Test-time Calibration Learning for Large Language Model Reasoning

Reliable large language models (LLMs) must not only produce accurate answers but also express confidence that faithfully reflects their probability of being correct.