Token-Level Hallucination Detection on RAGTruth QA
95.6AUROCLettuceDetect-L
Evaluation Results
| Method | Links | |
|---|---|---|
| LettuceDetect-LDetector Category=Black-box, Same-dataset evaluation=true2026.04 | 95.6 | |
| HDM-2-3BDetector Category=Black-box, Same-dataset evaluation=true2026.04 | 90.61 | |
| RAGognizerDetector Category=White-box, Proxy Generator=Qwen3-4B-Instruct-25072026.04 | 90.25 | |
| LookbackLensDetector Category=White-box, Same-dataset evaluation=true, Proxy Generator=Qwen3-4B-Instruct-25072026.04 | 85.42 | |
| HallucinationProbesDetector Category=White-box, Probe=LoRA-KL on LLaMA-3.1-8B (λKL = 0.05)2026.04 | 70.13 | |
| PerplexityDetector Category=White-box, Proxy Generator=Qwen3-4B-Instruct-25072026.04 | 64.59 | |
| HaloScopeDetector Category=White-box, Same-dataset evaluation=true, Proxy Generator=Qwen3-4B-Instruct-25072026.04 | 55.93 |