What's this about?
On-policy self-distillation — a model teaching itself from its own best outputs — works but is notoriously finicky. This paper shows vanilla OPSD is the β=1 corner case of a whole family of methods, and other members of the family train far more reliably.
What exactly is "self-distillation"?
Distillation normally means a small model learning from a big teacher. Self-distillation makes the model its own teacher: generate answers, keep the good ones, train on them. The known failure mode is feedback loops amplifying the model's own quirks — the β parameter effectively controls how much it trusts itself.
Why it matters
Self-improvement pipelines are how labs squeeze gains without new data; a principled fix for their brittleness (21 upvotes) feeds directly into that machinery.