1 paper · 1 filter
Qijia He, Zequn Li, Spencer Luo +6
On-policy distillation (OPD) provides dense token-level supervision on student rollouts, but this supervision can become unreliable when the student and teacher diverge, either bec…