memit-multi-layer-update-layers-5-9

Summary

MEMIT rewrites the internal weights of a language model across five consecutive middle layers (5 through 9) at once, rather than tweaking a single layer. This means the edit is spread deeper into the network, making it harder for the surrounding layers to simply override the change, but it also means the intervention touches more of the model's structure than a shallow, one-layer patch would.