Hallucination Detection on CoQA (AUROC, PD)
91.74AUROCMultiple Hypothesis Testing for Hallucination Detection
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Multiple Hypothesis Testing for Hallucination DetectionModel=DeepSeek-v2-Lite2025.08 | 91.74 | 76.45 | |
| Multiple Hypothesis Testing for Hallucination DetectionModel=Mistral-7B2025.08 | 91.23 | 75.81 | |
| EigVModel=DeepSeek-v2-Lite2025.08 | 91.06 | 74.34 | |
| Multiple Hypothesis Testing for Hallucination DetectionModel=Llama-2-13B2025.08 | 90.85 | 74.9 | |
| α_clustered_SEModel=DeepSeek-v2-Lite2025.08 | 90.8 | 74.78 | |
| Multiple Hypothesis Testing for Hallucination DetectionModel=Llama-3.1-8B2025.08 | 90.44 | 74.23 | |
| EigVModel=Mistral-7B2025.08 | 90.28 | 73.49 | |
| α_clustered_SEModel=Mistral-7B2025.08 | 90.04 | 73.62 | |
| EigVModel=Llama-2-13B2025.08 | 90.03 | 72.08 | |
| LSModel=DeepSeek-v2-Lite2025.08 | 89.83 | 70.94 | |
| α_clustered_SEModel=Llama-2-13B2025.08 | 89.79 | 72.48 | |
| EigVModel=Llama-3.1-8B2025.08 | 89.79 | 72.07 | |
| α_clustered_SEModel=Llama-3.1-8B2025.08 | 89.46 | 71.89 | |
| LSModel=Mistral-7B2025.08 | 89.28 | 68.25 | |
| KSEModel=DeepSeek-v2-Lite2025.08 | 89.05 | 69.53 | |
| KSEModel=Llama-2-13B2025.08 | 88.72 | 70.5 | |
| KSEModel=Mistral-7B2025.08 | 88.43 | 68.59 | |
| LSModel=Llama-2-13B2025.08 | 88.36 | 64.33 | |
| LSModel=Llama-3.1-8B2025.08 | 87.87 | 61.92 | |
| KSEModel=Llama-3.1-8B2025.08 | 87.77 | 66.96 | |
| αSEModel=DeepSeek-v2-Lite2025.08 | 86.55 | 66.73 | |
| clustered_SEModel=DeepSeek-v2-Lite2025.08 | 86.42 | 61.03 | |
| clustered_SEModel=Llama-2-13B2025.08 | 85.56 | 57.05 | |
| clustered_SEModel=Mistral-7B2025.08 | 85.55 | 58.23 | |
| αSEModel=Llama-2-13B2025.08 | 85.19 | 65.1 | |
| clustered_SEModel=Llama-3.1-8B2025.08 | 85.18 | 58.2 | |
| αSEModel=Mistral-7B2025.08 | 85.15 | 65.42 | |
| αSEModel=Llama-3.1-8B2025.08 | 84.17 | 61.96 | |
| IDBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 82.64 | — | |
| FEPoIDBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 82.64 | — | |
| Val LossBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 81.62 | — | |
| CurvatureBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 79.51 | — | |
| RankMEBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 76.24 | — | |
| RGNBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 72.85 | — | |
| SEModel=DeepSeek-v2-Lite2025.08 | 69.05 | 41.98 | |
| SEModel=Llama-2-13B2025.08 | 66.68 | 38.93 | |
| SNRBackbone=LlaMA-3.1-8B, Forward horizon (w)=7, Representation extraction method=FST, Model tuning status=non-instruction-tuned (base model)2026.05 | 65.65 | — | |
| SEModel=Mistral-7B2025.08 | 65.41 | 36.62 | |
| SEModel=Llama-3.1-8B2025.08 | 65.32 | 37.11 |