Agentic on-policy distillation (OPD) trains a student agent on its own trajectories while a teacher scores every token. A natural explanation of where it goe…
机构:NVIDIA
来源:arXiv 2610.28586 | AI4Papers 论文推荐平台