The key point: Reinforcement Learning with Metacognitive Feedback (RLMF) enables LLMs to express their own uncertainty in a calibrated manner and outperforms standard RL methods by up to 63 percent.
Researchers have developed a method that trains Large Language Models through Reinforcement Learning with metacognition – to improve their self-assessment and express uncertainty more reliably. The goal: hallucinating AI models should recognise their own knowledge boundaries.
Current Large Language Models suffer from fundamental metacognitive deficits: they generate hallucinations with high confidence, do not reliably recognise their knowledge boundaries, and misrepresent their internal uncertainty. This undermines their trustworthiness, particularly in safety-critical applications. Metacognition – the ability to monitor and regulate one’s own cognitive processes – is a central aspect of intelligence that previous training paradigms have not deliberately addressed.
Researchers have developed two new mechanisms to solve this problem: “Reinforcement Learning with Metacognitive Feedback” (RLMF) refines completion rankings during preference optimisation based on the quality of the model’s self-assessments. A second approach uses similar self-assessments to select high-quality training examples and outperforms naive active learning methods. The method works in two stages: first, the model’s confidence values are calibrated, then these are translated into natural language, context-adaptive uncertainty expressions.
In extensive experiments, RLMF achieved improved calibration of uncertainty representation (Faithful Calibration) across diverse tasks, while preserving model accuracy. The method outperforms standard RL methods by up to 63 percent and demonstrably improves the ability of LLMs to assess and express their own capability boundaries.
Source: arxiv.org · Published 29 June 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.2.