Katarzyna Filus

TMS reveals that traditional metrics for assessing monosemanticity in Sparse Autoencoders are vulnerable to encoder geometry, highlighting a more robust, label-free alternative.