Gao et al. (2024)

New research reveals that sparse autoencoders (SAEs), often used to interpret neural networks, may produce up to 77% functionally inert features in degraded models and 9% in well-trained ones, raising…