Abstract
ROME and MEMIT are largely believed to be two different model editingalgorithms, with the major difference between them being the ability to performbatched edits. In this paper, we unify these two algorithms under a singleconceptual umbrella, optimizing for the same goal, which we call thepreservation-memorization objective. ROME uses an equality constraint tooptimize this objective to perform one edit at a time, whereas MEMIT employs amore flexible least-square constraint that allows for batched edits. Wegeneralize ROME and enable batched editing with equality constraint in the formof EMMET - an Equality-constrained Mass Model Editing algorithm forTransformers, a new batched memory-editing algorithm. EMMET can performbatched-edits up to a batch-size of 10,000, with very similar performance toMEMIT across multiple dimensions. With the introduction of EMMET, we trulyunify ROME and MEMIT and show that both algorithms are equivalent in terms oftheir optimization objective, their abilities (singular and batched editing),their model editing performance and their limitations.