Fine-grained hallucination evaluation on AMBER
8CHAIRQwen2.5-VL-3B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen2.5-VL-3BBackbone=Qwen2.5-VL-3B2026.06 | 8 | 44.4 | 4.3 | 86 | |
| LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B2026.06 | 7.8 | 36.4 | 4.2 | 71.7 | |
| VCDBackbone=Qwen2.5-VL-3B2026.06 | 7.7 | 40.8 | 4.2 | 85.3 | |
| VCDBackbone=LLaVA-v1.5-7B2026.06 | 7.4 | 34.7 | 4.1 | 71.8 | |
| VCDBackbone=LLaVA-v1.5-13B2026.06 | 7.2 | 28.7 | 3.3 | 73.5 | |
| LLaVA-v1.5-13BBackbone=LLaVA-v1.5-13B2026.06 | 7 | 29 | 3.2 | 71.4 | |
| OPERABackbone=LLaVA-v1.5-7B2026.06 | 6.9 | 30.3 | 3.2 | 75.2 | |
| OPERABackbone=LLaVA-v1.5-13B2026.06 | 6.8 | 28.5 | 3.1 | — | |
| DPOBackbone=LLaVA-v1.5-7B2026.06 | 6.5 | 34.5 | 2.3 | 71.3 | |
| AVISCBackbone=LLaVA-v1.5-7B2026.06 | 6.25 | 25.6 | 2 | — | |
| ADAPT-TFBackbone=Qwen2.5-VL-3B2026.06 | 6.2 | 23.7 | 1.1 | 88.9 | |
| DPOBackbone=LLaVA-v1.5-13B2026.06 | 6.1 | 26.3 | 2.7 | — | |
| DPOBackbone=Qwen2.5-VL-3B2026.06 | 5.7 | 22.6 | 1.4 | 82.8 | |
| ADAPT-TFBackbone=Qwen2.5-VL-7B2026.06 | 5.3 | 19.5 | 0.8 | 86.3 | |
| VCDBackbone=Qwen2.5-VL-7B2026.06 | 4.8 | 23.9 | 1.1 | 74.2 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.06 | 4.7 | 21.9 | 1.2 | 87.6 | |
| mDPOBackbone=LLaVA-v1.5-7B2026.06 | 4.4 | 24.5 | 2.4 | 71.1 | |
| ADAPTBackbone=LLaVA-v1.5-13B2026.06 | 4.2 | 17.7 | 1.4 | 82.4 | |
| ADAPT-TFBackbone=LLaVA-v1.5-7B2026.06 | 4 | 16.7 | 1.1 | 73.5 | |
| ADAPT-TFBackbone=LLaVA-v1.5-13B2026.06 | 4 | 18 | 1.5 | 82.4 | |
| DPOBackbone=Qwen2.5-VL-7B2026.06 | 3.9 | 17.3 | 0.6 | 85.8 | |
| ADAPTBackbone=LLaVA-v1.5-7B2026.06 | 3.8 | 15.2 | 1.2 | 79.6 | |
| ADAPTBackbone=Qwen2.5-VL-7B2026.06 | 3.5 | 13.1 | 0.7 | 88.1 | |
| ADAPTBackbone=Qwen2.5-VL-3B2026.06 | 3.1 | 13.5 | 0.5 | 88.3 | |
| RLHF-VBackbone=LLaVA-v1.5-13B2026.06 | — | — | — | 72.6 |