Hallucination Detection on HaluEvalQA
89ROC-AUCSinkProbe
Evaluation Results
| Method | Links | |
|---|---|---|
| SinkProbeLLM=Llama3.1-8B2026.04 | 89 | |
| MTopDivLLM=Llama3.1-8B2026.04 | 88.1 | |
| LookbackLensLLM=Llama3.1-8B2026.04 | 87.9 | |
| LapEigvalLLM=Llama3.1-8B2026.04 | 87.8 | |
| AttnEigvalLLM=Llama3.1-8B2026.04 | 85 | |
| AttnLogDetLLM=Llama3.1-8B2026.04 | 84.9 | |
| SinkProbeLLM=Mistral-Nemo2026.04 | 84.9 | |
| SinkProbeLLM=Llama3.2-3B2026.04 | 84.7 | |
| MTopDivLLM=Llama3.2-3B2026.04 | 84.6 | |
| SinkProbeLLM=Phi3.52026.04 | 84.6 | |
| LapEigvalLLM=Llama3.2-3B2026.04 | 83.7 | |
| LookbackLensLLM=Mistral-Nemo2026.04 | 83.7 | |
| MTopDivLLM=Mistral-Nemo2026.04 | 83.7 | |
| LookbackLensLLM=Llama3.2-3B2026.04 | 83.6 | |
| MTopDivLLM=Phi3.52026.04 | 83.5 | |
| LapEigvalLLM=Mistral-Nemo2026.04 | 83.1 | |
| LookbackLensLLM=Phi3.52026.04 | 82.8 | |
| LapEigvalLLM=Phi3.52026.04 | 82.2 | |
| AttnLogDetLLM=Llama3.2-3B2026.04 | 81.7 | |
| AttnEigvalLLM=Llama3.2-3B2026.04 | 81.7 | |
| AttnLogDetLLM=Phi3.52026.04 | 80.5 | |
| AttnEigvalLLM=Phi3.52026.04 | 80.3 | |
| AttnLogDetLLM=Mistral-Nemo2026.04 | 79.7 | |
| AttnEigvalLLM=Mistral-Nemo2026.04 | 79.1 | |
| AttnScoreLLM=Llama3.1-8B2026.04 | 77 | |
| AttnScoreLLM=Llama3.2-3B2026.04 | 73.7 | |
| AttnScoreLLM=Phi3.52026.04 | 71.2 | |
| AttnScoreLLM=Mistral-Nemo2026.04 | 68.8 | |
| HallucinationProbesDetector category=White-box, Probe=LoRA-KL probe on LLaMA-3.1-8B (lambda_KL = 0.05)2026.04 | 0.9493 | |
| RAGognizerDetector category=White-box, Proxy generator=Qwen3-4B-Instruct-25072026.04 | 0.9038 | |
| HDM-2-3BDetector category=Black-box2026.04 | 0.8385 | |
| SelfCheckGPT (NLI)Detector category=Black-box2026.04 | 0.8343 | |
| MiniCheck-7BDetector category=Black-box2026.04 | 0.8138 | |
| DeBERTa-v3 (Ent.)Detector category=Black-box2026.04 | 0.7832 | |
| LettuceDetect-LDetector category=Black-box2026.04 | 0.7702 | |
| HHEM-2.1-OpenDetector category=Black-box2026.04 | 0.77 | |
| Semantic EntropyDetector category=White-box2026.04 | 0.7407 | |
| DeBERTa-v3 (Con.)Detector category=Black-box2026.04 | 0.681 | |
| INSIDE (EigenScore)Detector category=White-box, Proxy generator=Qwen3-4B-Instruct-25072026.04 | 0.4746 |