Contribution
A localization method (causal tracing), an editing method (rank-one update), and a benchmark (CounterFact) that measures generalization and specificity together.
Key insight
If an MLP is a linear associative memory, one fact is one key-value pair, and inserting it is a closed-form rank-one update that minimally disturbs other keys.
Key results
- CounterFact Score: ROME 89.2 (GPT-2 XL) and 91.5 (GPT-J). The best baseline is 66.9 on GPT-2 XL (FT+L) and 68.7 on GPT-J (FT+L)
- Fine-tuning reaches full efficacy but destroys neighbors on GPT-J (NS 10.3, against 83.0 unedited)
- The edit works best at the last subject token in middle layers (Figure 5)
Limitations
- One fact per edit, and directional, as §3.7 says
- Neighborhood success falls below the unedited model’s on both models
- The edit layer comes from tracing averaged over facts. Whether the average tells you anything about a given fact is the question Hase et al. answer: it doesn’t (LIT-574)