List your company

Harmfulness Detection on BeaverTails bottom 30% uncertainty slice (test)

85.1AUROCGeometry-Lite
68.9873.16577.3581.535May 18, 2026

Evaluation results

MethodLinks
Geometry-LiteBackbone=Qwen3-14B2026.0585.1
MultiLayer-LinearBackbone=Qwen3-1.7B2026.0583.8
Geometry-LiteBackbone=Qwen3-1.7B2026.0583.8
Geometry-LiteBackbone=Llama-70B2026.0583.7
MultiLayer-LinearBackbone=Llama-70B2026.0583.6
Geometry-LiteBackbone=Llama-8B2026.0583.4
MultiLayer-LinearBackbone=Llama-1B2026.0583
MultiLayer-LinearBackbone=Llama-8B2026.0582.9
MultiLayer-LinearBackbone=Qwen3-14B2026.0582.7
MultiLayer-LinearBackbone=Avg2026.0582.7
Geometry-LiteBackbone=Llama-1B2026.0582.6
Geometry-LiteBackbone=Avg2026.0582.6
TaT-Disp-LSTMBackbone=Qwen3-14B2026.0582.5
MultiLayer-LinearBackbone=Qwen3-8B2026.0582.4
TaT-Disp-LSTMBackbone=Qwen3-1.7B2026.0582.3
MultiLayer-LinearBackbone=Gemma-9B2026.0582.3
Geometry-LiteBackbone=Qwen3-8B2026.0582.2
MultiLayer-LinearBackbone=Gemma-2B2026.0582
MultiLayer-LinearBackbone=Gemma-27B2026.0581.9
Geometry-LiteBackbone=Gemma-2B2026.0581
Geometry-LiteBackbone=Gemma-9B2026.0581
Geometry-LiteBackbone=Gemma-27B2026.0580.7
Best single-layer probeBackbone=Llama-70B2026.0580.6
Final-layer probeBackbone=Llama-70B2026.0580
Refusal-style probeBackbone=Llama-70B2026.0579.9
TaT-Disp-LSTMBackbone=Llama-1B2026.0579.8
TaT-Disp-LSTMBackbone=Qwen3-8B2026.0579.5
TaT-Disp-LSTMBackbone=Avg2026.0579.4
TaT-Disp-LSTMBackbone=Gemma-2B2026.0579.1
Final-layer probeBackbone=Llama-8B2026.0578.9
TaT-Disp-LSTMBackbone=Llama-8B2026.0578.3
TaT-Disp-LSTMBackbone=Gemma-9B2026.0578.1
TaT-Disp-LSTMBackbone=Llama-70B2026.0577.4
TaT-Disp-LSTMBackbone=Gemma-27B2026.0577.4
Best single-layer probeBackbone=Gemma-9B2026.0577.3
Final-layer probeBackbone=Qwen3-14B2026.0577.1
Best single-layer probeBackbone=Llama-8B2026.0576.9
MultiLayer-DIMBackbone=Gemma-2B2026.0576.6
Best single-layer probeBackbone=Gemma-27B2026.0576.3
Refusal-style probeBackbone=Llama-8B2026.0576.3
Final-layer probeBackbone=Llama-1B2026.0576.1
Final-layer probeBackbone=Avg2026.0576.1
Best single-layer probeBackbone=Avg2026.0575.8
Best single-layer probeBackbone=Gemma-2B2026.0575.7
Refusal-style probeBackbone=Qwen3-1.7B2026.0575.7
Final-layer probeBackbone=Gemma-27B2026.0575.5
MultiLayer-DIMBackbone=Gemma-9B2026.0575.2
Final-layer probeBackbone=Gemma-9B2026.0575
Final-layer probeBackbone=Qwen3-1.7B2026.0575
Final-layer probeBackbone=Gemma-2B2026.0574.9
Best single-layer probeBackbone=Qwen3-8B2026.0574.9
Best single-layer probeBackbone=Qwen3-14B2026.0574.8
Refusal-style probeBackbone=Llama-1B2026.0574.5
Refusal-style probeBackbone=Avg2026.0574.1
Best single-layer probeBackbone=Llama-1B2026.0573.5
Refusal-style probeBackbone=Qwen3-8B2026.0573.5
Refusal-style probeBackbone=Gemma-27B2026.0573.1
Refusal-style probeBackbone=Gemma-9B2026.0572.9
Final-layer probeBackbone=Qwen3-8B2026.0572.8
MultiLayer-DIMBackbone=Gemma-27B2026.0572.5
MultiLayer-DIMBackbone=Qwen3-14B2026.0572.3
Best single-layer probeBackbone=Qwen3-1.7B2026.0572.2
MultiLayer-DIMBackbone=Llama-70B2026.0572
MultiLayer-DIMBackbone=Avg2026.0572
Refusal-style probeBackbone=Qwen3-14B2026.0570.9
MultiLayer-DIMBackbone=Llama-8B2026.0570.3
MultiLayer-DIMBackbone=Llama-1B2026.0569.9
MultiLayer-DIMBackbone=Qwen3-1.7B2026.0569.8
MultiLayer-DIMBackbone=Qwen3-8B2026.0569.7
Refusal-style probeBackbone=Gemma-2B2026.0569.6