Hallucination Detection on TruthfulQA
0.9417AUC (ROC)ARS (CCS)
Evaluation results
| Method | Links | |
|---|---|---|
| ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.01 | 0.9417 | |
| ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.01 | 0.9375 | |
| TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.01 | 0.9293 | |
| HALLUSHIFT++Model=OPT-6.7B2025.12 | 0.9268 | |
| TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.01 | 0.913 | |
| HALLUSHIFT++Model=LLaMA-2-7B2025.12 | 0.9082 | |
| HALLUSHIFTModel=LLaMA-2-7B2025.12 | 0.8993 | |
| HALLUSHIFTModel=OPT-6.7B2025.12 | 0.8991 | |
| SAPLMAModel=Qwen2.5-72B2026.05 | 0.874 | |
| B: CAABackbone=Llama-3.3-70B2026.05 | 0.87 | |
| ARS (CCS)Model=Qwen3-8B, Single Sampling=true, Supervision=false2026.01 | 0.8664 | |
| Adaptive Bayesian EstimationLLM=Mistral-Small-24B, Sampling budget (N)=52026.03 | 0.839 | |
| ARS (Probing)Model=Qwen3-8B, Single Sampling=true, Supervision=true2026.01 | 0.8366 | |
| SinkProbeLLM=Mistral-Nemo2026.04 | 0.821 | |
| G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.01 | 0.8195 | |
| Adaptive Bayesian EstimationLLM=Llama-3.1-8B, Sampling budget (N)=52026.03 | 0.818 | |
| DRIFT-concatModel=Qwen3-32B, Quantization=8-bit re-extraction2026.05 | 0.817 | |
| DRIFT-concatModel=Qwen2.5-72B2026.05 | 0.812 | |
| MTopDivLLM=Mistral-Nemo2026.04 | 0.811 | |
| ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=false2026.01 | 0.8089 | |
| DRIFTModel=Qwen2.5-72B2026.05 | 0.808 | |
| LapEigvalLLM=Mistral-Nemo2026.04 | 0.806 | |
| LookbackLensLLM=Mistral-Nemo2026.04 | 0.804 | |
| AttnLogDetLLM=Mistral-Nemo2026.04 | 0.798 | |
| MTopDivLLM=Phi3.52026.04 | 0.797 | |
| Adaptive Bayesian EstimationLLM=Llama-2-7B, Sampling budget (N)=52026.03 | 0.795 | |
| ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=DeepSeek-R1-Distill-Qwen-32B2026.01 | 0.7946 | |
| HaloScopeModel=LLaMA-2-7B2025.12 | 0.7864 | |
| SinkProbeLLM=Llama3.2-3B2026.04 | 0.785 | |
| AttnEigvalLLM=Mistral-Nemo2026.04 | 0.779 | |
| SinkProbeLLM=Llama3.1-8B2026.04 | 0.778 | |
| MTopDivLLM=Llama3.1-8B2026.04 | 0.775 | |
| AttnLogDetLLM=Phi3.52026.04 | 0.774 | |
| Adaptive Bayesian EstimationLLM=Mistral-Small-24B, Sampling budget (N)=22026.03 | 0.773 | |
| LookbackLensLLM=Phi3.52026.04 | 0.773 | |
| AttnEigvalLLM=Llama3.1-8B2026.04 | 0.773 | |
| SESDLGLLM=Mistral-Small-24B, Sampling budget (N)=52026.03 | 0.772 | |
| AttnLogDetLLM=Llama3.2-3B2026.04 | 0.771 | |
| TSVModel=Qwen3-8B, Single Sampling=true, Supervision=true2026.01 | 0.7708 | |
| AttnEigvalLLM=Phi3.52026.04 | 0.77 | |
| ARS (Probing)Model=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.01 | 0.7698 | |
| LookbackLensLLM=Llama3.1-8B2026.04 | 0.767 | |
| SELLM=Mistral-Small-24B, Sampling budget (N)=52026.03 | 0.765 | |
| AttnEigvalLLM=Llama3.2-3B2026.04 | 0.765 | |
| AttnLogDetLLM=Llama3.1-8B2026.04 | 0.765 | |
| SinkProbeLLM=Phi3.52026.04 | 0.761 | |
| MTopDivLLM=Llama3.2-3B2026.04 | 0.757 | |
| LapEigvalLLM=Llama3.1-8B2026.04 | 0.757 | |
