KronSAE uses Kronecker Factorization for efficient Sparse Autoencoders

KronSAE introduces a compositional co-activation prior by factorizing the latent space into heads and using mAND for pairwise compositions, improving interpretability and reducing computational costs compared to flat SAEs.