Pass the Baton: Trajectory-Relayed On-Policy Distillation
Relay-OPD lets the teacher briefly take over when a student’s reasoning prefix starts to fail.
The paper targets “prefix failure” in on-policy distillation, where a wrong early step makes later supervision unreliable. Its trigger is label-free: it uses the observed gap between how the teacher redirects from a failed prefix and how the student keeps going. In math-reasoning tests with Qwen3 teachers and smaller Qwen3 students, the method beats standard OPD by 5.73% on average for the 1.7B student and cuts training trajectory length by more than half. HF Daily Papers' note
The paper targets “prefix failure” in on-policy distillation, where a wrong early step makes later supervision unreliable. Its trigger is label-free: it uses the observed gap between how the teacher redirects from a failed prefix and how the student keeps going. In math-reasoning tests with Qwen3 teachers and smaller Qwen3 students, the method beats standard OPD by 5.73% on average for the 1.7B student and cuts training trajectory length by more than half. HF Daily Papers' note
score 5