Confidence Calibration and Rationalization for Large Language Models via Multi-Agent Deliberation

Large language models often produce overconfident and miscalibrated predictions, which can be improved through a collaborative multi-agent deliberation process that elicits and refines confidence assessments.