Hallucination Detection on MHaluBench Image-to-Text (Claim-level)
86.54Hallucinatory PrecisionGPT-based Self-Check
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-based Self-Checkn (number of few-shot examples)=2, Backbone=GPT2025.06 | 86.54 | 85.13 | 85.83 | 69.05 | 71.48 | 70.24 | 80.8 | 77.8 | 78.3 | 78.04 | |
| ZINA2025.06 | 84.91 | 89.52 | 87.15 | 84.91 | 89.52 | 87.15 | 85.39 | 86.07 | 80.28 | 83.07 | |
| GPT-based Self-Checkn (number of few-shot examples)=0, Backbone=GPT2025.06 | 84.78 | 80.07 | 82.35 | 61.64 | 69.01 | 65.12 | 76.56 | 73.21 | 74.54 | 73.73 | |
| Gemini-based UniHDBackbone=Gemini2025.06 | 84.44 | 72.44 | 77.98 | 71.08 | 83.54 | 76.8 | 77.41 | 77.76 | 77.99 | 77.39 | |
| Gemini-based Self-Checkn (number of few-shot examples)=2, Backbone=Gemini2025.06 | 84.24 | 66.75 | 74.48 | 67.35 | 84.6 | 75 | 74.74 | 75.8 | 75.68 | 74.74 | |
| Gemini-based Self-Checkn (number of few-shot examples)=0, Backbone=Gemini2025.06 | 83.17 | 42.15 | 55.95 | 55.64 | 89.48 | 68.61 | 63.34 | 69.41 | 65.82 | 62.28 | |
| GPT-based UniHDBackbone=GPT2025.06 | 82.54 | 85.29 | 83.89 | 81.08 | 77.74 | 79.38 | 81.91 | 81.81 | 81.52 | 81.63 |