saey-trigram-features-implemented-in-mlp
IN premise — summaries/2026/08/24/bricken-2023-monosemanticity-chunk-5.md
Created 2026-08-25T02:57:55+00:00
Trigram features (predicting a specific token given two preceding tokens, e.g., '19' in 'COVID-19') are implemented via MLP layers even though attention heads could in principle perform this function.
Summary
Trigram prediction, like guessing a specific token from the two tokens immediately before it, is carried out by the MLP layers rather than by attention heads, even though attention is technically capable of doing the same job. This reveals a concrete division of labor in the model: local, repetitive n-gram patterns are offloaded to the feed-forward pathway, which implies that attention heads are being reserved for other kinds of relationships rather than redundant local prediction.