Hallucination Detection on NQ-Open
0.9304AUROCHCPD
Evaluation results
| Method | Links | |
|---|---|---|
| HCPDSource Model=Qwen-3-8b, Target Model=LLaMA-2-13b2026.06 | 0.9304 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=LLaMA-2-7b2026.06 | 0.9294 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=Qwen-3-8b2026.06 | 0.9245 | |
| HCPDSource Model=Qwen-3-8b, Target Model=LLaMA-2-7b2026.06 | 0.916 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=LLaMA-2-13b2026.06 | 0.9042 | |
| HCPDModel=Llama-3.1-8b, Supervision=Zero-source2026.06 | 0.9038 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=LLaMA-3.1-8b2026.06 | 0.9038 | |
| DRIFTModel=Qwen-2.5-7B-Instruct, Input=answer2026.01 | 0.8843 | |
| HCPDModel=Qwen-3-8b, Supervision=Zero-source2026.06 | 0.8735 | |
| HCPDSource Model=Qwen-3-8b, Target Model=Qwen-3-8b2026.06 | 0.8735 | |
| DRIFTModel=Gemma-3-4b-it, Input=answer2026.01 | 0.8604 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=Qwen-2.5-7b2026.06 | 0.8588 | |
| HaloScopeModel=Qwen-2.5-7B-Instruct, Input=answer2026.01 | 0.8584 | |
| HaloScopeModel=Gemma-3-4b-it, Input=answer2026.01 | 0.8584 | |
| HCPDSource Model=Qwen-3-8b, Target Model=Qwen-2.5-7b2026.06 | 0.857 | |
| DRIFTModel=LLaMA 2 Chat 7B, Input=answer2026.01 | 0.8376 | |
| DRIFTModel=Qwen-2.5-7B-Instruct, Input=question2026.01 | 0.8328 | |
| DRIFTModel=Gemma-3-4b-it, Input=question2026.01 | 0.8296 | |
| SinkProbeLLM=Phi3.52026.04 | 0.821 | |
| LapEigvalLLM=Phi3.52026.04 | 0.82 | |
| MTopDivLLM=Phi3.52026.04 | 0.818 | |
| LookbackLensLLM=Phi3.52026.04 | 0.816 | |
| SelfCKGPTModel=Qwen-3-8b, Supervision=Zero-source2026.06 | 0.8051 | |
| Semantic EntropyModel=Qwen-2.5-7B-Instruct2026.01 | 0.8022 | |
| AttnLogDetLLM=Phi3.52026.04 | 0.801 | |
| DRIFTModel=LLaMA 2 Chat 7B, Input=question2026.01 | 0.7987 | |
| HaloScopeModel=Gemma-3-4b-it, Input=question2026.01 | 0.7962 | |
| HCPDSource Model=Qwen-3-8b, Target Model=LLaMA-3.1-8b2026.06 | 0.7942 | |
| SinkProbeLLM=Llama3.1-8B2026.04 | 0.789 | |
| LapEigvalLLM=Llama3.1-8B2026.04 | 0.787 | |
| SinkProbeLLM=Mistral-Nemo2026.04 | 0.787 | |
| HCPDSource Model=LLaMA-3.1-8b, Target Model=Qwen-2.5-14b2026.06 | 0.7862 | |
| EigenWD M1Target Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.03 | 0.7861 | |
| LookbackLensLLM=Mistral-Nemo2026.04 | 0.782 | |
| AttnEigvalLLM=Phi3.52026.04 | 0.779 | |
| LookbackLensLLM=Llama3.1-8B2026.04 | 0.776 | |
| MTopDivLLM=Mistral-Nemo2026.04 | 0.776 | |
| Semantic EntropyModel=LLaMA 2 Chat 7B2026.01 | 0.7757 | |
| FESSignal=Free-energy + SFF2026.06 | 0.775 | |
| MTopDivLLM=Llama3.1-8B2026.04 | 0.772 | |
| TADModel=Qwen-3-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.7682 | |
| LapEigvalLLM=Mistral-Nemo2026.04 | 0.765 | |
| SAPLMAModel=Llama-3.1-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.7623 | |
| SAPLMASource Model=LLaMA-3.1-8b, Target Model=LLaMA-3.1-8b2026.06 | 0.7623 | |
| AttnLogDetLLM=Llama3.1-8B2026.04 | 0.759 | |
| AttnEigvalLLM=Llama3.1-8B2026.04 | 0.759 | |
| SinkProbeLLM=Llama3.2-3B2026.04 | 0.758 | |
| Semantic EntropyModel=Gemma-3-4b-it2026.01 | 0.7548 | |
| MTopDivLLM=Llama3.2-3B2026.04 | 0.754 | |
| Semantic EntropyModel=Qwen-3-8b, Supervision=Zero-source2026.06 | 0.751 | |
| AttnLogDetLLM=Mistral-Nemo2026.04 | 0.75 | |
