Definition
Plain language
Teaching a model by letting it make its own attempts and correcting each one, instead of just having it copy a master's work.
As stated in the literature
A distillation method where the student generates its own trajectories and training nudges them toward a teacher's outputs, keeping supervision on the student's own distribution; in self-trained verification it succeeds where plain SFT imitation collapses because the deployed verifier walks its own path.
Why it matters: By training a model on its own attempts rather than imitation alone, it avoids the breakdown that happens when a model is deployed and walks paths its copied examples never covered.
For example, instead of handing a student the master's finished essay to copy, you let the student write its own and correct each line as it goes.
Heard on the show
“So the machinery everyone's using here is called on-policy distillation, and Bella, this is the part where the details matter, because the whole paper is a hack on one hidden assumption inside it.”Episode 242 — Making a Vision Model Better by Showing It Blurry Images