Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs

Hugging Face Daily Papers Papers
reinforcement-learning metacognitive-feedback large-language-models calibration uncertainty-expression faithful-calibration active-learning

Summary

This paper introduces reinforcement learning with metacognitive feedback (RLMF) and metacognitive data selection to improve large language model calibration, enabling faithful expression of intrinsic uncertainty and surpassing standard RL by up to 63%.

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.
Original Article
View Cached Full Text

Cached at: 07/01/26, 07:41 AM

Paper page - Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs

Source: https://huggingface.co/papers/2606.32032

Abstract

Reinforcement learning with metacognitive feedback and metacognitive data selection improve large language model calibration by enabling accurate self-assessment of performance and uncertainty.

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 learningwithmetacognitive feedback(RLMF), a paradigm to refine completion rankings duringpreference optimizationbased on the quality of a model’s self-judgments of performance, andmetacognitive data selection, which uses similar self-judgments to identify high-value training examples, outperforming naiveactive learning. We apply these innovations to the problem offaithful calibration(FC), a task that is itself fundamentally metacognitive: the goal is to align expressed withintrinsic 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 priorintrinsic feedback methods.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2606\.32032

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.32032 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.32032 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.32032 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Retrieval-Augmented Linguistic Calibration

arXiv cs.CL

This paper proposes Retrieval-Augmented Linguistic Calibration (RALC), a post-hoc pipeline for calibrating confidence signals in LLMs by modeling linguistic confidence as a distribution and using retrieval-augmented rewriting. It introduces Faithfulness Divergence metric and shows significant improvements across benchmarks.