Jul 2026
Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence
The Tversky Monosemanticity Score (TMS) is introduced, a label-free metric that operationalizes monosemanticity as activation-set coherence of binarized SAE latents, and does not require external embedding encoders.
Katarzyna Filus, Sebastian Pokuciński
· arXiv.org · 0 citations