Multi-modal Hallucination Evaluation on AMBER
9.2CHAIRZero-shot
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Zero-shotModel=LLaVA-1.5-7B, Strategy=Zero-shot2026.05 | 9.2 | — | 38.8 | — | — | — | — | — | — | — | — | |
| DoLaModel=LLaVA-v1.52026.05 | 7.6 | — | 36 | 51.6 | 4 | — | — | — | — | — | — | |
| BaselineModel=LLaVA-v1.52026.05 | 7.3 | — | 33.7 | 50.7 | 3.8 | — | — | — | — | — | — | |
| OPERAModel=LLaVA-v1.52026.05 | 7.3 | — | 32 | 49.6 | 3.5 | — | — | — | — | — | — | |
| VCDModel=LLaVA-1.5-7B, Strategy=VCD2026.05 | 7.2 | — | 33.33 | — | — | — | — | — | — | — | — | |
| DMASModel=LLaVA-1.5-7B, Strategy=DMAS2026.05 | 7.1 | — | 33.5 | — | — | — | — | — | — | — | — | |
| VISTAModel=LLaVA-1.5-7B, Strategy=VISTA2026.05 | 6.9 | — | 31.94 | — | — | — | — | — | — | — | — | |
| VCDModel=LLaVA-v1.52026.05 | 6.8 | — | 30.4 | 49.6 | 3.5 | — | — | — | — | — | — | |
| Training-Free Debiasing Inference StrategyModel=LLaVA-1.5-7B, Strategy=Ours2026.05 | 6.74 | — | 29.87 | — | — | — | — | — | — | — | — | |
| MaskCDModel=LLaVA-v1.52026.05 | 6.6 | — | 31.6 | 52.4 | 3 | — | — | — | — | — | — | |
| SIDModel=LLaVA-1.5-7B, Strategy=SID2026.05 | 6.24 | — | 33.73 | — | — | — | — | — | — | — | — | |
| SIDModel=LLaVA-v1.52026.05 | 6.1 | — | 27.5 | 50.2 | 2.8 | — | — | — | — | — | — | |
| DoLaModel=Qwen2.5-VL2026.05 | 5.5 | — | 25.1 | 46.8 | 2.2 | — | — | — | — | — | — | |
| BaselineModel=Qwen2.5-VL2026.05 | 5.3 | — | 23.6 | 47.2 | 1.9 | — | — | — | — | — | — | |
| ICTModel=LLaVA-v1.52026.05 | 5.2 | — | 23.1 | 51.3 | 2.1 | — | — | — | — | — | — | |
| Zero-shotModel=Qwen2.5-VL-7B, Strategy=Zero-shot2026.05 | 5.14 | — | 28.62 | — | — | — | — | — | — | — | — | |
| OPERAModel=Qwen2.5-VL2026.05 | 5.1 | — | 21.6 | 46.2 | 1.7 | — | — | — | — | — | — | |
| SIDModel=Qwen2.5-VL2026.05 | 4.9 | — | 21.5 | 46.3 | 1.4 | — | — | — | — | — | — | |
| ICTModel=Qwen2.5-VL2026.05 | 4.9 | — | 21.8 | 45.1 | 1.2 | — | — | — | — | — | — | |
| MaskCDModel=Qwen2.5-VL2026.05 | 4.8 | — | 22 | 47.4 | 1.5 | — | — | — | — | — | — | |
| SIRAModel=LLaVA-v1.52026.05 | 4.8 | — | 21.6 | 53.7 | 1.8 | — | — | — | — | — | — | |
| VCDModel=Qwen2.5-VL-7B, Strategy=VCD2026.05 | 4.72 | — | 26.35 | — | — | — | — | — | — | — | — | |
| VCDModel=Qwen2.5-VL2026.05 | 4.7 | — | 20.5 | 45.9 | 1.6 | — | — | — | — | — | — | |
| SIRAModel=Qwen2.5-VL2026.05 | 4.6 | — | 20.2 | 47.8 | 0.9 | — | — | — | — | — | — | |
| SIDModel=Qwen2.5-VL-7B, Strategy=SID2026.05 | 4.58 | — | 25.9 | — | — | — | — | — | — | — | — | |
| VISTAModel=Qwen2.5-VL-7B, Strategy=VISTA2026.05 | 4.46 | — | 25.79 | — | — | — | — | — | — | — | — | |
| DMASModel=Qwen2.5-VL-7B, Strategy=DMAS2026.05 | 4.4 | — | 25.6 | — | — | — | — | — | — | — | — | |
| Training-Free Debiasing Inference StrategyModel=Qwen2.5-VL-7B, Strategy=Ours2026.05 | 4.23 | — | 25.07 | — | — | — | — | — | — | — | — | |
| (1)Adaptation strategy=(1)2025.10 | — | 76.9 | — | — | — | — | — | — | — | — | — | |
| (2)Adaptation strategy=Model parameter adaptation2025.10 | — | 72.8 | — | — | — | — | — | — | — | — | — | |
| Baselinetest-time scaling=none2025.10 | — | 68.7 | — | — | — | — | — | — | — | — | — | |
| Baseline2025.10 | — | 68.7 | — | — | — | — | — | — | — | — | — | |
| CAPLBackbone=Qwen2.5-VL2026.03 | — | 85.25 | — | — | — | — | — | — | — | — | — | |
| CAPLBackbone=InternVL2.52026.03 | — | 89.79 | — | — | — | — | — | — | — | — | — | |
| CAPLBackbone=GLM4.1VBase2026.03 | — | 88.49 | — | — | — | — | — | — | — | — | — | |
| DPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 76.92 | 86.11 | 78.71 | — | — | — | |
| DPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 78.16 | 92.47 | 76.2 | — | — | — | |
| DPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 87.98 | 92.37 | 82.85 | — | — | — | |
| DPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 86.01 | 94.24 | 80.99 | — | — | — | |
