1 paper
Jinwei Xiao, Zhuowen Han, Yueqing Sun +6
On-policy distillation transfers reasoning ability through dense token-level supervision, yet the nature of the transferable signal remains unclear. We discover that reasoning chai…