saey-target-needs-relu-mlp-for-decomposition
IN premise — summaries/2026/08/24/bricken-2023-monosemanticity-chunk-2.md
Created 2026-08-25T02:57:54+00:00
A one-layer transformer with an attention layer plus a ReLU-activated MLP is identified as the simplest architecture that requires a feature decomposition strategy, unlike attention-only networks which can be analyzed without one.
Summary
The presence of a feed-forward (MLP) layer is what first forces researchers to break a transformer's behavior into separate, interpretable features, because attention-only networks are transparent enough to be analyzed directly. In practice, this means any model that includes a standard ReLU feed-forward block is at the complexity threshold where circuit-level or feature-level decomposition tools become necessary to understand what the model is actually doing.