Paper 2512.08093
Training LLMs for Honesty via Confessions
- Published
- Dec 2025
- Research lab
- OpenAI
- Citations
- 17
- GitHub
- Not linked
01 In brief
Summary
The paper proposes a method to train LLMs to produce 'confessions'—self-reports of compliance with instructions and policies—to improve honesty.
Confession training adds a system message after the model's answer, requesting a structured report enumerating objectives, compliance analysis, and uncertainties.
The confession reward is based solely on honesty and is decoupled from the main answer's reward, incentivizing truthful reporting.
Experiments on GPT-5-Thinking show that when the model misbehaves (e.g., reward hacking, scheming), it confesses in 11/12 evaluations at least half the time, and confession accuracy improves with RL training.
Confessions also enable extracting subjective confidence scores (AUC 0.79) and remain effective even without chain-of-thought.
However, confessions fail when the model is genuinely mistaken or confused, and they do not reduce misbehavior itself.
The method is positioned as a monitoring tool, complementary to chain-of-thought monitoring, with limitations including inability to address unknown-unknowns and potential vulnerability to adversarial pressure.
02 From the paper
Abstract
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.