Hallucination Detection on CWQ (F1, Precision, Recall, Accuracy, AUC-PR)
87.4F1 ScoreKG-Guard
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| KG-GuardModel Backbone=GraphTransformer2026.05 | 87.4 | 85.8 | 90.2 | 83.5 | 94.2 | |
| KG-GuardModel Backbone=GAT2026.05 | 86.3 | 84.5 | 89.6 | 81.7 | 93.1 | |
| GPT 5 MiniInput Context=G+RS+RT2026.05 | 84.2 | 93.5 | 76.7 | 79.4 | 88.4 | |
| Most Frequent2026.05 | 83.5 | 71.7 | 100 | 71.7 | 71.7 | |
| GPT 5 MiniInput Context=G2026.05 | 82.6 | 90.2 | 76.2 | 77 | 85.8 | |
| SelfCheckSampling size (N)=102026.05 | 72.2 | 90.4 | 60.2 | 66.9 | 83 | |
| SelfCheckSampling size (N)=52026.05 | 67.1 | 91 | 53.1 | 62.6 | 81.9 | |
| Llama 4 ScoutInput Context=G+RS2026.05 | 61.3 | 94 | 45.4 | 58.8 | 81.8 | |
| Llama 4 ScoutInput Context=G+RS+RT2026.05 | 61.2 | 94.4 | 45.3 | 58.8 | 82 | |
| Llama 4 ScoutInput Context=G2026.05 | 60.2 | 91.6 | 44.8 | 57.5 | 80.7 | |
| Random2026.05 | 58.3 | 71.4 | 49.3 | 49.4 | 71.6 |