Rank-One Model Editing (ROME) changes one fact inside a model. It finds the MLP layer that stores the fact. Then it adds one rank-one matrix to that layer's weights. All numbers here are illustrative.