sae-families

This paper investigates whether single-token sparse autoencoder features are causally necessary for model outputs, finding that causal roles vary across SAE families and are sensitive to training methodology rather than activation function or scale.