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

Harmfulness detection on BeaverTails, ToxicChat, PKU-SafeRLHF Hard (test)

90.2AUROC

MultiLayer-Linear

Updated 13d ago

Evaluation Results

Method	Links
MultiLayer-Linear 2026.05		90.2
Geometry-Lite 2026.05		89.8
Geometry-Lite 2026.05		89.8
MultiLayer-Linear 2026.05		89.5
MultiLayer-Linear 2026.05		89.5
Geometry-Lite 2026.05		89.4
MultiLayer-Linear 2026.05		89.1
Best single-layer probe 2026.05		89
Final-layer probe 2026.05		88.9
Final-layer probe 2026.05		88.8
Best single-layer probe 2026.05		88.8
MultiLayer-Linear 2026.05		88.8
Geometry-Lite 2026.05		88.8
TaT-Disp-LSTM 2026.05		88.7
MultiLayer-Linear 2026.05		88.7
Geometry-Lite 2026.05		88.6
MultiLayer-Linear 2026.05		88.4
Geometry-Lite 2026.05		88.2
Refusal-style probe 2026.05		87.9
Best single-layer probe 2026.05		87.8
MultiLayer-DIM 2026.05		87.8
TaT-Disp-LSTM 2026.05		87.8
Geometry-Lite 2026.05		87.8
Best single-layer probe 2026.05		87.7
TaT-Disp-LSTM 2026.05		87.6
TaT-Disp-LSTM 2026.05		87.6
Best single-layer probe 2026.05		87.5
Geometry-Lite 2026.05		87.5
Best single-layer probe 2026.05		87.4
TaT-Disp-LSTM 2026.05		87.4
TaT-Disp-LSTM 2026.05		87.4
TaT-Disp-LSTM 2026.05		87.3
MultiLayer-Linear 2026.05		87.2
MultiLayer-DIM 2026.05		87.1
MultiLayer-DIM 2026.05		87
MultiLayer-DIM 2026.05		86.9
Final-layer probe 2026.05		86.2
TaT-Disp-LSTM 2026.05		86.2
Final-layer probe 2026.05		86.1
Best single-layer probe 2026.05		86
MultiLayer-DIM 2026.05		85.8
MultiLayer-DIM 2026.05		85.8
MultiLayer-DIM 2026.05		85.8
Final-layer probe 2026.05		85.7
MultiLayer-Linear 2026.05		85.6
Final-layer probe 2026.05		85.4
Final-layer probe 2026.05		85.2
Geometry-Lite 2026.05		85.1
Final-layer probe 2026.05		84.8
Best single-layer probe 2026.05		84.8
Refusal-style probe 2026.05		84.7
Refusal-style probe 2026.05		84.5
TaT-Disp-LSTM 2026.05		83.9
MultiLayer-DIM 2026.05		82.8
Best single-layer probe 2026.05		82.5
Refusal-style probe 2026.05		82.4
Refusal-style probe 2026.05		82.3
Refusal-style probe 2026.05		82.1
Refusal-style probe 2026.05		81.7
Final-layer probe 2026.05		81.6
MultiLayer-DIM 2026.05		79.8
Refusal-style probe 2026.05		79.7
Refusal-style probe 2026.05		79.4