---
id: 20260417-T0-04
title: "LLM行为定位新突破：权重补丁实现源级机制定位"
title_en: "Weight Patching Enables Source-Level Mechanistic Localization in LLMs"
url: https://ai.daily.yangsir.net/daily/20260417-T0-04
issue_date: 2026-04-17
publish_date: 2026-04-16T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2604.13694
---

# LLM行为定位新突破：权重补丁实现源级机制定位

新研究提出权重补丁方法，首次实现LLM的源级机制定位。传统激活空间定位无法区分相关性与因果性，该方法通过权重层面的干预，准确定位控制模型行为的关键组件。为神经网络内部可解释性研究开辟新路径。

## English Version

**Weight Patching Enables Source-Level Mechanistic Localization in LLMs**

New research introduces weight patching to achieve source-level mechanistic localization in LLMs. Unlike activation-based methods, this approach identifies causally important components by intervening at the weight level, advancing neural network interpretability research.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2604.13694)

**详情页**：https://ai.daily.yangsir.net/daily/20260417-T0-04

---

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