Harmfulness Detection on Full in-distribution (test)
0.964AUROCMultiLayer-Linear
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MultiLayer-LinearBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.964 | — | |
| Geometry-LiteBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.964 | — | |
| Geometry-LiteBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.961 | — | |
| Geometry-LiteBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.96 | — | |
| MultiLayer-LinearBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.959 | — | |
| MultiLayer-LinearBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.959 | — | |
| MultiLayer-LinearBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.959 | — | |
| Best single-layer probeBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.958 | — | |
| MultiLayer-LinearBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.956 | — | |
| MultiLayer-LinearBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.956 | — | |
| Geometry-LiteBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.956 | — | |
| Geometry-LiteBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.956 | — | |
| Geometry-LiteBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.956 | — | |
| Geometry-LiteBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.954 | — | |
| Final-layer probeBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.953 | — | |
| Best single-layer probeBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.951 | — | |
| MultiLayer-LinearBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.951 | — | |
| Best single-layer probeBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.95 | — | |
| TaT-Disp-LSTMBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.95 | — | |
| TaT-Disp-LSTMBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.95 | — | |
| Refusal-style probeBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.949 | — | |
| TaT-Disp-LSTMBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.948 | — | |
| TaT-Disp-LSTMBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.948 | — | |
| TaT-Disp-LSTMBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.948 | — | |
| Geometry-LiteBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.948 | — | |
| Final-layer probeBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.947 | — | |
| Best single-layer probeBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.947 | — | |
| Best single-layer probeBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.947 | — | |
| TaT-Disp-LSTMBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.947 | — | |
| TaT-Disp-LSTMBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.945 | — | |
| MultiLayer-LinearBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.945 | — | |
| Geometry-LiteBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.939 | — | |
| Best single-layer probeBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.938 | — | |
| Best single-layer probeBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.938 | — | |
| TaT-Disp-LSTMBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.937 | — | |
| Final-layer probeBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.936 | — | |
| Final-layer probeBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.935 | — | |
| MultiLayer-LinearBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.935 | — | |
| Final-layer probeBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.934 | — | |
| Refusal-style probeBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.934 | — | |
| Final-layer probeBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.933 | — | |
| Final-layer probeBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.933 | — | |
| TaT-Disp-LSTMBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.932 | — | |
| MultiLayer-DIMBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.93 | — | |
| MultiLayer-DIMBackbone Family=Llama-3, Backbone Scale=70B2026.05 | 0.928 | — | |
| MultiLayer-DIMBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.927 | — | |
| Best single-layer probeBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.926 | — | |
| MultiLayer-DIMBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.926 | — | |
| Refusal-style probeBackbone Family=Gemma-2, Backbone Scale=27B2026.05 | 0.925 | — | |
| Final-layer probeBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.921 | — | |
| MultiLayer-DIMBackbone Family=Llama-3, Backbone Scale=8B2026.05 | 0.915 | — | |
| Best single-layer probeBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.911 | — | |
| Final-layer probeBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.908 | — | |
| MultiLayer-DIMBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.908 | — | |
| MultiLayer-DIMBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.905 | — | |
| Refusal-style probeBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.901 | — | |
| Refusal-style probeBackbone Family=Qwen3, Backbone Scale=14B2026.05 | 0.9 | — | |
| Refusal-style probeBackbone Family=Gemma-2, Backbone Scale=9B2026.05 | 0.898 | — | |
| Refusal-style probeBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.898 | — | |
| Refusal-style probeBackbone Family=Qwen3, Backbone Scale=8B2026.05 | 0.897 | — | |
| MultiLayer-DIMBackbone Family=Llama-3, Backbone Scale=1B2026.05 | 0.888 | — | |
| MultiLayer-DIMBackbone Family=Qwen3, Backbone Scale=1.7B2026.05 | 0.874 | — | |
| Refusal-style probeBackbone Family=Gemma-2, Backbone Scale=2B2026.05 | 0.863 | — | |
| Best single-layer probeBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.941 | |
| Final-layer probeBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.933 | |
| Geometry-LiteBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.955 | |
| MultiLayer-DIMBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.911 | |
| MultiLayer-LinearBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.954 | |
| Refusal-style probeBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.907 | |
| TaT-Disp-LSTMBackbone Family=All, Backbone Scale=Average2026.05 | — | 0.945 |