Hallucination Detection on HaluBench (AUROC)
97AUROCF: Ans. Expect.
Evaluation Results
| Method | Links | |
|---|---|---|
| F: Ans. Expect.Backbone=Llama-3.3-70B2026.05 | 97 | |
| Ensemble A+C+E+FBackbone=Llama-3.3-70B2026.05 | 96 | |
| E: DRIFTBackbone=Llama-3.3-70B2026.05 | 91 | |
| DRIFT-concat (ablation of E)Backbone=Llama-3.3-70B2026.05 | 91 | |
| SAPLMABackbone=Llama-3.3-70B2026.05 | 91 | |
| HaMIBackbone=Llama-3.3-70B2026.05 | 87 | |
| D: Act. VarianceBackbone=Llama-3.3-70B2026.05 | 82 | |
| PRISMBackbone=Llama-3.3-70B2026.05 | 80 | |
| TXTEMB (ctrl)Backbone=Llama-3.3-70B2026.05 | 77 | |
| DRIFT-logpBackbone=Llama-3.3-70B2026.05 | 76 | |
| A: PerturbationBackbone=Llama-3.3-70B2026.05 | 68 | |
| IRIS⋆ (self-verify)Backbone=Llama-3.3-70B2026.05 | 67 | |
| MINDBackbone=Llama-3.3-70B2026.05 | 67 | |
| HaloScopeBackbone=Llama-3.3-70B2026.05 | 67 | |
| SEPsBackbone=Llama-3.3-70B2026.05 | 67 | |
| C: Sem. EntropyBackbone=Llama-3.3-70B2026.05 | 62 | |
| DoLaBackbone=Llama-3.3-70B2026.05 | 62 | |
| ACT (detect)Backbone=Llama-3.3-70B2026.05 | 62 | |
| P(True)Backbone=Llama-3.3-70B2026.05 | 60 | |
| Log-ProbBackbone=Llama-3.3-70B2026.05 | 56 | |
| SelfCheckGPT-NLIBackbone=Llama-3.3-70B2026.05 | 54 | |
| INSIDEBackbone=Llama-3.3-70B2026.05 | 50 | |
| HalluShiftBackbone=Llama-3.3-70B2026.05 | 31 | |
| SAPLMAModel=Gemma-3-12b2026.05 | 0.935 | |
| DRIFT-concatModel=Gemma-3-12b2026.05 | 0.928 | |
| DRIFTModel=Llama-3.3-70B2026.05 | 0.915 | |
| DRIFTModel=Gemma-3-12b2026.05 | 0.913 | |
| SAPLMAModel=Llama-3.3-70B2026.05 | 0.911 | |
| DRIFT-concatModel=Llama-3.3-70B2026.05 | 0.911 | |
| DRIFT-concatModel=Mistral-24B2026.05 | 0.91 | |
| DRIFT-concatModel=Gemma-3-4b2026.05 | 0.904 | |
| DRIFTModel=Qwen2.5-72B2026.05 | 0.9 | |
| DRIFTModel=Mistral-24B2026.05 | 0.9 | |
| SAPLMAModel=Mistral-24B2026.05 | 0.9 | |
| DRIFTModel=Phi-4-mini2026.05 | 0.9 | |
| DRIFT-concatModel=Phi-4-mini2026.05 | 0.9 | |
| DRIFT-concatModel=Qwen2.5-72B2026.05 | 0.898 | |
| DRIFT-concatModel=Mixtral-8x7B2026.05 | 0.898 | |
| DRIFTModel=Gemma-3-4b2026.05 | 0.896 | |
| SAPLMAModel=Gemma-3-4b2026.05 | 0.895 | |
| SAPLMAModel=Qwen3-8B2026.05 | 0.894 | |
| SAPLMAModel=Mixtral-8x7B2026.05 | 0.893 | |
| SAPLMAModel=Phi-4-mini2026.05 | 0.891 | |
| DRIFTModel=Mixtral-8x7B2026.05 | 0.889 | |
| SAPLMAModel=Qwen2.5-72B2026.05 | 0.885 | |
| DRIFT-concatModel=Qwen3-32B2026.05 | 0.883 | |
| DRIFT-concatModel=Mean2026.05 | 0.883 | |
| DRIFTModel=Qwen3-32B2026.05 | 0.88 | |
| DRIFT-concatModel=Qwen3-8B2026.05 | 0.879 | |
| DRIFTModel=Mean2026.05 | 0.879 | |
| SAPLMAModel=Mean2026.05 | 0.879 | |
| DRIFT-concatModel=DS-R1-Llama-70B2026.05 | 0.873 | |
| DRIFTModel=DS-R1-Llama-70B2026.05 | 0.871 | |
| DRIFTModel=Qwen3-8B2026.05 | 0.868 | |
| SAPLMAModel=DS-R1-Llama-70B2026.05 | 0.865 | |
| SAPLMAModel=Qwen3-32B2026.05 | 0.855 | |
| DRIFTModel=Llama-3.1-8B2026.05 | 0.839 | |
| SAPLMAModel=Llama-3.1-8B2026.05 | 0.824 | |
| DRIFT-concatModel=Llama-3.1-8B2026.05 | 0.823 | |
| SAPLMAModel=DS-R1-Qwen-7B2026.05 | 0.792 | |
| DRIFT-concatModel=DS-R1-Qwen-7B2026.05 | 0.788 | |
| DRIFTModel=DS-R1-Qwen-7B2026.05 | 0.774 | |
| SelfCheckGPT-NLIModel=Qwen3-32B, Passes=102026.05 | 0.679 | |
| SelfCheckGPT-NLIModel=Qwen3-8B, Passes=102026.05 | 0.678 | |
| SelfCheckGPT-NLIModel=Gemma-3-4b, Passes=102026.05 | 0.637 | |
| SelfCheckGPT-NLIModel=DS-R1-Llama-70B, Passes=102026.05 | 0.635 | |
| SelfCheckGPT-NLIModel=Qwen2.5-72B, Passes=102026.05 | 0.613 | |
| SelfCheckGPT-NLIModel=Mixtral-8x7B, Passes=102026.05 | 0.607 | |
| SelfCheckGPT-NLIModel=Llama-3.1-8B, Passes=102026.05 | 0.579 | |
| SelfCheckGPT-NLIModel=Mistral-24B, Passes=102026.05 | 0.564 | |
| SelfCheckGPT-NLIModel=Mean, Passes=102026.05 | 0.557 | |
| SelfCheckGPT-NLIModel=Llama-3.3-70B, Passes=102026.05 | 0.535 | |
| SelfCheckGPT-NLIModel=DS-R1-Qwen-7B, Passes=102026.05 | 0.524 | |
| SelfCheckGPT-NLIModel=Phi-4-mini, Passes=102026.05 | 0.48 | |
| SelfCheckGPT-NLIModel=Gemma-3-12b, Passes=102026.05 | 0.462 |