Apple ML Research analyzes on-policy distillation for reasoning models

Study examines where on-policy distillation helps or hurts training of reasoning models. Investigation explores teacher model choice and self-distillation contexts.