arxiv:2605.31159

Trust-Region Behavior Blending for On-Policy Distillation

Published on May 29

· Submitted by

Alexey Gorbatovski on Jun 1

#2 Paper of the day

T-Tech

Upvote

Authors:

Alexey Gorbatovski ,

Alexey Malakhov ,

Nikita Balagansky ,

Daniil Gavrilov

Abstract

Trust-Region behavior Blending improves on-policy distillation by replacing early poor-quality student rollouts with teacher-like behavior within a KL trust region during warmup.

AI-generated summary

On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose Trust-Region behavior Blending (TRB), a warmup method that replaces the early rollout policy with the closest-to-teacher behavior policy inside a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD loss unchanged. The KL budget is annealed to zero, so training returns to pure student rollouts after warmup. Across two math-reasoning distillation settings, TRB attains the strongest average among the compared methods.