On-policy distillation gains come from suppressing low-prob tokens, not the teacher
burny_tech · x · 2026-09-03
The paper "Does On-Policy Distillation Really Distill?" (Yi Ding, Ruqi Zhang) quantitatively dissects where on-policy distillation (OPD) gains actually come from, with surprising results:
- Teacher signal is noisy: teachers score student-generated trajectories that are inherently off-policy, and supervision noise actually increases with teacher scale.
- Students are insensitive to the noise: removing noisy supervision yields comparable performance — the teacher barely distills anything.
- Real source of gains: learning concentrates on low log-probability tokens; a single fixed negative advantage matches teacher-provided scores. OPD largely works by suppressing tail tokens.
The authors propose On-Policy Self-Adaptation (OPSA), a supervision-free method using entropy-adaptive negative advantages: stronger signals at high-entropy positions, tail-token suppression, and even redistribution of probability mass among head tokens.
Results: versus base Qwen3-1.7B, OPSA improves Avg@32 on AIME24 by 35.41 points (263% relative gain), more than doubles Pass@32 across three benchmarks, and beats OPD by 16.77 points on AIME24 Avg@32, generalizing across model families and tasks.
More from Research
- Information Bottleneck podcast to livestream Jonas Geiping on recurrent-depth models — ziv_ravid · 2026-09-03
- ICLR 2027's deanonymization policy wins fans; researcher urges slashing submission cap to 5 — peter_richtarik · 2026-09-03
- motion-bricks.cpp hits GitHub: tiny model generates animations in realtime on a desktop CPU — zhengyiluo · 2026-09-03
- Mostik bridges frontier and small models in latent space, tops ARC-AGI 3 at 1/20th the cost — SimplyAnnisa · 2026-09-03
- Mechanism Design Could Shape AI Behavior Without Understanding Neural Nets, Argues Economist — morqon · 2026-09-03
- Chris Potts Explains His Interpretability Paper in IPAM Workshop Talk — ChrisGPotts · 2026-09-03