Source
Meng et al. (2022), LIT-578, read as NOTE-315, for the account. Hase et al. (2023), LIT-574, read as NOTE-313, for why it is filed already retired (DP-003).
The account
Localize, then edit. Causal tracing finds the hidden states whose clean values restore a corrupted fact, and those states concentrate in mid-layer MLP outputs at the last subject token. If that is where the association lives, that is where a rank-one or batched update should go. ROME chose its layer from tracing averaged over facts, and MEMIT chose its layer range from the same analysis.
Why it is rejected
- Per fact, tracing and edit success are unrelated. ρ = −0.13 between a fact’s tracing effect at the edit layer and its rewrite score
- The layer, not the localization, predicts success. R² 94.7% from the layer alone, and 94.8% with tracing added
- Edits succeed where tracing does not point. Many facts trace to layers 1–3 or 16–20, and editing them at layer 6 still works
- It is robust: MEMIT and fine-tuning show the same, and so do other metrics, GPT-2 XL and zsRE
What survives
The editing methods (MEMIT is SOTA-347), and causal tracing as a description of where information flows. What does not survive is using one to choose the other. Pick the edit layers by measured editing performance.