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, w

---
**📖 中文解读**
以上内容由AI翻译自英文原文,可能存在不准确之处。建议阅读[原文](https://arxiv.org/abs/2607.26057v1)获取最准确的信息。

---
🔗 **原文链接**: [Pass the Baton: Trajectory-Relayed On-Policy Distillation](https://arxiv.org/abs/2607.26057v1)
🏷️ **转载来源**: ArXiv cs.AI
> 本文由小九AI技术站翻译整理,内容版权归原作者所有。
👤 作者: Haolei Xu, Xiaowen Xu, Haiwen Hong, 倪梓萱, Hongxing Li, Yiwen Qiu, Weiming Lu, Yongliang Shen

---
🐾 **小九锐评**

这篇论文来自arXiv预印本,虽然还没有经过同行评审,但选题方向值得关注。
建议先读中文摘要判断是否相关,再看全文细节。

你对这个话题有什么看法?欢迎在评论区讨论 💬

> _转载自 ArXiv cs.AI,内容版权归原作者所有_

---
⏱️ 2026-07-29 14:01