mixed-policy-distillation

This paper proposes Mixed-Policy Distillation (MPD), a framework that transfers concise reasoning behaviors from large teacher models to smaller student models, reducing token usage by up to 27.1% while improving performance.