| LapEigvalLLM=Phi3.52026.04 | 0.755 | |
| Adaptive Bayesian EstimationLLM=Llama-3.1-8B, Sampling budget (N)=22026.03 | 0.753 | |
| SAPLMAModel=Llama-3.3-70B2026.05 | 0.751 | |
| SESDLGLLM=Llama-3.1-8B, Sampling budget (N)=52026.03 | 0.75 | |
| SAPLMABackbone=Llama-3.3-70B2026.05 | 0.75 | |
| FESSignal=Free-energy + SFF2026.06 | 0.75 | |
| LapEigvalLLM=Llama3.2-3B2026.04 | 0.748 | |
| DRIFT-concatModel=Qwen3-8B2026.05 | 0.744 | |
| SELLM=Llama-3.1-8B, Sampling budget (N)=52026.03 | 0.741 | |
| LookbackLensLLM=Llama3.2-3B2026.04 | 0.741 | |
| DRIFT-concat (ablation of E)Backbone=Llama-3.3-70B2026.05 | 0.74 | |
| DRIFT-concatModel=Llama-3.3-70B2026.05 | 0.737 | |
| DRIFT-concatModel=DS-R1-Llama-70B2026.05 | 0.735 | |
| Adaptive Bayesian EstimationLLM=Llama-2-7B, Sampling budget (N)=22026.03 | 0.732 | |
| HaloScopeModel=OPT-6.7B2025.12 | 0.7317 | |
| G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.01 | 0.7308 | |
| E: DRIFTBackbone=Llama-3.3-70B2026.05 | 0.73 | |
| DRIFTModel=Llama-3.3-70B2026.05 | 0.729 | |
| SAPLMAModel=DS-R1-Llama-70B2026.05 | 0.729 | |
| TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=DeepSeek-R1-Distill-Qwen-32B2026.01 | 0.7258 | |
| SESDLGLLM=Llama-2-7B, Sampling budget (N)=52026.03 | 0.724 | |
| SAPLMA2025.12 | 0.72 | |
| DRIFTModel=DS-R1-Llama-70B2026.05 | 0.72 | |
| Best spec.Description=Strongest non-FES in-tree attention-spectral baseline2026.06 | 0.72 | |
| G-DetectorModel=Qwen3-8B, Single Sampling=true, Supervision=true2026.01 | 0.7186 | |
| DRIFT-concatModel=Mean2026.05 | 0.718 | |
| SELLM=Llama-2-7B, Sampling budget (N)=52026.03 | 0.713 | |
| SAPLMAModel=DS-R1-Qwen-7B2026.05 | 0.711 | |
| DRIFT-concatModel=Mistral-24B2026.05 | 0.707 | |
| SAPLMAModel=Qwen3-32B, Quantization=8-bit re-extraction2026.05 | 0.706 | |
| SAPLMAModel=Mean2026.05 | 0.706 | |
| SAPLMAModel=Mistral-24B2026.05 | 0.705 | |
| AttnScoreLLM=Llama3.1-8B2026.04 | 0.704 | |
| SAPLMAModel=Phi-4-mini2026.05 | 0.703 | |
| G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.01 | 0.7001 | |
| Ensemble A+C+E+FBackbone=Llama-3.3-70B2026.05 | 0.7 | |
| DRIFT-concatModel=Mixtral-8x7B2026.05 | 0.7 | |
| TSVModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.01 | 0.6949 | |
| DRIFT-concatModel=Phi-4-mini2026.05 | 0.694 | |
| AttnScoreLLM=Phi3.52026.04 | 0.693 | |
| DRIFT-concatModel=DS-R1-Qwen-7B2026.05 | 0.693 | |
| TXTEMB (ctrl)Backbone=Llama-3.3-70B2026.05 | 0.69 | |
| DRIFTModel=Mistral-24B2026.05 | 0.683 | |
| P(True)Backbone=Llama-3.3-70B2026.05 | 0.68 | |
| CCS*Model=LLaMA-2-7B2025.12 | 0.6795 | |
| AttnScoreLLM=Mistral-Nemo2026.04 | 0.678 | |
| SAPLMAModel=Mixtral-8x7B2026.05 | 0.677 | |
| RACEModel=Qwen3-8B, Single Sampling=false, Supervision=false2026.01 | 0.6757 | |
| DRIFTModel=Qwen3-8B2026.05 | 0.675 | |
| SESDLGLLM=Mistral-Small-24B, Sampling budget (N)=22026.03 | 0.673 | |
| SAPLMAModel=Qwen3-8B2026.05 | 0.671 | |
| DRIFTModel=Qwen3-32B, Quantization=8-bit re-extraction2026.05 | 0.671 |