跳到主要内容Skip to main content

模型编辑Model editing

Model editing

  • 核心概念Core concepts
  • 第四站 · 看懂Station 4 · Under the Hood
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

不改全量权重,只精准改掉模型里某条事实或某个行为。Change one fact or behaviour in a model without retraining all weights.

这是什么What it is

先定位存储该事实的层与模块(例如 MLP 的某几个键值)。

再对该位置的权重做局部更新,让模型回答新事实。

理想情况是只改目标事实,不影响其他知识。

Locate the layer and module that stores the fact, such as a few MLP key-value slots.

Apply a local weight update so the model answers with the new fact.

The ideal edit changes only the target fact and leaves other knowledge intact.

为什么需要知道Why it matters

模型知识过期时,全量重训太贵。模型编辑提供「热修复」的可能,但容易连带改坏相邻知识。Retraining for stale knowledge is expensive. Editing offers a hot fix, but can damage neighbouring knowledge.

出处Sources
  • Locating and Editing Factual Associations in GPT(arXiv)Locating and Editing Factual Associations in GPT (arXiv 2202.05262) ↗