Sebastian Pokuciński

Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence

Katarzyna Filus, Sebastian Pokuciński

Within Explainable Artificial Intelligence, mechanistic interpretability uses Sparse Autoencoders (SAEs) to extract more interpretable features from neural representations. However, assessing their monosemanticity, and thus explanation quality, remains challenging. Existing metri…Read more