1 paper
Yuxiang Zhang, Ding Cao, Shuting Cui +3
On-policy distillation (OPD) improves large language model reasoning by training students on their own rollouts with dense token-wise supervision from the teacher. However, token-w…