---
id: 20260726-T0-04
title: "Moir：让模型自己“规划”修改路径，减轻知识编辑对数学推理的破坏"
title_en: "Moir Lets LLMs Direct Their Own Edits, Reducing Knowledge Update Side Effects"
url: https://ai.daily.yangsir.net/daily/20260726-T0-04
issue_date: 2026-07-26
publish_date: 2026-07-25T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2607.20433
---

# Moir：让模型自己“规划”修改路径，减轻知识编辑对数学推理的破坏

知识编辑KEE会导致模型在数学和逻辑推理能力上的退化？新论文Moir提出了一种“轨迹规划式”知识编辑方法：让LLM通过自注意力机制选择最适合的修改路径，而非对所有层进行统一修改。实验表明，Moir在保持知识更新准确率的同时，将数学推理能力的退化幅度降低了60%以上。该方法试图在“学会新知识”和“不忘旧本领”之间找到更好的平衡点。

## English Version

**Moir Lets LLMs Direct Their Own Edits, Reducing Knowledge Update Side Effects**

Knowledge editing often degrades LLMs' math and logical reasoning. Moir, a new approach, lets the model choose its own editing path via self-attention rather than modifying all layers uniformly. Experiments show Moir reduces reasoning degradation by over 60% while maintaining update accuracy. It aims to find a better balance between learning new facts and preserving core capabilities.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2607.20433)

**详情页**：https://ai.daily.yangsir.net/daily/20260726-T0-04

---

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