Trust-Region Behavior Blending for On-Policy Distillation

Trust-Region Behavior Blending is a new warmup method for on-policy distillation.

It replaces weak early student rollouts with the closest teacher-like behavior within a student-centered KL trust region, while keeping the distillation loss unchanged.

The KL constraint is gradually annealed to zero, so training returns to pure student rollouts later.

In math-reasoning distillation tests, it achieved the best average performance among the compared methods.