Object Hallucination Evaluation on AMBER
42.3Hallucination Score (Hal)PAI
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| PAIModel=LLaVA-1.5-7B2026.05 | 42.3 | 6.7 | 48.8 | 1.9 | 74.3 | 87.2 | 73.8 | 74.8 | 84.05 | |
| RegularModel=LLaVA-1.5-7B2026.05 | 36.7 | 8.3 | 49.4 | 4.4 | 72.7 | 84.7 | 62.1 | 73.7 | 82.7 | |
| SFT onlyBase Model=LLaVA-OneVision-7B [3]2026.06 | 35.4 | — | — | — | — | — | — | — | — | |
| SFT + DPOBase Model=LLaVA-OneVision-7B [3]2026.06 | 34.3 | — | — | — | — | — | — | — | — | |
| SFT onlyBase Model=InternVL2.5-26B [4]2026.06 | 31.2 | — | — | — | — | — | — | — | — | |
| SFT + DPOBase Model=InternVL2.5-26B [4]2026.06 | 29.8 | — | — | — | — | — | — | — | — | |
| SFT + DA-DPOBase Model=LLaVA-OneVision-7B [3]2026.06 | 28 | — | — | — | — | — | — | — | — | |
| SFT + DA-DPOBase Model=InternVL2.5-26B [4]2026.06 | 25.5 | — | — | — | — | — | — | — | — | |
| SFT + VIGIL (Ours)Base Model=LLaVA-OneVision-7B [3]2026.06 | 25.1 | — | — | — | — | — | — | — | — | |
| CRGModel=LLaVA-1.5-7B2026.05 | 23.2 | 4.6 | 53.3 | 2.1 | 77.4 | 92.3 | 72.1 | 77.5 | 86.45 | |
| SFT + VIGIL (Ours)Base Model=InternVL2.5-26B [4]2026.06 | 22.3 | — | — | — | — | — | — | — | — |