Harmfulness detection on BeaverTails, ToxicChat, PKU-SafeRLHF Hard (test)
90.2AUROCMultiLayer-Linear
Evaluation Results
| Method | Links | |
|---|---|---|
| MultiLayer-LinearBackbone family=Llama, Model size=70B2026.05 | 90.2 | |
| Geometry-LiteBackbone family=Llama, Model size=8B2026.05 | 89.8 | |
| Geometry-LiteBackbone family=Llama, Model size=70B2026.05 | 89.8 | |
| MultiLayer-LinearBackbone family=Llama, Model size=8B2026.05 | 89.5 | |
| MultiLayer-LinearBackbone family=Gemma, Model size=27B2026.05 | 89.5 | |
| Geometry-LiteBackbone family=Qwen3, Model size=14B2026.05 | 89.4 | |
| MultiLayer-LinearBackbone family=Gemma, Model size=9B2026.05 | 89.1 | |
| Best single-layer probeBackbone family=Llama, Model size=70B2026.05 | 89 | |
| Final-layer probeBackbone family=Llama, Model size=70B2026.05 | 88.9 | |
| Final-layer probeBackbone family=Llama, Model size=8B2026.05 | 88.8 | |
| Best single-layer probeBackbone family=Gemma, Model size=27B2026.05 | 88.8 | |
| MultiLayer-LinearBackbone family=Qwen3, Model size=8B2026.05 | 88.8 | |
| Geometry-LiteBackbone family=Gemma, Model size=27B2026.05 | 88.8 | |
| TaT-Disp-LSTMBackbone family=Llama, Model size=8B2026.05 | 88.7 | |
| MultiLayer-LinearBackbone family=Gemma, Model size=2B2026.05 | 88.7 | |
| Geometry-LiteBackbone family=Gemma, Model size=9B2026.05 | 88.6 | |
| MultiLayer-LinearBackbone family=Qwen3, Model size=14B2026.05 | 88.4 | |
| Geometry-LiteBackbone family=Qwen3, Model size=8B2026.05 | 88.2 | |
| Refusal-style probeBackbone family=Llama, Model size=70B2026.05 | 87.9 | |
| Best single-layer probeBackbone family=Gemma, Model size=9B2026.05 | 87.8 | |
| MultiLayer-DIMBackbone family=Gemma, Model size=2B2026.05 | 87.8 | |
| TaT-Disp-LSTMBackbone family=Qwen3, Model size=14B2026.05 | 87.8 | |
| Geometry-LiteBackbone family=Gemma, Model size=2B2026.05 | 87.8 | |
| Best single-layer probeBackbone family=Gemma, Model size=2B2026.05 | 87.7 | |
| TaT-Disp-LSTMBackbone family=Llama, Model size=70B2026.05 | 87.6 | |
| TaT-Disp-LSTMBackbone family=Gemma, Model size=9B2026.05 | 87.6 | |
| Best single-layer probeBackbone family=Llama, Model size=8B2026.05 | 87.5 | |
| Geometry-LiteBackbone family=Llama, Model size=1B2026.05 | 87.5 | |
| Best single-layer probeBackbone family=Qwen3, Model size=14B2026.05 | 87.4 | |
| TaT-Disp-LSTMBackbone family=Gemma, Model size=2B2026.05 | 87.4 | |
| TaT-Disp-LSTMBackbone family=Gemma, Model size=27B2026.05 | 87.4 | |
| TaT-Disp-LSTMBackbone family=Qwen3, Model size=8B2026.05 | 87.3 | |
| MultiLayer-LinearBackbone family=Llama, Model size=1B2026.05 | 87.2 | |
| MultiLayer-DIMBackbone family=Gemma, Model size=9B2026.05 | 87.1 | |
| MultiLayer-DIMBackbone family=Gemma, Model size=27B2026.05 | 87 | |
| MultiLayer-DIMBackbone family=Qwen3, Model size=8B2026.05 | 86.9 | |
| Final-layer probeBackbone family=Gemma, Model size=27B2026.05 | 86.2 | |
| TaT-Disp-LSTMBackbone family=Llama, Model size=1B2026.05 | 86.2 | |
| Final-layer probeBackbone family=Qwen3, Model size=14B2026.05 | 86.1 | |
| Best single-layer probeBackbone family=Qwen3, Model size=8B2026.05 | 86 | |
| MultiLayer-DIMBackbone family=Llama, Model size=8B2026.05 | 85.8 | |
| MultiLayer-DIMBackbone family=Llama, Model size=70B2026.05 | 85.8 | |
| MultiLayer-DIMBackbone family=Qwen3, Model size=14B2026.05 | 85.8 | |
| Final-layer probeBackbone family=Gemma, Model size=9B2026.05 | 85.7 | |
| MultiLayer-LinearBackbone family=Qwen3, Model size=1.7B2026.05 | 85.6 | |
| Final-layer probeBackbone family=Qwen3, Model size=8B2026.05 | 85.4 | |
| Final-layer probeBackbone family=Llama, Model size=1B2026.05 | 85.2 | |
| Geometry-LiteBackbone family=Qwen3, Model size=1.7B2026.05 | 85.1 | |
| Final-layer probeBackbone family=Gemma, Model size=2B2026.05 | 84.8 | |
| Best single-layer probeBackbone family=Llama, Model size=1B2026.05 | 84.8 | |
| Refusal-style probeBackbone family=Llama, Model size=8B2026.05 | 84.7 | |
| Refusal-style probeBackbone family=Gemma, Model size=27B2026.05 | 84.5 | |
| TaT-Disp-LSTMBackbone family=Qwen3, Model size=1.7B2026.05 | 83.9 | |
| MultiLayer-DIMBackbone family=Llama, Model size=1B2026.05 | 82.8 | |
| Best single-layer probeBackbone family=Qwen3, Model size=1.7B2026.05 | 82.5 | |
| Refusal-style probeBackbone family=Qwen3, Model size=8B2026.05 | 82.4 | |
| Refusal-style probeBackbone family=Qwen3, Model size=14B2026.05 | 82.3 | |
| Refusal-style probeBackbone family=Gemma, Model size=9B2026.05 | 82.1 | |
| Refusal-style probeBackbone family=Llama, Model size=1B2026.05 | 81.7 | |
| Final-layer probeBackbone family=Qwen3, Model size=1.7B2026.05 | 81.6 | |
| MultiLayer-DIMBackbone family=Qwen3, Model size=1.7B2026.05 | 79.8 | |
| Refusal-style probeBackbone family=Gemma, Model size=2B2026.05 | 79.7 | |
| Refusal-style probeBackbone family=Qwen3, Model size=1.7B2026.05 | 79.4 |