Search papers, labs, and topics across Lattice.
This paper introduces reinforcement learning with metacognitive feedback (RLMF) to enhance the metacognitive abilities of large language models (LLMs), addressing their tendency to hallucinate and misrepresent uncertainty. By refining completion rankings based on self-judged performance and selecting high-value training examples, RLMF significantly improves the calibration of models' confidence scores. Experimental results demonstrate that RLMF achieves state-of-the-art performance in faithful calibration while maintaining accuracy, surpassing standard reinforcement learning methods by up to 63%.
RLMF not only boosts LLMs' ability to accurately express uncertainty but also enhances their self-assessment capabilities, fundamentally reshaping their trustworthiness.
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.
Google Research