🛰️ Daily AI Frontier
‹ back to 2026-09-01

Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement

Research LLMs & Foundation Models

Ranking

Overall 91
Content 100
Popularity 69

Observed public metrics from 1 member.

Merged summary

TL;DR - This paper finds that on-policy distillation gains largely come from suppressing low-probability tokens rather than learning from teacher-provided advantages. It introduces On-Policy Self-Adaptation (OPSA), a teacher-free alternative that substantially improves mathematical reasoning performance.

  • Teacher supervision during on-policy distillation is substantially noisy, with noise increasing as teacher scale grows, yet removing it produces comparable student performance.
  • A fixed negative advantage matches teacher-provided advantages, suggesting the core benefit is suppression of low-probability tail tokens.
  • OPSA applies entropy-adaptive negative advantages at high-entropy positions, suppressing tail tokens and redistributing probability among likely head tokens.
  • On AIME24, OPSA improves Qwen3-1.7B Avg@32 by 35.41 points and outperforms on-policy distillation by 16.77 points.

Sources (1)

Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement

arXiv cs.LG Yi Ding, Ruqi Zhang 2026-08-31 arXiv:2608.31046
Public signals Hugging Face upvotes 96
Providers: Hugging Face · Upvotes 96 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-25 14:25:24.652135 UTC

TL;DR - This paper finds that on-policy distillation gains largely come from suppressing low-probability tokens rather than learning from teacher-provided advantages. It introduces On-Policy Self-Adaptation (OPSA), a teacher-free alternative that substantially improves mathematical reasoning performance.

  • Teacher supervision during on-policy distillation is substantially noisy, with noise increasing as teacher scale grows, yet removing it produces comparable student performance.
  • A fixed negative advantage matches teacher-provided advantages, suggesting the core benefit is suppression of low-probability tail tokens.
  • OPSA applies entropy-adaptive negative advantages at high-entropy positions, suppressing tail tokens and redistributing probability among likely head tokens.
  • On AIME24, OPSA improves Qwen3-1.7B Avg@32 by 35.41 points and outperforms on-policy distillation by 16.77 points.
item →