memit-rome-sequential-degradation

IN premise — summaries/2026/08/24/meng-2022-memit-sR-references.md

Created 2026-08-25T02:58:14+00:00

ROME processes edits sequentially (~4 sec/edit) and degrades beyond approximately 32 edits, while MEMIT scales to 10,000+ edits via parallel z_i computation and a shared matrix inverse.

Summary

ROME can only handle a small batch of knowledge edits because it works one edit at a time and starts producing worse results after around thirty-two changes. MEMIT removes that ceiling by computing edits in parallel and reusing shared math, so a system built on it can apply thousands of corrections without the sequential bottleneck that caps ROME.