rome-factual-knowledge-location
IN premise — summaries/2026/08/24/meng-2022-rome-s0-abstract.md
Created 2026-08-25T02:58:14+00:00
In GPT-2 XL, factual associations are stored in mid-layer feed-forward MLP modules, not in attention layers, as identified by causal mediation analysis.
Summary
In GPT-2 XL, the specific facts the model remembers live in the feed-forward calculation blocks in the middle layers, while the attention layers handle routing and combining rather than storage. This gives a clear target for anyone wanting to probe, audit, or edit what the model has memorized, since it tells you exactly which components carry the knowledge and which ones do not.