Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs

2026-06-30

AI Summary

This work addresses the reliability challenges of large language models (LLMs) stemming from high-confidence hallucinations, ambiguous knowledge boundaries, and inaccurate uncertainty reporting. To mitigate these issues, the authors propose a Reinforcement Learning framework with Metacognitive Feedback (RLMF), employing a two-stage decoupled approach: first, model self-evaluation scores are leveraged as ranking signals for preference optimization to calibrate confidence faithfulness; second, a metacognitive data selection mechanism identifies high-value samples, and the calibrated confidence is mapped into context-adaptive natural language expressions of uncertainty. Experimental results demonstrate that RLMF achieves state-of-the-art calibration performance across diverse tasks, improving up to 63% over standard reinforcement learning baselines while preserving accuracy and significantly enhancing the model’s ability to assess and articulate its own competence boundaries.

Abstract

Metacognition is a critical component of intelligence that describes the ability to monitor and regulate one's own cognitive processes. Yet LLMs exhibit systemic deficiencies in key metacognitive faculties: they hallucinate with high confidence, fail to recognize knowledge boundaries, and misrepresent their internal uncertainty--undermining trustworthiness and reliability. Since monitoring task performance and adapting behavior accordingly are central to metacognition, we posit that models capable of accurately judging their own performance are better positioned to improve it. We operationalize this idea via two novel mechanisms: reinforcement learning with metacognitive feedback (RLMF), a paradigm to refine completion rankings during preference optimization based on the quality of a model's self-judgments of performance, and metacognitive data selection, which uses similar self-judgments to identify high-value training examples, outperforming naive active learning. We apply these innovations to the problem of faithful calibration (FC), a task that is itself fundamentally metacognitive: the goal is to align expressed with intrinsic uncertainty, difficult even for frontier LLMs. We adopt a two-stage, decoupled approach, first using these methods to calibrate the faithfulness of models' self-reported confidence scores, then mapping to natural, context-adaptable linguistic uncertainty via targeted output editing. Extensive experiments show RLMF achieves generalizable, state-of-the-art FC on diverse tasks while preserving accuracy. Further, RLMF surpasses standard RL by up to 63% while enhancing models' ability to assess and express their own capability limits. This positions RLMF as a promising paradigm to enhance LLM metacognition toward improved abilities and alignment, and suggests metacognitive performance as an effective RL signal to overcome limits of prior intrinsic feedback methods.

Problem

• metacognition
• faithful calibration
• uncertainty expression
• large language models
• hallucination

Innovation

• metacognition
• reinforcement learning
• faithful calibration
• uncertainty expression
• self-judgment