REHEARSE: Experiential Rehearsal for Verbal Confidence Calibration in Large Language Models
The episode discusses a paper titled "REHEARSE: Experiential Rehearsal for Verbal Confidence Calibration in Large Language Models." The hosts explain how this method improves LLM performance by having the model play a scoring game, learning from its own mistakes, and then applying that learned experience to new questions. This training-free approach significantly reduces calibration error and increases accuracy across various models.