Locating and Editing Factual Associations in GPT

Summary

Using causal interventions, the authors find that factual associations in a GPT model correspond to localized computations in middle-layer feed-forward modules, and show that a single association can be rewritten by directly editing those weights (Rank-One Model Editing).

CommentaryShows that a model’s knowledge does live in its weights, in specific places — which is exactly why memory held there is bound to that model.