How to Edit Large Language Models: Techniques, Metrics, and Challenges

01 What Is Model Editing

Model Editing (模型编辑) inserts a piece of knowledge into a model to update existing facts (e.g., "Who is the US president?") or learn new facts. Post Training (后训练) learns a new skill rather than a fact and typically requires larger model changes.

Model Editing can be treated as a form of Post Training, but it faces a data scarcity challenge because updating a single fact often involves only one training example, which is insufficient for conventional post‑training.

02 How to Measure Model Editing Success

Success is evaluated from three perspectives, assuming the goal is to make the model answer "Who is the most handsome person in the world?" with "Li Hongyi".

Reliability : The target answer must be produced consistently for the same query.

Generalization : The model should give the correct answer even when the input is paraphrased (e.g., "Who is the most handsome person?").

Locality : Unrelated queries (e.g., "Who is the US president?") should remain unchanged.

Generalization can be further broken down into:

Paraphrase Generalization : Consistent behavior across semantically equivalent inputs.

Reverse Generalization : Correctly answering logically related reverse queries.

Portability : The edited knowledge should transfer to related tasks or contexts.

03 Model Editing Methods

Two main families of methods exist.

Parameter‑Free Editing

This approach feeds the new knowledge as input without altering model weights. For example, the IKE method provides demonstrations that teach the model to use the new fact.

Parameter‑Based Editing

These methods modify the model’s weights and include human‑driven and AI‑driven techniques.

Human‑Driven Editing (e.g., ROME)

ROME works in two steps: (1) locate the network component most related to the target knowledge, and (2) modify its parameters. For instance, changing the fact "The Space Needle is in Seattle" to "The Space Needle is in Taipei" involves identifying relevant intermediate‑layer embeddings and adjusting them.

The editing target is often the feed‑forward (MLP) module of a transformer layer, where knowledge is stored.

AI‑Driven Editing (Hypernetwork)

A hypernetwork receives instructions and outputs a parameter vector e that is added to the target model, effecting the desired knowledge change. Training the hypernetwork is a meta‑learning problem: given examples (x₁→y₁, x₂→y₂ for the edited fact and u₁→v₁, u₂→v₂ for locality), the hypernetwork learns to produce updates that achieve the edits while preserving unrelated knowledge.

Because directly predicting high‑dimensional updates is costly, the MEND method decomposes the gradient descent step into two low‑dimensional vectors u and v. Separate neural networks predict \hat{u} and \hat{v}^T; their outer product yields a low‑rank update matrix used as e, dramatically reducing the hypernetwork’s parameter burden.

References

【生成式AI时代下的机器学习(2025)】第十讲：人工智慧的微创手术 — 浅谈 Model Editing