Trust-Region Behavior Blending for On-Policy Distillation
This work provides an incremental improvement for researchers working on on-policy distillation in reinforcement learning, specifically for math-reasoning tasks.
This paper addresses the issue of poor early student rollouts in on-policy distillation (OPD) by proposing Trust-Region behavior Blending (TRB). TRB replaces the early rollout policy with a teacher-proximal behavior policy within a KL trust region, which is then annealed to zero, leading to the strongest average performance across two math-reasoning distillation settings.
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.