OPSD Paper: Self-Distillation Gains Don't Need Correct Reference Solutions — None Is Best (+2.2pp)
HanGuo97 · x · 2026-10-28
The OPSD paper (arXiv 2601.18734) lets a single LLM act as both teacher and student: the teacher conditions on privileged information while the student sees only the question, trained via per-token divergence on the student's own rollouts. Ablations comparing correct, absent, and unrelated reference solutions show gains of +1.9/+2.2/+2.0 pp — the correct reference isn't necessary, and none works best. Across math reasoning benchmarks, OPSD beats RL methods on token efficiency and off-policy distillation on accuracy.
More from Research
- NVIDIA's MotionBricks generates game animation frames live, no animation clips needed — TinfoilTricorn · 2026-10-28
- AI Researcher on the Job Market Reflects on the Art of Problem Selection — simi_97k · 2026-10-28
- Open-source model turns any single-camera video into real-time 3D maps at 20 FPS — bennash · 2026-10-28
- Niko McCarty proposes PARADOX, a journal built around single decisive experiments — NikoMcCarty · 2026-10-28
- "3SUM was solved 5 days ago, feels like 10 years": the breakthrough everyone keeps reposting — gleech · 2026-10-28
- Gym-Anything-RT: best frontier model solves only 9% of real-time tasks vs 45% for humans — wellecks · 2026-10-28