Causal tracing of factual recall — where is the fact stored?
The field-standard causal trace (Meng et al., ROME), run across six models (ROME only did GPT-2/GPT-J). Corrupt the subject tokens with Gaussian noise (3× the embedding std) — the fact’s probability drops — then in the corrupted run restore the clean MLP output at each layer (at the subject’s last token) and measure how much the object’s probability recovers. The MLP whose restoration recovers the most is where the fact is enriched; ROME’s headline is an early-mid MLP site at the subject’s last token.
Two sites are expected: an early MLP store at the subject’s last token (restore the clean MLP output) and a late attention readout at the last token (restore the clean attention output — the heads that copy the enriched fact to the prediction).
| model | facts used | MLP store — peak @ subject (depth, recovery) | attention readout — peak @ last token (depth, recovery) |
|---|---|---|---|
| gpt2 | 17 | L0 (0.00, +75%) | L9 (0.82, +78%) |
| gpt2-medium | 17 | L0 (0.00, +97%) | L20 (0.87, +31%) |
| gpt2-large | 15 | L1 (0.03, +104%) | L26 (0.74, +37%) |
| Llama-3.2-1B | 18 | L0 (0.00, +94%) | L9 (0.60, +55%) |
| Qwen2.5-1.5B | 18 | L0 (0.00, +87%) | L22 (0.81, +57%) |
Finding — the two-site flow is architecture-invariant. Every model traced shows the canonical ROME structure: an early MLP store at the subject (peak depth ≈ 0.00–0.03) feeding a late attention readout at the last token (peak depth ≈ 0.60–0.87). The fact is enriched into the subject’s residual by the early MLPs, then copied to the prediction by late-layer attention — the same early-MLP → late-attention information flow in GPT-2 (small/medium/large), Llama, and Qwen. Recovered cross-model (ROME only did GPT-2/GPT-J).
Scale note. Factual recall recovers from the early MLPs at the subject’s last token in every model — and the early-mid MLP plateau widens with scale: GPT-2-small is a sharp single L0 spike, while gpt2-large and Llama show a broad L0–3 early-mid plateau (the same embedding-block-widens-with-scale pattern as the extended-embedding test). This is the rigorous (corruption + restoration) confirmation of the cheaper ablation-contrast — facts are enriched in the early MLPs at the subject, ROME’s store, now cross-model. Excluded: gemma-2-2b — Gemma scales its embeddings by √d, so the standard 3×-std noise barely corrupts the fact (denom < 0.2) and no clean trace is obtained.
Recovery = fraction of the corruption-induced probability drop that restoring that layer’s MLP output (at the subject’s last token) recovers. Provisional, ~18 capital-city facts, single-token objects, one noise sample per fact. Data: causal_tracing_summary.json. Regenerate: causal_tracing.py.