1 paper · 1 filter
Zhennan Shen, Yanshu Li, Qingyu Yin +6
On-policy distillation (OPD) is increasingly used to improve large language model reasoning, but its training dynamics remain poorly understood. We characterize the trajectory of O…