Large language models (LLMs) can be dishonest when reporting on their actions and beliefs -- for example, they may overstate their confidence in factual claims or cover up evidence of covert actions. Such dishonesty may arise due to the effects of reinforcement learning (RL), where challenges with reward shaping can result in a training process that inadvertently incentivizes the model to lie or misrepresent its actions.
In this work we propose a method for eliciting an honest expression of an LLM's shortcomings via a self-reported *confession*. A confession is an output, provided upon request after a model's original answer, that is meant to serve as a full account of the model's compliance with the letter and spirit of its policies and instructions. The reward assigned to a confession during training is solely based on its honesty, and does not impact positively or negatively the main answer's reward. As long as the "path of least resistance" for maximizing confession reward is to surface misbehavior rather than covering it up, this incentivizes models to be honest in their confessions. Our findings provide some justification this empirical assumption, especially in the case of egregious model misbehavior.
To demonstrate the viability of our approach, we train GPT-5-Thinking to produce confessions, and we evaluate its honesty in out-of-distribution scenarios measuring hallucination, instruction following, scheming, and reward hacking. We find that when the model lies or omits shortcomings in its "main" answer, it often confesses to these behaviors honestly, and this confession honesty modestly improves with training. Confessions can enable a number of inference-time interventions including monitoring, rejection sampling, and surfacing issues to the user.
BeHonest: Benchmarking Honesty in Large Language Models
Shanghai Jiao Tong University（上海交通大学）；Carnegie Mellon University（卡内基梅隆大学）；Fudan University（复旦大学）；Shanghai AI Laboratory（上海人工智能实验室）；Generative AI Research Lab (GAIR)（生成式人工智能研究实验室（GAIR））
The University of Hong Kong（香港大学）；Tsinghua University（清华大学）；University of Illinois at Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；University of Virginia（弗吉尼亚大学）；Peking University（北京大学）；WeChat AI（微信人工智能）；The Chinese University of Hong Kong（香港中文大学）
HonestLLM: Toward an Honest and Helpful Large Language Model
MBZUAI（Mohamed bin Zayed 人工智能大学）；Huazhong University of Science and Technology（华中科技大学）；University of Notre Dame（圣母大学）；University of Washington（华盛顿大学）；Peking University（北京大学）；Lehigh University（理海大学）
Do Large Language Models Know What They Don't Know II? A Fully Behavioral, Non-Cognitive Measure of Epistemic Honesty
Tarsus University（塔尔苏斯大学）；Arizona State University (ASU)（亚利桑那州立大学）；Institute for Social Science Research（社会科学研究所）；School of Human Evolution and Social Change（人类进化与社会变革学院）；School of Computing and Augmented Intelligence (SCAI)（计算与增强智能学院）
Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Control
State Key Lab of CAD&CG, Zhejiang University（浙江大学CAD&CG国家重点实验室）；Alibaba Cloud（阿里云）；Hong Kong Baptist University（香港浸会大学）；School of Software Technology, Zhejiang University（浙江大学软件学院）；Zhiyuan Research Institute（智源研究院）；Fabu Inc.（fabu公司）
From Natural Language to Verified Code: Toward AI Assisted Problem-to-Code Generation with Dafny-Based Formal Verification
Department of Computer Science, The University of Alabama（阿拉巴马大学计算机科学系）；The University of Alabama（阿拉巴马大学）；Alabama Water Institute, The University of Alabama（阿拉巴马水研究院，阿拉巴马大学）
Crisis-Bench: Benchmarking Strategic Ambiguity and Reputation Management in Large Language Models
Hong Kong University of Science and Technology（香港科技大学）；Hong Kong Baptist University（香港 Baptist 大学）；National University of Singapore（新加坡国立大学）；Imperial College London（伦敦帝国理工学院）
