ripple-edits-rome-mechanism
IN premise — summaries/2026/08/24/cohen-2023-ripple-effects-s5-experiments.md
Created 2026-08-25T02:57:57+00:00
ROME performs rank-one updates to Transformer MLP layer weights to modify specific factual associations in a language model.
Summary
ROME can surgically change a single fact a language model "knows" by making one tiny, mathematically minimal adjustment to the weights inside one of its feed-forward layers, rather than retraining the model. This means the system can correct or swap a specific association on the fly with very low computational cost and without disturbing the thousands of other things the model already handles well.