重新思考跨分词器在线策略蒸馏：从对齐覆盖到监督可靠性

AI 摘要

该论文研究跨分词器（cross-tokenizer）的在线策略蒸馏（On-Policy Distillation），探讨扩大师生预测对齐覆盖范围是否能提升学习效果。作者在数学推理和代码生成任务上测试了三组异构师生模型对，发现严格的1:1对齐分组已覆盖大部分学生生成token，且共享词表保留了几乎全部概率质量。实验表明，将反向KL限制在学生选定的共享词表top-16子集即可达到与全共享词表OPD相当的准确率，而通过MSE对不匹配组的span对数概率进行监督虽实现完全覆盖却反而降低准确率。作者据此提出应从最大化对齐覆盖转向优先保证监督可靠性。

正文节选

Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability Abstract On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels. In this paper, we examine whether expanding this alignment coverage improves learning. Across three heterogeneous teacher--student pairs on mathematical reasoning and c

发布时间：—

抓取时间：2026-10-07 15:17

来源机构：Hugging Face