Redistribution Sensitivity Score for Sparse Autoencoders
October 14, 2026
The Redistribution Sensitivity Score (RSS) identifies how prompt transformations redistribute activations across sparse autoencoder (SAE) features. Evaluating 27 SAEs across eight language models, the study finds that prompt changes produce reproducible redistribution profiles even on excluded variants.
HOW THIS AFFECTS YOU
●
researcherThis provides a metric to better understand how feature activation shifts during prompt transformations.