rome-aie-mlp-vs-attention

IN premise — summaries/2026/08/24/meng-2022-rome-s2-interventions-on-activations-for-tracing-information-flow.md

Created 2026-08-25T02:58:15+00:00

Causal Tracing measures the Average Indirect Effect (AIE) of MLP at the last subject token as approximately 6.6%, compared to approximately 1.6% for attention at the same position.

Summary

When you measure how much each part of a transformer actually shapes the output at the subject-word position, the feed-forward layer is doing roughly four times more indirect work than the attention layer. This matters because it suggests the model's behavior at the most critical position is driven more by its internal computation than by the information-routing mechanism that is typically considered the headline feature of transformers.