On-policy distillation (OPD) improves reasoning models by learning the token-level discrepancy between a stronger teacher and an on-policy student. However, …
机构:南大
来源:arXiv 2609.21619 | AI4Papers 论文推荐平台