Despite the recent advancement in large language models (LLMs) and their high performances across numerous benchmarks, recent research has unveiled that LLMs suffer from hallucinations and unfaithful reasoning. This work studies a specific type of hallucination induced by semantic associations. Specifically, we investigate to what extent LLMs take shortcuts from certain keyword/entity biases in the prompt instead of following the correct reasoning path. To quantify this phenomenon, we propose a novel probing method and benchmark called EureQA. We start from questions that LLMs will answer correctly with utmost certainty, and mask the important entity with evidence sentence recursively, asking models to find masked entities according to a chain of evidence before answering the question.
During the construction of the evidence, we purposefully replace semantic clues (entities) that may lead to the correct answer with distractor clues (evidence) that will not directly lead to the correct answer but require a chain-like reasoning process. We evaluate if models can follow the correct reasoning chain instead of short-cutting through distractor clues. We find that existing LLMs lack the necessary capabilities to follow correct reasoning paths and resist the attempt of greedy shortcuts. We show that the distractor semantic associations often lead to model hallucination, which is strong evidence that questions the validity of current LLM reasoning.
Exploring and Mitigating Fawning Hallucinations in Large Language Models
School of Medical Technology, Beijing Institute of Technology(北京理工大学医学技术学院);Artificial Intelligence Research Institute, Shenzhen MSU-BIT University(深圳MSU-BIT大学人工智能研究院);Tianjin University(天津大学);The Hong Kong University of Science(香港科学大学)
How to Detect and Defeat Molecular Mirage: A Metric-Driven Benchmark for Hallucination in LLM-based Molecular Comprehension
Shenzhen Graduate School, Peking University(北京大学深圳研究生院);International Digital Economy Academy (IDEA)(国际数字经济学院(IDEA));Pengcheng Laboratory(鹏城实验室);Hong Kong University of Science and Technology(香港科技大学)
Investigating the Influence of Prompt-Specific Shortcuts in AI Generated Text Detection
Seoul National University(首尔大学);Hanyang University(汉阳大学);Ewha Womans University(梨花女子大学);NAVER Search US(NAVER美国搜索部门);NAVER AI LAB(NAVER AI实验室);NAVER Cloud(NAVER云)
A Progressive Training Strategy for Embodied Vision-Language Models to Mitigate Spatio-Temporal Hallucinations
Zhejiang University(浙江大学);The Hong Kong University of Science and Technology(香港科技大学);Qingdao University(青岛大学);The University of Hong Kong(香港大学);Shanghai Jiao Tong University(上海交通大学);The Chinese University of Hong Kong(香港中文大学);University of YYY(YYY大学);Institute of WWW(WWW研究所)
Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs
School of Engineering, Westlake University, Hangzhou, China(西湖大学工程学院);School of Cyberspace Security, Nanjing University of Science and Technology, Nanjing, China(南京理工大学网络安全学院)
Domain Knowledge-Enhanced LLMs for Fraud and Concept Drift Detection
School of Computing and Augmented Intelligence (SCAI), Arizona State University (ASU)(计算与增强智能学院(SCAI),亚利桑那州立大学);Department of Computer Engineering, Tarsus University(计算机工程系,塔鲁斯大学);Minerva CQ and HumaConn AI Consulting(Minerva CQ和HumaConn人工智能咨询);School of Computing and Augmented Intelligence (SCAI), Arizona State University(计算与增强智能学院(SCAI),亚利桑那州立大学)
Language Models can Subtly Deceive Without Lying: A Case Study on Strategic Phrasing in Legislation
Centre for Responsible AI, IIT Madras(IIT马德拉斯责任人工智能中心);Indian Institute of Technology Madras(印度马德拉斯理工学院);Princeton University(普林斯顿大学);Stanford University(斯坦福大学)
What External Knowledge is Preferred by LLMs? Characterizing and Exploring Chain of Evidence in Imperfect Context for Multi-Hop QA
State Key Laboratory of Intelligent Game(智能游戏国家重点实验室);Science and Technology on Integrated Information System Laboratory(集成信息系统技术实验室);Institute of Software Chinese Academy of Sciences(中国科学院软件研究所);University of Chinese Academy of Sciences(中国科学院大学);Nanyang Technological University(南洋理工大学)
Dark Patterns Meet GUI Agents: LLM Agent Susceptibility to Manipulative Interfaces and the Role of Human Oversight
University of Notre Dame(诺丁汉大学);University of Michigan(密歇根大学);Northeastern University(东北大学);University of Washington(华盛顿大学);Johns Hopkins University(约翰霍普金斯大学)
Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective
Gaoling School of Artificial Intelligence(北京理工大学人工智能学院);Engineering Research Center of Next-Generation Intelligent Search and Recommendation(下一代智能搜索与推荐工程研究中心)
Reason-KE++: Aligning the Process, Not Just the Outcome, for Faithful LLM Knowledge Editing
Shanghai Jiao Tong University(上海交通大学);The University of Sydney(悉尼大学);Shenzhen Campus of Sun Yat-sen University(中山大学深圳校区);Nanyang Technological University(南洋理工大学)
Deep Expert Injection for Anchoring Retinal VLMs with Domain-Specific Knowledge
Beijing Institute of Technology, Beijing, China(北京理工大学);National University of Singapore, Singapore(新加坡国立大学);Tsinghua University, Beijing, China(清华大学);The Hong Kong Polytechnic University, Hong Kong(香港理工大学)