1 paper · 1 filter
Zhuo Sun, Entong Li, Yanlong Zhao +9
On-policy self-distillation (OPSD) converts feedback into dense token-level supervision on student-generated trajectories, complementing reinforcement learning with sparse outcome…