c2023-rome-rank-one-weight-edit

IN premise — summaries/2026/08/24/cohen-2023-ripple-effects-sR-references.md

Created 2026-08-25T02:57:57+00:00

ROME performs knowledge editing via rank-one matrix (weight) updates to the model

Summary

ROME corrects a model's stored fact by adjusting only a tiny, focused slice of one weight matrix rather than retraining the model. This matters because it keeps the edit cheap and surgical, so you can fix one piece of knowledge without accidentally breaking all the other things the model already knows.