Ziye Mei
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Representation Redistribution in Sparse Autoencoders
According to the abstract excerpt, the authors study how rewording a prompt shifts activity across sparse autoencoder features. They introduce a redistribution sensitivity score and report that, on two frozen safety readouts, it predicts recurring failures with AUROCs of 0.935 and 0.958 using about 20 features.