memit-vicuna-layers-5-to-9

Summary

MEMIT spreads its knowledge edit across five consecutive mid-to-late layers of the model, while ROME targets a single late layer, meaning the two techniques intervene at very different points in the network's reasoning chain. This matters because spreading edits over multiple layers (MEMIT) risks more collateral changes to other knowledge, whereas concentrating them in one layer (ROME) is more surgical but may limit how well the edit integrates with the model's broader representations.