Analysis of On-Policy Distillation Noise and Student Convergence
August 30, 2026
Quantitative analysis of on-policy distillation (OPD) reveals substantial noise in teacher supervision that increases with teacher scale. Despite this, student policies remain insensitive to the noise, converging to comparable performance levels regardless of whether noisy supervision is removed.
HOW THIS AFFECTS YOU
●
researcherThis suggests that OPD might be more robust to noisy teacher signals than previously assumed.