1 paper
Shangjian Yin, Zehao Zhao, Kavosh Asadi +7
On-policy distillation (OPD) trains a student model by having it generate trajectories, then matching its next-token predictions with an external teacher's next-token predictions.…