Evaluating SAE interpretability without explanations

Gonçalo Paulo, Nora Belrose

Sparse autoencoders (SAEs) and transcoders have become important tools for machine learning interpretability. However, measuring how interpretable they are remains challenging, wit…