On-policy distillation (OPD) supplies dense token-level supervision by aligning a student to the teacher's next-token distribution on its own generated rollo…
机构:华为
来源:arXiv 2611.07731 | AI4Papers 论文推荐平台