Singular Vectors of Attention Heads Align with Features

Gabriel Franco, Carson Loughridge, Mark Crovella

Identifying feature representations in language models is a central task in mechanistic interpretability. Several recent studies have made the observation that feature representati…