Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic Calibration in LLMs

Large Language Models (LLMs) often lack meaningful confidence estimates for their outputs. While base LLMs are known to exhibit next-token…