| GLM4.1VBase2026.03 | — | 89.2 | — | — | — | — | — | — | — | — | — | |
| IC-VCOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 81.81 | 90.48 | 75.56 | — | — | — | |
| IC-VCOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 82.24 | 92.73 | 70.47 | — | — | — | |
| IC-VCOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 88.06 | 95.98 | 73.2 | — | — | — | |
| IC-VCOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 88 | 97.12 | 73.96 | — | — | — | |
| Idefics22026.03 | — | 86.4 | — | — | — | — | — | — | — | — | — | |
| Idefics32026.03 | — | 84.09 | — | — | — | — | — | — | — | — | — | |
| InternVL22026.03 | — | 86 | — | — | — | — | — | — | — | — | — | |
| InternVL2.52026.03 | — | 88.99 | — | — | — | — | — | — | — | — | — | |
| LLaVA-1.5-7BVision Encoder=vanilla CLIP2026.06 | — | — | — | — | — | 75.8 | — | 68.4 | 73.9 | 78.2 | 81.1 | |
| LLaVA-1.5-7BVision Encoder=MACCO-CLIP2026.06 | — | — | — | — | — | 76.5 | — | 69.3 | 74.8 | 78.2 | 81.7 | |
| LLaVA-Next2026.03 | — | 84.96 | — | — | — | — | — | — | — | — | — | |
| LLaVA-NeXT-Interleave-Qwen-7BBackbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 79.97 | 89.03 | 74.51 | — | — | — | |
| LLaVA-OneVision-Qwen2-7BBackbone=LLaVA-OneVision-Qwen2-7B2026.05 | — | — | — | — | — | 84.05 | 91.67 | 75.98 | — | — | — | |
| LLaVA-OV2026.03 | — | 83.99 | — | — | — | — | — | — | — | — | — | |
| mDPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 80.27 | 88.36 | 77.16 | — | — | — | |
| mDPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 80.31 | 92.55 | 74.64 | — | — | — | |
| mDPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 87.12 | 95.68 | 74.8 | — | — | — | |
| mDPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 86.11 | 96.28 | 76.39 | — | — | — | |
| Method ①test-time scaling=Other method 12025.10 | — | 70.4 | — | — | — | — | — | — | — | — | — | |
| Method ②test-time scaling=Other method 22025.10 | — | 64.5 | — | — | — | — | — | — | — | — | — | |
| Method ③test-time scaling=Other method 32025.10 | — | 53.5 | — | — | — | — | — | — | — | — | — | |
| Method ④test-time scaling=Other method 42025.10 | — | 67.8 | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-VL2026.03 | — | 85.13 | — | — | — | — | — | — | — | — | — | |
| Qwen2VL2026.03 | — | 85.96 | — | — | — | — | — | — | — | — | — | |
| S-VCOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 77.88 | 86.17 | 79.21 | — | — | — | |
| S-VCOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 79.72 | 91.41 | 79.15 | — | — | — | |
| S-VCOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 88.04 | 92.94 | 81.71 | — | — | — | |
| S-VCOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 86.36 | 95.36 | 79.1 | — | — | — | |
| SymMPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 80.41 | 88.63 | 76.54 | — | — | — | |
| SymMPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 80.39 | 92.89 | 74.52 | — | — | — | |
| SymMPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 87.21 | 95.82 | 75.46 | — | — | — | |
| SymMPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 86.06 | 96.24 | 76.03 | — | — | — | |
| TTAugtest-time scaling=Method 52025.10 | — | 75.4 | — | — | — | — | — | — | — | — | — | |
| TTAugAdaptation strategy=Test-time Augmentation2025.10 | — | 75.9 | — | — | — | — | — | — | — | — | — | |
| V-DPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 76.48 | 86.21 | 78.47 | — | — | — | |
| V-DPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | — | — | — | — | — | 77.77 | 91.98 | 76.5 | — | — | — | |
| V-DPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Synthetic2026.05 | — | — | — | — | — | 88.01 | 92.02 | 83.69 | — | — | — | |
| V-DPOBackbone=LLaVA-OneVision-Qwen2-7B, Contrastive Sample Source=Contrastive Sample Editing2026.05 | — | — | — | — | — | 86.42 | 94.65 | 81.56 | — | — | — |