liveBook · Manning

Knowledge editing

As explained in LLMs in Production

Knowledge editing is the technique of modifying specific factual or behavioral associations in a trained model without retraining the entire model.

When a model gives an outdated fact, retraining or fine-tuning can be excessive. Knowledge editing aims to surgically adjust a specific behavior or piece of information while avoiding broader, unexpected changes. section 12.2.5

Knowledge editing is used to:

  • Combat factual decay, such as outdated information about a current Super Bowl winner or country president. section 12.2.5
  • Insert, update, or erase information by targeting the model weights associated with incorrect responses. section 12.2.5
  • Explore specialized techniques including ROME, MEND, and GRACE rather than applying a heavier full-model update. section 12.2.5

Updating one stale fact

Suppose a model still answers with a former Super Bowl winner. Knowledge editing seeks to change that association directly, instead of retraining or fine-tuning the model and risking unrelated changes. section 12.2.5

How LLMs in Production covers it

How deep it goes
Defines the surgical goal and names techniques, but defers algorithmic details and implementation guidance.
Examples
Uses changing Super Bowl winners and country presidents as examples of facts that can become outdated.
What you can do after
Evaluate knowledge-editing approaches and begin with ROME, MEND, or GRACE through EasyEdit.
How it connects
Connects hallucinations and factual decay with retraining, fine-tuning, model weights, and knowledge graphs.
Cover of LLMs in Production
liveBookmanning.com
Read LLMs in Production on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLMs in Production. The section links point to the parts of the book it draws on.