1 paper
Wilson E. Marcílio-Jr, Danilo M. Eler
Sparse autoencoders (SAEs) trained on large language model activations output thousands of features that enable mapping to human-interpretable concepts. The current practice for an…