rome-attnedit-layer-33-fails-generalization
IN premise — summaries/2026/08/24/meng-2022-rome-sR-references-chunk-2.md
Created 2026-08-25T02:58:17+00:00
AttnEdit (constrained fine-tuning of W_Q, W_K, W_V at layer 33) succeeds on direct prompts but fails paraphrase generalization, supporting the hypothesis that middle-layer MLPs store factual recall while late-layer attention handles word prediction.
Summary
Editing the attention weights in the model's late layers only changes how it predicts the next word for a specific phrasing, not what it actually "knows." This means that if you want a model to genuinely update a fact rather than just rephrase its answer to one prompt, the real work has to happen in the middle-layer feed-forward circuits, not in late-layer attention.