Georg Lange

Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control

Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching