Confidence Calibration

2 items

RESEARCHDEV.to AI·5/10/2026

Distillation that keeps confidence honest

Traditional on-policy distillation (OPD) causes smaller student models to exhibit overconfidence due to the larger teacher model's access to privileged context. New research formalizes this mismatch and proposes CaOPD to rectify this certainty illusion without sacrificing accuracy gains.

Confidence Calibration distillation model training machine learning

RESEARCHarXiv CS.AI·14d ago

Confidence Calibration in Large Language Models

This study investigates confidence calibration in Large Language Models (LLMs) across diverse tasks, finding that current LLMs are overconfident on difficult tests and underconfident on easy ones. The researchers developed LifeEval, a new test to evaluate model calibration across varying levels of difficulty.

Confidence Calibration Overconfidence machine learning large language models