Unsupervised Features Mining via Activation Geometry

Amit LeVi, Elad David, Max Fomin

Interpretability methods aim to reveal the features represented inside large language models (LLMs). Many existing methods begin with labeled examples of a human-defined concept th…