1 paper · 1 filter
Xincheng Wei, Yifan Ding, Yoshua Li +6
On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard…