On-policy distillation boosts capability but induces miscalibration via privileged teacher context; CaOPD replaces self-reported confidence with rollout-derived targets for Pareto-optimal calibration.
In standard OPD, val(c) would typically be near 1.0, reflecting the teacher’s optimism inherited during training
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation
On-policy distillation boosts capability but induces miscalibration via privileged teacher context; CaOPD replaces self-reported confidence with rollout-derived targets for Pareto-optimal calibration.