A Survey of On-Policy Distillation for LLMs 学习(一)On-Policy Distillation (OPD) trains large language models with teacher guidance on student-generated trajectories, addressing the mismatch between fixed demonstrations and the histories encountered during deployment.