On-policy power distillation (OPPD) trains a model to approximate a sharpened sequence-level answer distribution in one generation.
OPPD distills expensive answer search into one model generation
2026-10-06Frontier watchTRIAL
On-policy power distillation (OPPD) trains a model to approximate a sharpened sequence-level answer distribution in one generation. A frozen teacher weights candidates from sequential Monte Carlo, and those weights supervise the student without reference answers. The authors report gains up to 23 points on MATH-500 and 27.3 on GSM8K over the base checkpoint, plus improvements beyond matched-budget GRPO. Code is public.
What this means for you
For post-training teams, OPPD may move inference-time search cost into training while remaining complementary to verified-reward RL. The open question is whether gains survive different checkpoints and compute accounting.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | sequential Monte Carlo samples candidates; a frozen teacher power distribution weights maximum-likelihood updates |
|---|---|
| Reported Gains | up to +23.0 MATH-500 and +27.3 GSM8K over the untrained model |
| Comparison | one generation beats published 64-candidate power sampling by 2.4 and 3.5 points; exceeds matched-budget GRPO on three math benchmarks |
| Artifact | public code repository; results author-reported |
Technical impact 4
Engineering value 3
Adoption signal 1
Maturity 2
Verification cost 3
Risk 3