Hallucination detection on FAVA Annotated Dataset
83.28AUCROCPersImg
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| PersImgBackbone=Vicuna-7b2026.01 | 83.28 | 77.17 | 35.71 | 82.35 | |
| PersImgBackbone=Llama-3-8b2026.01 | 82.64 | 73.91 | 46.43 | 80.33 | |
| PersImgBackbone=Llama-2-7b2026.01 | 82.09 | 75 | 26.79 | 80.67 | |
| Betti CurveBackbone=Vicuna-7b2026.01 | 76.71 | 70.65 | 24.59 | 80.29 | |
| PersEntropyBackbone=Llama-2-7b2026.01 | 75.67 | 72.82 | 21.43 | 78.26 | |
| PersEntropyBackbone=Vicuna-7b2026.01 | 75.47 | 68.47 | 17.86 | 76.03 | |
| Betti CurveBackbone=Llama-2-7b2026.01 | 74.75 | 68.47 | 23.21 | 77.86 | |
| Betti CurveBackbone=Llama-3-8b2026.01 | 74.33 | 70.65 | 21.43 | 77.69 | |
| PersEntropyBackbone=Llama-3-8b2026.01 | 74.09 | 71.73 | 19.29 | 78.69 | |
| LLM-Check (Attn Score)Backbone=Llama-2-7b2026.01 | 72.34 | 67.96 | 14.97 | 69.27 | |
| LLM-Check (Attn Score)Backbone=Vicuna-7b2026.01 | 71.69 | 66.47 | 24.55 | 62 | |
| LLM-Check (Attn Score)Backbone=Llama-3-8b2026.01 | 68.19 | 65.87 | 15.57 | 70.53 | |
| INSIDE2026.01 | 59.03 | 57.98 | 13.17 | 39.66 | |
| FAVA Model2026.01 | 53.29 | 53.29 | — | 43.88 | |
| Self-Prompt2026.01 | 50.3 | 50.3 | — | 66.53 | |
| SelfCheckGPT-Prompt2026.01 | 50.08 | 54.19 | — | 67.24 |