OASIS fixes on-policy self-distillation's scale collapse, gaining 3+ points over OPSD at 8B
Md. Ismail Hossain · hf · 2026-10-01
This work diagnoses why on-policy self-distillation (OPSD) fails to scale for LLM reasoning.
- A factorial analysis shows scaffold correctness matters more than teacher-context correctness; unverified scaffolds create an imitation gap since the teacher can use information unavailable to the student.
- OPSD's gain shrinks from 3.05 points at 1.7B to 0.14 at 8B because it mostly supervises unverified trajectories.
- OASIS keeps the OPSD objective but supervises mostly label-verified on-policy trajectories and replaces written reference solutions with unverified model-generated attempts as teacher context — requiring only final-answer labels.
- On Qwen3-1.7B/4B/8B across AIME 2024/2025 and HMMT 2025, OASIS improves over base by 3.2–3.8 points; at 8B it beats OPSD by 3.05 points, showing verified scaffolds preserve self-distillation at scale.
More from Research
- In business, Type 2 errors often cost more than Type 1 errors — mdancho84 · 2026-11-16
- In risk management, missing a real risk can sink the business — mdancho84 · 2026-11-16
- Cost of Type 2 error: a missed pregnancy delays prenatal care — mdancho84 · 2026-11-16
- Cost of Type 1 error: a false alarm is usually temporary and clarifiable — mdancho84 · 2026-11-16
- Thread defines Type 2 error: false negative means missing a real effect — mdancho84 · 2026-11-16
- Thread defines Type 1 error via pregnancy test: false positive explained — mdancho84 · 2026-11-16