大模型对齐与安全 - arXivDaily 专题

2512.09340 2025-12-11 cs.AI cs.CV cs.LG 62%

Visual Categorization Across Minds and Models: Cognitive Analysis of Human Labeling and Neuro-Symbolic Integration

跨心灵与模型的视觉分类：人类标注与神经符号整合的认知分析

Chethana Prasad Kabgere

机构 * Georgia Institute of Technology（佐治亚理工学院）

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

AI总结本文通过对比人类与AI在低分辨率图像标注中的表现，探讨了认知策略与神经符号整合方法的异同，旨在推动更可解释和认知对齐的AI系统发展。

Comments 12 pages, 3 figures. Research manuscript based on the final project for CS6795 (Introduction to Cognitive Science), Georgia Tech

2502.12992 2025-12-10 cs.CL cs.AI 62%

B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability

B-cos LM：高效地将预训练语言模型转换以提高可解释性

Yifan Wang, Sukrut Rao, Ji-Ung Lee, Mayank Jobanputra, Vera Demberg

机构 * Saarland University（萨尔兰大学）； Max Planck Institute for Informatics（马克斯·普朗克信息研究所）； Saarland Informatics Campus（萨尔兰计算机科学校区）

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

AI总结本文提出B-cos LM，通过结合B-cos转换和任务微调，提高预训练语言模型的可解释性与效率。

Comments TMLR 12/2025

2512.06406 2025-12-09 cs.AI cs.LG 62%

UncertaintyZoo: A Unified Toolkit for Quantifying Predictive Uncertainty in Deep Learning Systems

UncertaintyZoo: 一个统一的工具包，用于量化深度学习系统中的预测不确定性

Xianzong Wu, Xiaohong Li, Lili Quan, Qiang Hu

机构 * Tianjin University（天津大学）

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

AI总结 UncertaintyZoo是一个统一的工具包，整合了29种不确定性量化方法，用于量化深度学习系统中的预测不确定性。

2510.02324 2025-12-09 cs.CL cs.AI 62%

Hallucination reduction with CASAL: Contrastive Activation Steering For Amortized Learning

通过CASAL减少幻觉：对比激活引导用于近似学习

Wannan, Yang, Xinchi Qiu, Lei Yu, Yuchen Zhang, Aobo Yang, Narine Kokhlikyan, Nicola Cancedda, Diego Garcia-Olano

机构 * Meta Superintelligence Labs（Meta超级智能实验室）； New York University（纽约大学）

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

AI总结 CASAL通过对比激活引导减少LLM幻觉，提升模型可解释性和实用性，适用于多种模型架构。

2512.01420 2025-12-02 cs.CL cs.AI 62%

PromptBridge: Cross-Model Prompt Transfer for Large Language Models

PromptBridge: 跨模型提示迁移用于大型语言模型

Yaxuan Wang, Quan Liu, Zhenting Wang, Zichao Li, Wei Wei, Yang Liu, Yujia Bao

机构 * University of California, Santa Cruz（加州大学圣克ruz分校）； Center for Advanced AI, Accenture（Accenture高级人工智能研究中心）

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

AI总结 PromptBridge通过跨模型提示迁移技术，解决模型切换时提示效果下降的问题，提升下游任务准确性并减少迁移成本。

2511.21762 2025-12-01 cs.CL cs.AI 62%

Factors That Support Grounded Responses in LLM Conversations: A Rapid Review

支持LLM对话中基础响应的因素：一项快速回顾

Gabriele Cesar Iwashima, Claudia Susie Rodrigues, Claudio Dipolitto, Geraldo Xexéo

