On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training …
机构:Salesforce
来源:arXiv 2611.08453 | AI4Papers 论文推荐平台