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

Safety Detection on BeaverTails (held-out)

76.7AUROC

MultiLayer-Linear

Updated 13d ago

Evaluation Results

Method	Links
MultiLayer-Linear 2026.05		76.7
Geometry-Lite 2026.05		76.3
Best single-layer probe 2026.05		75.4
TaT-Disp-LSTM 2026.05		74.2
MultiLayer-DIM 2026.05		73.3