Relay-OPD: Trajectory relay for on-policy distillation¶
Candidate-policy relay state machine, not eight-GPU training.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Zhejiang University / Alibaba Group Yuvion Team. |
| First public release | 2026-07-28 |
| Author code | ZJU-REAL/Relay-OPD. |
| Local adapter or method key | relay-opd |
| Local reproduction code | src/auto_research/post_training/ |
Paper method and evidence¶
Detect a failing student prefix, let the teacher take over briefly, then return control to the student. A finite relay budget concentrates supervision on critical early positions.
flowchart LR
A[Detect failing prefix] --> B[Bounded teacher relay]
B[Bounded teacher relay] --> C[Resume student generation]Original paper figure¶
Figure shows teacher handoff and student continuation. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Qwen3-1.7B student averages +5.73% versus standard OPD and +1.49% versus FastOPD on eight math benchmarks, with more than 50% shorter training trajectories. No production A/B.
Local reproduction¶
GSM8K candidates exercise prefix failure, teacher handoff, finite relay budget and student resumption against the same untrained policy. Accuracy 0.1719→0.8906, reward 0.3124→0.8867, KL 0.0000→0.5014. Metrics. The teacher is a reproducible candidate-distribution cache, not Qwen3-4B.