On-policy distillation (OPD) provides dense token-level supervision on student-generated trajectories. Although OPD performs strongly when teacher and studen…
机构:中科院
来源:arXiv 2609.10154 | AI4Papers 论文推荐平台