ICA Lens: Interpreting Language Models Without Training Another Dictionary

Sida Liu, Feijiang Han

Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior. Sparse autoencoders (SAEs) have become the standard…