Share your thoughts, 1 month free Claude Pro on usSee more

Safety Detection on PKU-SafeRLHF (held-out)

90.6AUROC

MultiLayer-DIM

Updated 13d ago

Evaluation Results

Method	Links
MultiLayer-DIM 2026.05		90.6
Geometry-Lite 2026.05		90.3
TaT-Disp-LSTM 2026.05		90.2
MultiLayer-Linear 2026.05		90
Best single-layer probe 2026.05		87