机构 * (June 2025)（（2025年6月））

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

AI总结本文通过快速回顾识别了支持LLM对话中基础响应的因素，重点分析了推理时间、训练后及强化学习方法，旨在提升LLM响应的准确性和可靠性。

Comments 28 pages, 1 figure, 3 tables

2508.18847 2025-11-26 cs.CL cs.AI 62%

ConfTuner: Training Large Language Models to Express Their Confidence Verbally

ConfTuner: 训练大型语言模型以口头表达其置信度

Yibo Li, Miao Xiong, Jiaying Wu, Bryan Hooi

机构 * National University of Singapore（新加坡国立大学）

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

AI总结 ConfTuner通过引入标记化布里尔分数损失函数，有效提升大型语言模型的置信度校准能力，适用于多种推理任务并泛化至黑盒模型。

Comments Accepted by NeurIPS 2025

2505.16690 2025-11-26 cs.LG cs.AI 62%

Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator

你的预训练大语言模型实际上是一个未监督的置信度校准器

Beier Luo, Shuoyuan Wang, Sharon Li, Hongxin Wei

机构 * Department of Statistics and Data Science, Southern University of Science and Technology（统计与数据科学系，南方科技大学）； Department of Computer Sciences, University of Wisconsin-Madison（计算机科学系，威斯康星大学麦迪逊分校）

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

AI总结本文提出DACA方法，通过无监督方式优化后训练模型的置信度校准，减少不一致预测带来的过度自信问题，提升模型可靠性。

Comments NeurIPS 2025

2511.18874 2025-11-25 cs.AI cs.CV cs.LG cs.MA cs.RO cs.SI 62%

GContextFormer: A global context-aware hybrid multi-head attention approach with scaled additive aggregation for multimodal trajectory prediction

GContextFormer: 一种基于全局上下文的混合多头注意力方法，通过缩放加法聚合实现多模态轨迹预测

Yuzhi Chen, Yuanchang Xie, Lei Zhao, Pan Liu, Yajie Zou, Chen Wang

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

AI总结 GContextFormer通过全局上下文感知的混合多头注意力和缩放加法聚合，实现无地图依赖的多模态轨迹预测，提升鲁棒性和预测精度。

2410.13246 2025-11-21 cs.CL cs.AI 62%

Atomic Calibration of LLMs in Long-Form Generations

大语言模型在长文本生成中的原子校准

Caiqi Zhang, Ruihan Yang, Zhisong Zhang, Xinting Huang, Sen Yang, Dong Yu, Nigel Collier

机构 * University of Cambridge（剑桥大学）； Fudan University（复旦大学）； Tencent AI Lab（腾讯AI实验室）； The Chinese University of Hong Kong（香港中文大学）

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

AI总结本研究提出原子校准方法，用于改进大语言模型在长文本生成中的校准能力，揭示置信度方法与生成过程中置信度变化的关联。

Comments ACL 2025 KnowFM Oral / AACL-IJCNLP 2025

2503.15511 2025-11-20 cs.HC cs.CY cs.LG 62%

The Trust Calibration Maturity Model for Characterizing and Communicating Trustworthiness of AI Systems

Scott T Steinmetz, Asmeret Naugle, Paul Schutte, Matt Sweitzer, Alex Washburne, Lisa Linville, Daniel Krofcheck, Michal Kucer, Samuel Myren

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

Comments 19 pages, 4 figures, 3 tables

2511.15005 2025-11-20 cs.CL cs.AI 62%

Mathematical Analysis of Hallucination Dynamics in Large Language Models: Uncertainty Quantification, Advanced Decoding, and Principled Mitigation

Moses Kiprono

机构 * Catholic University of America（美国天主教大学）

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

Comments 10 pages, theoretical/mathematical LLM research, no figures, intended for peer-reviewed journal

2411.01956 2025-11-18 cs.LG cs.CY stat.ML 62%

EXAGREE: Mitigating Explanation Disagreement with Stakeholder-Aligned Models

Sichao Li, Tommy Liu, Quanling Deng, Amanda S. Barnard

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

2509.15901 2025-11-17 cs.CL cs.AI 62%

Re-FRAME the Meeting Summarization SCOPE: Fact-Based Summarization and Personalization via Questions

Frederic Kirstein, Sonu Kumar, Terry Ruas, Bela Gipp

机构 * University of Göttingen（哥廷根大学）

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

Comments Accepted at EMNLP 2025

2502.12767 2025-11-11 cs.CL cs.AI 62%

R2-KG: General-Purpose Dual-Agent Framework for Reliable Reasoning on Knowledge Graphs

Sumin Jo, Junseong Choi, Jiho Kim, Edward Choi

机构 * KAIST（韩国科学技术院）

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

Comments Accepted to IJCNLP-AACL 2025 Findings

2505.04847 2025-11-07 cs.CL cs.AI 62%

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

Manveer Singh Tamber, Forrest Sheng Bao, Chenyu Xu, Ge Luo, Suleman Kazi, Minseok Bae, Miaoran Li, Ofer Mendelevitch, Renyi Qu, Jimmy Lin

