RL with metacognitive feedback elicits faithful uncertainty expression in LLMs

arXiv·medium signalXBlueskyLinkedInCopy link

LLMs hallucinate with high confidence, fail to recognize knowledge boundaries, and misrepresent internal uncertainty. This paper trains models with metacognitive feedback so they monitor task performance and adapt confidence accordingly, producing more faithful uncertainty statements rather than superficially calibrated ones. For production deployments, better-calibrated 'I don't know' behavior directly reduces confidently-wrong failures in agent and RAG stacks.