rmovva/HypotheSAEs

[ICML 2025] HypotheSAEs: Hypothesizing interpretable relationships in text datasets using sparse autoencoders. https://arxiv.org/abs/2502.04382