Trust-Region Behavior Blending for On-Policy Distillation
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866918531440836608 |
|---|---|
| author | Plyusov, Daniil Gorbatovski, Alexey Malakhov, Alexey Balagansky, Nikita Shaposhnikov, Boris Korotyshova, Daria Gavrilov, Daniil |
| author_facet | Plyusov, Daniil Gorbatovski, Alexey Malakhov, Alexey Balagansky, Nikita Shaposhnikov, Boris Korotyshova, Daria Gavrilov, Daniil |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_31159 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Trust-Region Behavior Blending for On-Policy Distillation Plyusov, Daniil Gorbatovski, Alexey Malakhov, Alexey Balagansky, Nikita Shaposhnikov, Boris Korotyshova, Daria Gavrilov, Daniil Machine Learning Artificial Intelligence 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. |
| title | Trust-Region Behavior Blending for On-Policy Distillation |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2605.31159 |