| title | Trust-Region Behavior Blending for On-Policy Distillation |
|---|---|
| authors | Daniil Plyusov, Alexey Gorbatovski, Alexey Malakhov, Nikita Balagansky, Boris Shaposhnikov, Daria Korotyshova, Daniil Gavrilov |
| date | 2026-05-29 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.31159 |
| affiliation | T-Tech |
| upvotes | 64 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| abstract | 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. |
cite
@misc{plyusov2026trust,
title={Trust-Region Behavior Blending for On-Policy Distillation},
author={Daniil Plyusov and Alexey Gorbatovski and Alexey Malakhov and Nikita Balagansky and Boris Shaposhnikov and Daria Korotyshova and Daniil Gavrilov},
year={2026},
eprint={2605.31159},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.31159}
}