Zhejiang, Alibaba Relay-OPD Hands Off to Teacher, Beats OPD 5.73%

A team from Zhejiang University and Alibaba's Yuvion Team has a small but interesting result on how you actually get useful signal out of on-policy distillation. Their preprint, Pass the Baton, targets a failure mode they call prefix failure: once the small student model commits to a wrong reasoning step early on, everything it generates after that is wasted supervision. Their fix is to detect the moment the teacher and student would diverge, hand generation back to the teacher for a short leg, then let the student resume.

The measured effect is more concrete than the story often is with these training-side tweaks. On eight math reasoning benchmarks, using Qwen3-4B-Instruct-2507 as teacher and Qwen3-1.7B-Non-Thinking as student, Relay-OPD reaches an average accuracy of 46.96, beating standard on-policy distillation by +5.73% and the strongest baseline, FastOPD, by +1.49%. The 0.6B student also gains, +3.01% over OPD. What is arguably more interesting is the compute side: the average rollout drops from 4,658 tokens on OPD to 2,296, a 50.7% cut, and the best checkpoint arrives at step 35 instead of step 55.

Why this matters if you are not personally training small models: the trend across the last year has been that whoever can extract the most signal per training token from a stronger teacher wins the small-model race. A trick that cuts training trajectory length by more than half while lifting math scores is exactly the kind of unglamorous efficiency lever that moves the frontier for the 1-2B open-weights class most people can actually run.

The honest caveats are the ones the authors themselves flag. The evaluation is only mathematical reasoning, only Qwen3 teacher-student pairs, and the method presumes a teacher whose continuations redirect the student more reliably than the student's own, so gains are expected to shrink as that capability gap narrows. What the paper does not tell you is how it holds up on code or agentic tool-use tasks, whether the reflection-token trigger set transfers to model families other than Qwen3, or where the wall-clock savings actually land once you count the teacher's takeover cost.

For anyone distilling smaller open models, the useful signal is that the wasted-compute story on OPD is real, and there is now a plausible online mechanism for cutting it.