1 paper
Haofeng Xu, Junwei Su, Lansong Diao +2
On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level dis…