| Best spec.Description=Strongest non-FES in-tree attention-spectral baseline2026.06 | 0.75 | |
| LapEigvalsLLM=Llama-3.1-8b2026.01 | 0.748 | |
| HaloScopeSource Model=LLaMA-3.1-8b, Target Model=LLaMA-2-7b2026.06 | 0.7435 | |
| LookbackLensLLM=Llama3.2-3B2026.04 | 0.743 | |
| LapEigvalLLM=Llama3.2-3B2026.04 | 0.736 | |
| EmbRegressModel=gpt-4o-mini-2024-07-18, Category=Multi-trajectory / LLM-based2026.05 | 0.733 | |
| AttnEigvalsLLM=Llama-3.1-8b2026.01 | 0.732 | |
| AttnEigvalLLM=Llama3.2-3B2026.04 | 0.732 | |
| PersImgLLM=Llama-3.1-8b2026.01 | 0.73 | |
| AttnEigvalLLM=Mistral-Nemo2026.04 | 0.73 | |
| AttnScoreLLM=Phi3.52026.04 | 0.729 | |
| SAPLMAModel=Qwen-3-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.7286 | |
| SAPLMASource Model=Qwen-3-8b, Target Model=Qwen-3-8b2026.06 | 0.7286 | |
| AttnLogDetLLM=Llama3.2-3B2026.04 | 0.723 | |
| Betti CurveLLM=Llama-3.1-8b2026.01 | 0.715 | |
| PersImgLLM=Llama-3.2-3b2026.01 | 0.712 | |
| DSETarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.03 | 0.7114 | |
| SAPLMASource Model=Qwen-3-8b, Target Model=LLaMA-3.1-8b2026.06 | 0.7055 | |
| HaloScopeModel=Qwen-2.5-7B-Instruct, Input=question2026.01 | 0.7033 | |
| TSVModel=Llama-3.1-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.7017 | |
| TSVSource Model=LLaMA-3.1-8b, Target Model=LLaMA-3.1-8b2026.06 | 0.7017 | |
| Best rep.Description=Strongest reported reference hidden-state or sampling baseline2026.06 | 0.7 | |
| AttnEigvalsLLM=Llama-3.2-3b2026.01 | 0.694 | |
| LapEigvalsLLM=Llama-3.2-3b2026.01 | 0.693 | |
| TADModel=Llama-3.1-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.6888 | |
| SAPLMASource Model=Qwen-3-8b, Target Model=Qwen-2.5-14b2026.06 | 0.6887 | |
| Betti CurveLLM=Llama-3.2-3b2026.01 | 0.688 | |
| Best LMDescription=Best language-model probability baseline among MSP and PPL−12026.06 | 0.688 | |
| LN-EntropyModel=Qwen-3-8b, Supervision=Zero-source2026.06 | 0.687 | |
| SAPLMASource Model=Qwen-3-8b, Target Model=Qwen-2.5-7b2026.06 | 0.6852 | |
| PersEntropyLLM=Llama-3.2-3b2026.01 | 0.685 | |
| SAPLMASource Model=LLaMA-3.1-8b, Target Model=Qwen-3-8b2026.06 | 0.6827 | |
| PersEntropyLLM=Llama-3.1-8b2026.01 | 0.682 | |
| LSTarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.03 | 0.6804 | |
| ERTarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.03 | 0.6797 | |
| PerplexityModel=Qwen-3-8b, Supervision=Zero-source2026.06 | 0.6781 | |
| AttnScoreLLM=Llama3.1-8B2026.04 | 0.677 | |
| HaloScopeModel=LLaMA 2 Chat 7B, Input=answer2026.01 | 0.6705 | |
| HaloScopeModel=LLaMA 2 Chat 7B, Input=question2026.01 | 0.6697 | |
| SAPLMASource Model=Qwen-3-8b, Target Model=LLaMA-2-7b2026.06 | 0.6685 | |
| SAPLMASource Model=LLaMA-3.1-8b, Target Model=LLaMA-2-7b2026.06 | 0.6677 | |
| AttnScoreLLM=Mistral-Nemo2026.04 | 0.666 | |
| AttnScoreLLM=Llama3.2-3B2026.04 | 0.661 | |
| LNETarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.03 | 0.6585 | |
| SAPLMASource Model=LLaMA-3.1-8b, Target Model=LLaMA-2-13b2026.06 | 0.6555 | |
| HaloScopeSource Model=Qwen-3-8b, Target Model=LLaMA-2-7b2026.06 | 0.6484 | |
| SAPLMASource Model=LLaMA-3.1-8b, Target Model=Qwen-2.5-14b2026.06 | 0.6382 | |
| HaloScopeSource Model=LLaMA-3.1-8b, Target Model=Qwen-3-8b2026.06 | 0.6368 | |
| HaloScopeModel=Llama-3.1-8b, Supervision=Trained on fully labeled datasets2026.06 | 0.6338 |