Non-Gaussianity Reveals Interpretable Directions in Language Models
The study investigated whether non-Gaussianity in a language model's activation directions signals interpretable patterns. They found that projections along sparse autoencoder directions exhibit significantly more non-Gaussianity than random ones, suggesting this statistical signature can directly reveal useful structure. ICA Lens, a new workflow using Independent Component Analysis (ICA), successfully discovered these non-Gaussian directions across multiple models, yielding coherent lexical and semantic interpretations.