ResearchOct 27, 2026, 00:00 UTC

Research proposes redistribution sensitivity score to find recurring changes in sparse autoencoder features

The RSS method identifies recurring feature responses under prompt transformations and separates them from downstream failure. In two frozen safety readouts, RSS-based support erosion with the full clean margin gave recurrent failure AUROCs of 0.935 and 0.958.

Topicresearch
Topicsafety

1outlet ran itRead at Hugging Face Papers

Status: Research
Published: Oct 27, 2026, 00:00 UTC That day
Outlets: 1
Written by: Claude How

Every outlet that ran it 1 article

Oct 27, 2026, 00:00 UTC
Hugging Face PapersResearchLead
Representation Redistribution in Sparse Autoencoders

All news