mean squared error

New methods enhance sparse autoencoder interpretability and stability

Researchers have developed new methods to address limitations in sparse autoencoders (SAEs), which are used to interpret the internal representations of large language models. One paper introduces adaptive elastic net S…