Research question
Where are facts stored in a transformer and can they be edited directly?
Key findings
- Causal tracing localises factual recall to mid-layer MLPs - ROME edits a fact by a rank-one update to a single layer
Locating and Editing Factual Associations in GPT
Causal tracing and ROME, the model-editing technique later used in PoisonGPT and studied as a backdoor vector.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.