机构 * University of Waterloo（滑铁卢大学）； Vectara（Vectara公司）； Iowa State University（爱荷华州立大学）； Stanford University（斯坦福大学）

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

Comments EMNLP Industry Track 2025

2511.01902 2025-11-05 cs.CY cs.AI 62%

Before the Clinic: Transparent and Operable Design Principles for Healthcare AI

Alexander Bakumenko, Aaron J. Masino, Janine Hoelscher

机构 * Clemson University（克莱姆森大学）

专题命中幻觉与事实性：alignment(abstract)；分类 cs.AI、cs.CY

2412.15189 2025-10-30 cs.CL cs.CY 62%

Face the Facts! Evaluating RAG-based Pipelines for Professional Fact-Checking

Daniel Russo, Stefano Menini, Jacopo Staiano, Marco Guerini

机构 * Fondazione Bruno Kessler（布罗诺·凯塞勒基金会）； University of Trento（特伦托大学）

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

Comments Code and data at https://github.com/drusso98/face-the-facts - Accepted for publication at INLG 2025

2510.23264 2025-10-28 cs.LG cs.AI 62%

PAHQ: Accelerating Automated Circuit Discovery through Mixed-Precision Inference Optimization

Xinhai Wang, Shu Yang, Liangyu Wang, Lin Zhang, Huanyi Xie, Lijie Hu, Di Wang

机构 * King Abdullah University of Science and Technology（卡布斯大学）； Harbin Institute of Technology（哈尔滨工业大学）

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

2510.22261 2025-10-28 cs.LG cs.AI 62%

Epistemic Deep Learning: Enabling Machine Learning Models to Know When They Do Not Know

Shireen Kudukkil Manchingal

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

Comments PhD thesis

2510.22751 2025-10-28 cs.AI cs.CL 62%

Multi-Modal Fact-Verification Framework for Reducing Hallucinations in Large Language Models

Piyushkumar Patel

机构 * Microsoft（微软）

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

2510.22362 2025-10-28 cs.LG cs.CL 62%

Mapping Faithful Reasoning in Language Models

Jiazheng Li, Andreas Damianou, J Rosser, José Luis Redondo García, Konstantina Palla

机构 * King’s College London（伦敦国王学院）； Spotify UK（Spotify英国分公司）； University of Oxford（牛津大学）； Spotify Spain（Spotify西班牙分公司）

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

Comments 9 pages, Accepted to the Mechanistic Interpretability Workshop at NeurIPS 2025

2412.06771 2025-10-27 cs.AI cs.CV cs.LG 62%

Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty

Meera Hahn, Wenjun Zeng, Nithish Kannen, Rich Galt, Kartikeya Badola, Been Kim, Zi Wang

机构 * Google DeepMind（谷歌DeepMind）

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

Journal ref International Conference on Machine Learning, 2025

2510.18918 2025-10-23 cs.CL cs.AI 62%

Misinformation Detection using Large Language Models with Explainability

Jainee Patel, Chintan Bhatt, Himani Trivedi, Thanh Thi Nguyen

机构 * Department of Computer Engineering, LDRP Institute of Technology and Research（计算机工程系，LDRP技术与研究学院）； University of Wollongong（沃林根大学）； Monash University（莫纳什大学）

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

Comments Accepted for publication in the Proceedings of the 8th International Conference on Algorithms, Computing and Artificial Intelligence (ACAI 2025)

2510.15233 2025-10-20 cs.LG cs.AI 62%

Adaptive Individual Uncertainty under Out-Of-Distribution Shift with Expert-Routed Conformal Prediction

Amitesh Badkul, Lei Xie

机构 * Ph.D. Programs in Computer Science（计算机科学博士项目）； The Graduate Center, City University of New York（纽约城市大学研究生中心）； School of Pharmacy and Pharmaceutical Sciences（药学与制药科学学院）； Center for Drug Discovery（药物发现中心）； Northeastern University（东北大学）

