Relay-OPD Lets a Teacher Briefly Rescue a Small Model’s Failed Reasoning

The method detects when a student enters a bad reasoning direction, lets the teacher take a short leg, and hands the trajectory back. The authors report a 5.73% average gain over standard OPD for a 1.7B student.

The problem

The July 28 preprint Pass the Baton studies failed-prefix behavior in on-policy distillation. When a student makes an early reasoning error, later tokens build on the wrong direction and teacher feedback on that trajectory can become inefficient.

How Relay-OPD works

The researchers use disagreement between teacher and student continuation tendencies as a label-free handoff signal. At selected points, the teacher generates a short corrective leg before the student resumes. A limited relay budget avoids replacing the entire trajectory and keeps training close to the student policy.

Reported results

Experiments use Qwen3-4B-Instruct-2507 as teacher and 0.6B or 1.7B students across eight mathematical-reasoning benchmarks. The authors report that the 1.7B model improves by 5.73% on average over standard OPD and 1.49% over FastOPD, while training trajectories become more than 50% shorter.

Limits

This is a non-peer-reviewed preprint focused on mathematics and particular Qwen models. Transfer to code, dialogue, and tool use remains unproven, and total training economics must include teacher calls. Implementations should compare accuracy, teacher tokens, wall-clock time, and failure modes together.

Primary source