---
title: "传递接力棒：基于轨迹中继的在线蒸馏方法"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.26057v1"
permalink: "https://feed.astralor.com/items/itm_d3d3959cb5cfc5ea6d18ef5f"
published: "2026-07-28T17:59:46.000Z"
score: 78
channel: "ai"
tags: ["Qwen3","在线蒸馏","数学推理"]
---

# 传递接力棒：基于轨迹中继的在线蒸馏方法

## AI 点评

> 这篇文章对从事模型压缩和蒸馏的研究者尤其有价值，它直击在线蒸馏的痛点——前缀错误累积，并且巧妙地利用了师生行为差异作为触发信号，无需额外标注。如果你正在做小模型的推理能力提升，这里提供了一个低成本且有效的新思路。

## 速览

本文提出Relay-OPD，一种改进的在线蒸馏方法，解决传统方法中前缀失败导致的学生模型偏离正确推理方向问题。通过检测学生失败前缀并让教师模型短暂接管产生“教师腿”，然后学生继续生成并进行优化，从而有效纠正错误轨迹。在数学推理基准上，1.7B模型平均超越标准OPD 5.73%，训练轨迹长度减少超50%。

## 关键点

- 识别出前缀失败导致的师生延续不对称性：教师倾向于重定向，学生则沿原方向继续。
- 将这种不对称性转化为无标签的切换触发器，构建接力轨迹。
- 有限接力预算将干预集中在关键早期位置，同时保持对学生策略的偏离尽可能小。
- 在8个数学推理基准上，1.7B模型平均超越标准OPD 5.73%，超越最强基线FastOPD 1.49%。

## 全文

On-policy distillation (OPD) grounds token-level supervision in the student's own trajectory, yet suffers from prefix failure: once the student commits to a wrong reasoning direction, all subsequent generation builds on this deviation, producing misdirected continuations that elicit unreliable supervision and waste compute. We identify a teacher-student continuation asymmetry on failed prefixes, where the teacher tends to redirect while the student continues along the original direction, and convert it into a label-free handoff trigger in Relay On-Policy Distillation (Relay-OPD). During training, Relay-OPD constructs relay trajectories by letting the teacher briefly take over at detected trigger points to produce a teacher leg, after which the student resumes and is optimized on the resulting trajectory. A limited relay budget concentrates intervention on critical early positions while limiting departure from the student policy. With a Qwen3-4B-Instruct-2507 teacher and Qwen3-0.6B/1.7B-Non-Thinking students on eight mathematical reasoning benchmarks, Relay-OPD achieves the best or second-best results on every benchmark, outperforming standard OPD by +5.73% and the strongest baseline FastOPD by +1.49% on average for 1.7B, with consistent gains at 0.6B. Training trajectory length is reduced by over 50%.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_d3d3959cb5cfc5ea6d18ef5f