---
id: 20261024-T0-06
title: "小模型看图老出错？新方法用最小编辑做多模态蒸馏"
title_en: "Minimal-Edit On-Policy Distillation Fixes Small VLMs' Diagram Errors"
url: https://ai.daily.yangsir.net/daily/20261024-T0-06
issue_date: 2026-10-24
publish_date: 2026-10-23T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.30231
---

# 小模型看图老出错？新方法用最小编辑做多模态蒸馏

arXiv 新论文指出，视觉数学推理可能在推理开始前就失败：视觉语言模型读错图表后，后续推理都建立在错误前提上，小模型尤其明显。论文提出最小编辑的在线策略蒸馏方法，核心思路是“编辑而非模仿”，让学生模型在教师模型输出基础上做尽量小的修正，而不是整体照搬。实际影响：小模型在图表类数学题上的视觉读取准确率有望提升，降低对超大模型的依赖。

## English Version

**Minimal-Edit On-Policy Distillation Fixes Small VLMs' Diagram Errors**

A new arXiv paper notes that visual math reasoning can fail before reasoning even starts: when a vision-language model misreads a diagram, all subsequent reasoning builds on false premises, and smaller models are hit hardest. The paper proposes minimal-edit on-policy distillation, whose core idea is to edit rather than imitate — the student model makes the smallest possible corrections to teacher outputs instead of copying them wholesale. Practical impact: better diagram-reading accuracy for small models on visual math problems, reducing reliance on much larger models.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.30231)

**详情页**：https://ai.daily.yangsir.net/daily/20261024-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*