The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation

TL;DR

This paper identifies a calibration issue in on-policy distillation of language models, proposes a new framework CaOPD to improve confidence calibration without sacrificing task performance, and demonstrates its effectiveness across models and domains.

Contribution

It introduces CaOPD, a calibration-aware distillation method that aligns confidence estimates with deployment conditions, addressing overconfidence in existing approaches.

Findings

CaOPD achieves Pareto-optimal calibration and maintains competitive capability.

CaOPD generalizes robustly under out-of-distribution and continual learning.

Teacher supervision under privileged context causes systematic optimism bias.

Abstract

On-policy distillation (OPD) is an increasingly important paradigm for post-training language models. However, we identify a pervasive Scaling Law of Miscalibration: while OPD effectively improves task accuracy, it systematically traps models in severe overconfidence. We trace this failure to an information mismatch: teacher supervision is formed under privileged context available during training, whereas the deployed model must report confidence using only deployment-time information. We formalize this perspective theoretically, showing that teacher-conditioned success is generally not a valid target for deployment-time confidence and that helpful privileged context induces entropy collapse and a systematic optimism bias. To address this, we propose a calibration-aware OPD framework, CaOPD, that estimates empirical confidence from model rollouts, replaces self-reported confidence with this student-grounded target, and distills the revised response through the same self-distillation pipeline. Experiments across various models and domains show that CaOPD achieves Pareto-optimal calibration while maintaining competitive capability, generalizing robustly under out-of-distribution and continual learning. Our findings highlight that capability distillation does not imply calibrated confidence, and that confidence should be treated as an essential objective in post-training. Code: https://github.com/SalesforceAIResearch/CaOPD