What's this about?
On-policy distillation normally requires a teacher at least as strong as the student. This paper shows how *weaker* teachers can still improve frontier models (41↑).
What exactly is on-policy distillation?
The student generates its own attempts, and the teacher grades each token of those attempts — denser feedback than a single pass/fail reward. The catch was always needing a better teacher; at the frontier there isn't one.
Why does it matter?
If weak-to-strong transfer works in OPD, frontier labs get a scalable-oversight tool: today's model helps train tomorrow's without being smarter than it — a small piece of the alignment puzzle, from the training-efficiency direction. Related: SAF-OPD (24↑) fuses OPD with RL rewards to break the teacher ceiling the other way.