SOTA Harmfulness Detection on Full in-distribution (test) and PapersWithCode

0.964AUROC

MultiLayer-Linear

Updated 13d ago

Evaluation Results

Method
MultiLayer-Linear 2026.05	0.964	-
Geometry-Lite 2026.05	0.964	-
Geometry-Lite 2026.05	0.961	-
Geometry-Lite 2026.05	0.96	-
MultiLayer-Linear 2026.05	0.959	-
MultiLayer-Linear 2026.05	0.959	-
MultiLayer-Linear 2026.05	0.959	-
Best single-layer probe 2026.05	0.958	-
MultiLayer-Linear 2026.05	0.956	-
MultiLayer-Linear 2026.05	0.956	-
Geometry-Lite 2026.05	0.956	-
Geometry-Lite 2026.05	0.956	-
Geometry-Lite 2026.05	0.956	-
Geometry-Lite 2026.05	0.954	-
Final-layer probe 2026.05	0.953	-
Best single-layer probe 2026.05	0.951	-
MultiLayer-Linear 2026.05	0.951	-
Best single-layer probe 2026.05	0.95	-
TaT-Disp-LSTM 2026.05	0.95	-
TaT-Disp-LSTM 2026.05	0.95	-
Refusal-style probe 2026.05	0.949	-
TaT-Disp-LSTM 2026.05	0.948	-
TaT-Disp-LSTM 2026.05	0.948	-
TaT-Disp-LSTM 2026.05	0.948	-
Geometry-Lite 2026.05	0.948	-
Final-layer probe 2026.05	0.947	-
Best single-layer probe 2026.05	0.947	-
Best single-layer probe 2026.05	0.947	-
TaT-Disp-LSTM 2026.05	0.947	-
TaT-Disp-LSTM 2026.05	0.945	-
MultiLayer-Linear 2026.05	0.945	-
Geometry-Lite 2026.05	0.939	-
Best single-layer probe 2026.05	0.938	-
Best single-layer probe 2026.05	0.938	-
TaT-Disp-LSTM 2026.05	0.937	-
Final-layer probe 2026.05	0.936	-
Final-layer probe 2026.05	0.935	-
MultiLayer-Linear 2026.05	0.935	-
Final-layer probe 2026.05	0.934	-
Refusal-style probe 2026.05	0.934	-
Final-layer probe 2026.05	0.933	-
Final-layer probe 2026.05	0.933	-
TaT-Disp-LSTM 2026.05	0.932	-
MultiLayer-DIM 2026.05	0.93	-
MultiLayer-DIM 2026.05	0.928	-
MultiLayer-DIM 2026.05	0.927	-
Best single-layer probe 2026.05	0.926	-
MultiLayer-DIM 2026.05	0.926	-
Refusal-style probe 2026.05	0.925	-
Final-layer probe 2026.05	0.921	-
MultiLayer-DIM 2026.05	0.915	-
Best single-layer probe 2026.05	0.911	-
Final-layer probe 2026.05	0.908	-
MultiLayer-DIM 2026.05	0.908	-
MultiLayer-DIM 2026.05	0.905	-
Refusal-style probe 2026.05	0.901	-
Refusal-style probe 2026.05	0.9	-
Refusal-style probe 2026.05	0.898	-
Refusal-style probe 2026.05	0.898	-
Refusal-style probe 2026.05	0.897	-
MultiLayer-DIM 2026.05	0.888	-
MultiLayer-DIM 2026.05	0.874	-
Refusal-style probe 2026.05	0.863	-
Final-layer probe 2026.05	-	0.933
Best single-layer probe 2026.05	-	0.941
Refusal-style probe 2026.05	-	0.907
MultiLayer-DIM 2026.05	-	0.911
TaT-Disp-LSTM 2026.05	-	0.945
MultiLayer-Linear 2026.05	-	0.954
Geometry-Lite 2026.05	-	0.955