OPD (Rethinking On-Policy Distillation)

THUNLP's on-policy distillation recipe (~815★ in three months; adopted by MiniCPM5-1B): rethinks how a student learns from a teacher on its own rollouts, improving small-model distillation. Academic THUNLP work — filed under the university, distinct from the OpenBMB-branded MiniCPM line it feeds.