Tag
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%.