List your company

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

90.2AUROCMultiLayer-Linear
78.96881.88484.887.716May 18, 2026

Evaluation results

MethodLinks
MultiLayer-LinearBackbone family=Llama, Model size=70B2026.0590.2
Geometry-LiteBackbone family=Llama, Model size=8B2026.0589.8
Geometry-LiteBackbone family=Llama, Model size=70B2026.0589.8
MultiLayer-LinearBackbone family=Llama, Model size=8B2026.0589.5
MultiLayer-LinearBackbone family=Gemma, Model size=27B2026.0589.5
Geometry-LiteBackbone family=Qwen3, Model size=14B2026.0589.4
MultiLayer-LinearBackbone family=Gemma, Model size=9B2026.0589.1
Best single-layer probeBackbone family=Llama, Model size=70B2026.0589
Final-layer probeBackbone family=Llama, Model size=70B2026.0588.9
Final-layer probeBackbone family=Llama, Model size=8B2026.0588.8
Best single-layer probeBackbone family=Gemma, Model size=27B2026.0588.8
MultiLayer-LinearBackbone family=Qwen3, Model size=8B2026.0588.8
Geometry-LiteBackbone family=Gemma, Model size=27B2026.0588.8
TaT-Disp-LSTMBackbone family=Llama, Model size=8B2026.0588.7
MultiLayer-LinearBackbone family=Gemma, Model size=2B2026.0588.7
Geometry-LiteBackbone family=Gemma, Model size=9B2026.0588.6
MultiLayer-LinearBackbone family=Qwen3, Model size=14B2026.0588.4
Geometry-LiteBackbone family=Qwen3, Model size=8B2026.0588.2
Refusal-style probeBackbone family=Llama, Model size=70B2026.0587.9
Best single-layer probeBackbone family=Gemma, Model size=9B2026.0587.8
MultiLayer-DIMBackbone family=Gemma, Model size=2B2026.0587.8
TaT-Disp-LSTMBackbone family=Qwen3, Model size=14B2026.0587.8
Geometry-LiteBackbone family=Gemma, Model size=2B2026.0587.8
Best single-layer probeBackbone family=Gemma, Model size=2B2026.0587.7
TaT-Disp-LSTMBackbone family=Llama, Model size=70B2026.0587.6
TaT-Disp-LSTMBackbone family=Gemma, Model size=9B2026.0587.6
Best single-layer probeBackbone family=Llama, Model size=8B2026.0587.5
Geometry-LiteBackbone family=Llama, Model size=1B2026.0587.5
Best single-layer probeBackbone family=Qwen3, Model size=14B2026.0587.4
TaT-Disp-LSTMBackbone family=Gemma, Model size=2B2026.0587.4
TaT-Disp-LSTMBackbone family=Gemma, Model size=27B2026.0587.4
TaT-Disp-LSTMBackbone family=Qwen3, Model size=8B2026.0587.3
MultiLayer-LinearBackbone family=Llama, Model size=1B2026.0587.2
MultiLayer-DIMBackbone family=Gemma, Model size=9B2026.0587.1
MultiLayer-DIMBackbone family=Gemma, Model size=27B2026.0587
MultiLayer-DIMBackbone family=Qwen3, Model size=8B2026.0586.9
Final-layer probeBackbone family=Gemma, Model size=27B2026.0586.2
TaT-Disp-LSTMBackbone family=Llama, Model size=1B2026.0586.2
Final-layer probeBackbone family=Qwen3, Model size=14B2026.0586.1
Best single-layer probeBackbone family=Qwen3, Model size=8B2026.0586
MultiLayer-DIMBackbone family=Llama, Model size=8B2026.0585.8
MultiLayer-DIMBackbone family=Llama, Model size=70B2026.0585.8
MultiLayer-DIMBackbone family=Qwen3, Model size=14B2026.0585.8
Final-layer probeBackbone family=Gemma, Model size=9B2026.0585.7
MultiLayer-LinearBackbone family=Qwen3, Model size=1.7B2026.0585.6
Final-layer probeBackbone family=Qwen3, Model size=8B2026.0585.4
Final-layer probeBackbone family=Llama, Model size=1B2026.0585.2
Geometry-LiteBackbone family=Qwen3, Model size=1.7B2026.0585.1
Final-layer probeBackbone family=Gemma, Model size=2B2026.0584.8
Best single-layer probeBackbone family=Llama, Model size=1B2026.0584.8
Refusal-style probeBackbone family=Llama, Model size=8B2026.0584.7
Refusal-style probeBackbone family=Gemma, Model size=27B2026.0584.5
TaT-Disp-LSTMBackbone family=Qwen3, Model size=1.7B2026.0583.9
MultiLayer-DIMBackbone family=Llama, Model size=1B2026.0582.8
Best single-layer probeBackbone family=Qwen3, Model size=1.7B2026.0582.5
Refusal-style probeBackbone family=Qwen3, Model size=8B2026.0582.4
Refusal-style probeBackbone family=Qwen3, Model size=14B2026.0582.3
Refusal-style probeBackbone family=Gemma, Model size=9B2026.0582.1
Refusal-style probeBackbone family=Llama, Model size=1B2026.0581.7
Final-layer probeBackbone family=Qwen3, Model size=1.7B2026.0581.6
MultiLayer-DIMBackbone family=Qwen3, Model size=1.7B2026.0579.8
Refusal-style probeBackbone family=Gemma, Model size=2B2026.0579.7
Refusal-style probeBackbone family=Qwen3, Model size=1.7B2026.0579.4