---
id: 20261031-T0-19
title: "一次权重更新能换来什么：训练各阶段的杠杆效应"
title_en: "What One Weight Update Buys: Leverage Effects Across Training Stages"
url: https://ai.daily.yangsir.net/daily/20261031-T0-19
issue_date: 2026-10-31
publish_date: 2026-10-30T04:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.37719
---

# 一次权重更新能换来什么：训练各阶段的杠杆效应

arXiv:2610.37719v1 公告类型：新论文。摘要：后训练会将语言模型的权重移动约其范数的千分之一，却能显著改变模型行为；预训练将权重移动约三分之一，却只带来缓慢漂移。我们提出“杠杆”（leverage）这一精确度量，用于衡量单位权重变化能换来什么。

## English Version

**What One Weight Update Buys: Leverage Effects Across Training Stages**

arXiv:2610.37719v1 (announcement type: new paper). Post-training moves a language model's weights by about one-thousandth of their norm yet markedly changes model behavior, while pre-training moves weights by about one-third but produces only slow drift. The authors introduce "leverage," a precise metric for measuring what a unit of weight change buys.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.37719)

**详情页**：https://ai.daily.yangsir.net/daily/20261031-T0-19

---

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