New SAE paper: Redistribution Sensitivity Score predicts readout failures with AUROC up to 0.958
OpenSAE · hf · 2026-10-28
A paper by Ziye Mei and Haotian Ma (UW–Madison) tackles feature redistribution in sparse autoencoders: prompt transformations preserving the underlying request can redistribute activation across SAE features, and observed changes need not recur or affect downstream computation.
- Introduces the Redistribution Sensitivity Score (RSS), identifying large feature responses recurring in a consistent direction, validated on held-out variants
- Across 27 pretrained SAEs spanning 8 language models, transformations yield reproducible redistribution profiles
- Separates recurring representation change from signed readout contribution and decision failure via an exact decomposition for fixed linear readouts
- On two frozen safety readouts, combining RSS support erosion with clean margin achieves recurrent-failure AUROCs of 0.935 and 0.958 using only 20/10 features
More from Research
- UCLA and Tsinghua paper probes how far coding agents can go in visual generation — ucla-cmllab · 2026-10-28
- Author credits Claude and ChatGPT for writing most of his Stanford medical AI thread — jonc101x · 2026-10-28
- What would it take to trust an AI agent to carry out clinical work tasks? — jonc101x · 2026-10-28
- Medical AI is weakest at clinical decision support and admin workflows — exactly where hospitals need it most — jonc101x · 2026-10-28
- GPT-5.5 passes 46.3% of clinical tasks once, but only 28% across three runs — jonc101x · 2026-10-28
- Stanford's 3-year medical AI benchmark review: GPT-5.5 drops to 28% when tasks must succeed 3 runs in a row — jonc101x · 2026-10-28