1 paper
Changhui Sun, Lanbo Liu, Hang Lei +11
On-policy distillation (OPD) aligns a student model with a teacher's logit distribution on student-generated trajectories. This approach has achieved strong empirical gains and can…