li Pan

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

2508.06026 2025-08-11 cs.CL cs.AI 62%

Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future

Yidong Wang, Xin Wang, Cunxiang Wang, Junfeng Fang, Qiufeng Wang, Jianing Chu, Xuran Meng, Shuxun Yang, Libo Qin, Yue Zhang, Wei Ye, Shikun Zhang

专题命中偏好对齐：DPO(abstract)；分类 cs.CL、cs.AI

Comments 12 pages, 5 figures

1. 偏好对齐 3278 篇

Turning Internal Gap into Self-Improvement: Promoting the Generation-Understanding Unification in MLLMs

Cognitive-Level Adaptive Generation via Capability-Aware Retrieval and Style Adaptation

Cross-Cultural Transfer of Commonsense Reasoning in LLMs: Evidence from the Arab World

Brittleness and Promise: Knowledge Graph Based Reward Modeling for Diagnostic Reasoning

Mitigating Hallucinations in Large Vision-Language Models via Entity-Centric Multimodal Preference Optimization

Towards Universal Debiasing for Language Models-based Tabular Data Generation

Creative Preference Optimization

MT-RewardTree: A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling

Entropy-Regularized Process Reward Model

CoPL: Collaborative Preference Learning for Personalizing LLMs

One Goal, Many Challenges: Robust Preference Optimization Amid Content-Aware and Multi-Source Noise

Offline RLAIF: Piloting VLM Feedback for RL via SFO

RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution

ReST-RL: Achieving Accurate Code Reasoning of LLMs with Optimized Self-Training and Decoding

SUDER: Self-Improving Unified Large Multimodal Models for Understanding and Generation with Dual Self-Rewards

Icon$^{2}$: Aligning Large Language Models Using Self-Synthetic Preference Data via Inherent Regulation

Weights-Rotated Preference Optimization for Large Language Models

Understanding Action Effects through Instrumental Empowerment in Multi-Agent Reinforcement Learning

KG-o1: Enhancing Multi-hop Question Answering in Large Language Models via Knowledge Graph Integration

Score as Action: Fine-Tuning Diffusion Generative Models by Continuous-time Reinforcement Learning

Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection

POPri: Private Federated Learning using Preference-Optimized Synthetic Data

Explicit v.s. Implicit Memory: Exploring Multi-hop Complex Reasoning Over Personalized Information

Preference Models assume Proportional Hazards of Utilities

WeChat-YATT: A Scalable, Simple, Efficient, and Production Ready Training Library

EllieSQL: Cost-Efficient Text-to-SQL with Complexity-Aware Routing

REFN: A Reinforcement-Learning-From-Network Framework against 1-day/n-day Exploitations

FinSage: A Multi-aspect RAG System for Financial Filings Question Answering

MASteer: Multi-Agent Adaptive Steer Strategy for End-to-End LLM Trustworthiness Repair

Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future