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

2505.01997 2025-10-17 cs.LG cs.AI stat.ML 62%

Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach

Jiancong Xiao, Bojian Hou, Zhanliang Wang, Ruochen Jin, Qi Long, Weijie J. Su, Li Shen

机构 * University of Pennsylvania（宾夕法尼亚大学）

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

Journal ref ICML 2025

2505.19234 2025-10-16 cs.AI cs.CL cs.MA 62%

GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling

Jialong Zhou, Lichao Wang, Xiao Yang

机构 * King’s College London（伦敦国王学院）； Beijing Institute of Technology（北京理工大学）； Tsinghua University（清华大学）

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

2504.12324 2025-10-08 cs.CL cs.AI 62%

Cross-Document Cross-Lingual NLI via RST-Enhanced Graph Fusion and Interpretability Prediction

Mengying Yuan, Wenhao Wang, Zixuan Wang, Yujie Huang, Kangli Wei, Fei Li, Chong Teng, Donghong Ji

机构 * Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University（航空信息安全与可信计算重点实验室，教育部，网络安全与工程学院，武汉大学）； Zhejiang University（浙江大学）

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

Comments EMNLP 2025 Main (Camera Ready)

2510.02571 2025-10-06 cs.CV cs.AI cs.CL 62%

How Confident are Video Models? Empowering Video Models to Express their Uncertainty

Zhiting Mei, Ola Shorinwa, Anirudha Majumdar

机构 * Princeton University（普林斯顿大学）

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

2510.01288 2025-10-03 cs.LG cs.AI 62%

Microsaccade-Inspired Probing: Positional Encoding Perturbations Reveal LLM Misbehaviours

Rui Melo, Rui Abreu, Corina S. Pasareanu

机构 * Carnegie Mellon University（卡内基梅隆大学）； FEUP（费拉尔大学）； INESC-ID（葡萄牙里斯本信息技术与创新研究中心）

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

Comments 9 main pages, 13 appendix pages

1. 幻觉与事实性 1732 篇

Visual Categorization Across Minds and Models: Cognitive Analysis of Human Labeling and Neuro-Symbolic Integration

B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability

UncertaintyZoo: A Unified Toolkit for Quantifying Predictive Uncertainty in Deep Learning Systems

Hallucination reduction with CASAL: Contrastive Activation Steering For Amortized Learning

PromptBridge: Cross-Model Prompt Transfer for Large Language Models

Factors That Support Grounded Responses in LLM Conversations: A Rapid Review

ConfTuner: Training Large Language Models to Express Their Confidence Verbally

Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator

GContextFormer: A global context-aware hybrid multi-head attention approach with scaled additive aggregation for multimodal trajectory prediction

Atomic Calibration of LLMs in Long-Form Generations

The Trust Calibration Maturity Model for Characterizing and Communicating Trustworthiness of AI Systems

Mathematical Analysis of Hallucination Dynamics in Large Language Models: Uncertainty Quantification, Advanced Decoding, and Principled Mitigation

EXAGREE: Mitigating Explanation Disagreement with Stakeholder-Aligned Models

Re-FRAME the Meeting Summarization SCOPE: Fact-Based Summarization and Personalization via Questions

R2-KG: General-Purpose Dual-Agent Framework for Reliable Reasoning on Knowledge Graphs

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

Before the Clinic: Transparent and Operable Design Principles for Healthcare AI

Face the Facts! Evaluating RAG-based Pipelines for Professional Fact-Checking

PAHQ: Accelerating Automated Circuit Discovery through Mixed-Precision Inference Optimization

Epistemic Deep Learning: Enabling Machine Learning Models to Know When They Do Not Know

Multi-Modal Fact-Verification Framework for Reducing Hallucinations in Large Language Models

Mapping Faithful Reasoning in Language Models

Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty

Misinformation Detection using Large Language Models with Explainability

Adaptive Individual Uncertainty under Out-Of-Distribution Shift with Expert-Routed Conformal Prediction

Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach

GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling

Cross-Document Cross-Lingual NLI via RST-Enhanced Graph Fusion and Interpretability Prediction

How Confident are Video Models? Empowering Video Models to Express their Uncertainty

Microsaccade-Inspired Probing: Positional Encoding Perturbations Reveal LLM Misbehaviours