On-Policy Distillation Isn't a Free Lunch: Four Failure Modes Explained

A detailed analysis of four failure modes in OPD/OPSD: structurally uncorrectable early mistakes, stronger teachers being worse, privileged-information-conditioned OPSD failing to transfer, and thinking collapse. Proposes fixes like teacher-refined trajectories and picking teachers based on distributional closeness.

• 后训练
• 模型蒸馏
• On-Policy Distillation
• 失败模式