Object Hallucination Evaluation on MS-COCO (POPE Adversarial)
87.83AccuracySIRA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SIRABackbone=Qwen2.5-VL2026.05 | 87.83 | 87.39 | — | — | |
| AVISCModel=Qwen2.5-VL-7B2024.05 | 87.62 | 86.93 | — | — | |
| M3IDModel=Qwen2.5-VL-7B2024.05 | 87.3 | 84.96 | — | — | |
| PNDModel=Qwen3VL-2B2026.04 | 87.26 | 87.13 | 88.02 | 86.26 | |
| M3IDModel=LLaVA-OneVision (Qwen2-7B)2024.05 | 86.8 | 84.86 | — | — | |
| regularModel=Qwen3VL-2B2026.04 | 86.53 | 86.57 | 86.29 | 86.86 | |
| VCDModel=Qwen2.5-VL-7B2024.05 | 86.5 | 85.52 | — | — | |
| ICTBackbone=Qwen2.5-VL2026.05 | 86.16 | 84.32 | — | — | |
| AVISCModel=LLaVA-OneVision (Qwen2-7B)2024.05 | 86.12 | 85.83 | — | — | |
| MaskCDBackbone=Qwen2.5-VL2026.05 | 86.05 | 86.15 | — | — | |
| VCDModel=LLaVA-OneVision (Qwen2-7B)2024.05 | 86 | 85.42 | — | — | |
| ScalpelBase Model=LLaVA-1.5-7B2026.02 | 85.97 | 86 | — | — | |
| baseBase Model=Intern-VL22025.09 | 85.8 | 85 | — | — | |
| GACDBase Model=Intern-VL22025.09 | 85.8 | 85 | — | — | |
| WoodpeckerBase Model=Intern-VL22025.09 | 85.7 | 84.8 | — | — | |
| SchroMindModel=LLaVA-1.5-7B2026.02 | 85.43 | 84.48 | — | — | |
| AVISCBase Model=Intern-VL22025.09 | 85.3 | 84.6 | — | — | |
| CAIBackbone=Qwen-VL2026.06 | 84.97 | 84.69 | — | — | |
| PNDModel=LLAVA1.5-13B2026.04 | 84.96 | 83.59 | 91.99 | 76.6 | |
| VDD-NoneBackbone=Qwen-VL-Chat, Decoding Strategy=VDD-None2026.05 | 84.83 | 83.87 | — | — | |
| BRACSBackbone=Qwen-VL-Chat, Decoding Strategy=BRACS2026.05 | 84.8 | 83.86 | — | — | |
| OPERABackbone=Qwen2.5-VL2026.05 | 84.78 | 83.45 | — | — | |
| MESAModel=Qwen-VL2026.04 | 84.47 | 84.1 | 88.81 | 79.87 | |
| MESAModel=Qwen-VL2026.04 | 84.47 | 84.1 | — | — | |
| CoFi-DecModel=Qwen-VL2025.12 | 84.36 | 83.69 | 85.03 | — | |
| HDDModel=LLaVA-NeXT2025.12 | 84.3 | 84.5 | — | — | |
| ICTModel=Qwen-VL2026.04 | 84.3 | 83.51 | 87.91 | 79.53 | |
| ICTModel=Qwen-VL2026.04 | 84.3 | 83.51 | — | — | |
| VCDModel=Qwen-VL2026.04 | 84.26 | 83.9 | 85.84 | 82.05 | |
| VCDModel=Qwen-VL2026.04 | 84.26 | 83.9 | — | — | |
| VCDBackbone=Qwen2.5-VL2026.05 | 84.26 | 83.9 | — | — | |
| ONLYBackbone=Qwen-VL2026.06 | 84.23 | 83.67 | — | — | |
| GACDBase Model=mPLUG-Owl22025.09 | 84.2 | 83.7 | — | — | |
| mPLUG-Owl2 + GACDBase Model=mPLUG-Owl2, Setting=Adversarial, Gradient Method=Direct Gradient (Ours)2025.09 | 84.2 | 83.7 | — | — | |
| VCDBackbone=Qwen-VL2026.06 | 84.2 | 84.07 | — | — | |
| NulluModel=Qwen-VL2026.04 | 84.11 | 82.09 | 86.63 | 78 | |
| NulluModel=Qwen-VL2026.04 | 84.11 | 82.09 | — | — | |
| baseModel=Qwen2.5-VL-7B2024.05 | 84.07 | 82.73 | — | — | |
| SIRABackbone=LLaVA-v1.52026.05 | 84.07 | 83.41 | — | — | |
| PNDModel=LLaVA1.5-7B2026.04 | 84.03 | 83.48 | 90.55 | 77.43 | |
| VCDModel=Qwen-VL2025.05 | 84 | 82 | 90.6 | 74.9 | |
| MoDModel=Qwen-VL2025.05 | 84 | 82.3 | 92.4 | 74.2 | |
| SIDModel=LLaVA-NeXT2025.12 | 84 | 83 | — | — | |
| PAIBackbone=Qwen-VL-Chat, Decoding Strategy=PAI2026.05 | 84 | 83.15 | — | — | |
| VCDModel=Qwen-VL, Decoding=VCD2026.01 | 83.9 | 82.58 | 89.95 | 76.33 | |
| VAFModel=LLaVA-v1.52026.04 | 83.86 | 82.71 | 89.07 | 77.2 | |
| VAFModel=LLaVA-v1.52026.04 | 83.86 | 82.71 | — | — | |
| ICTBase Model=LLaVA-1.5-7B2026.02 | 83.83 | 83.84 | — | — | |
| ICTModel=LLaVA-1.5-7B2026.02 | 83.83 | 83.84 | — | — | |
| ICTModel=LLaVA-v1.52026.04 | 83.83 | 82.45 | 90.11 | 76 | |
| ICTModel=LLaVA-v1.52026.04 | 83.83 | 82.45 | — | — | |
| SIDBackbone=Qwen2.5-VL2026.05 | 83.83 | 83.03 | — | — | |
| SPINBackbone=LLaVA-1.5-7B, Decoding Strategy=SPIN2026.05 | 83.83 | 82.46 | — | — | |
| SPINBackbone=Qwen-VL-Chat, Decoding Strategy=SPIN2026.05 | 83.73 | 82.05 | — | — | |
| VAFBase Model=LLaVA-1.5-7B2026.05 | 83.72 | 83.05 | — | — | |
| VAFModel=Qwen-VL2026.04 | 83.7 | 82.09 | 91.06 | 74.73 | |
| VAFModel=Qwen-VL2026.04 | 83.7 | 82.09 | — | — | |
| VCDBackbone=Qwen-VL-Chat, Decoding Strategy=VCD2026.05 | 83.7 | 82.25 | — | — | |
| BRACSBackbone=LLaVA-1.5-7B, Decoding Strategy=BRACS2026.05 | 83.67 | 83.03 | — | — | |
| RVEBase Model=LLaVA-1.5-7B2026.05 | 83.65 | 82.45 | — | — | |
| GreedyBackbone=Qwen-VL-Chat, Decoding Strategy=Greedy2026.05 | 83.63 | 81.93 | — | — | |
| GreedyBackbone=LLaVA-1.5-7B, Decoding Strategy=Greedy2026.05 | 83.57 | 82.07 | — | — | |
| VTIModel=Qwen-VL2026.04 | 83.53 | 82.75 | 86.88 | 79 | |
| VTIModel=Qwen-VL2026.04 | 83.53 | 82.75 | — | — | |
| GACDBase Model=LLaVA-v1.52025.09 | 83.5 | 82.1 | — | — | |
| M3IDBase Model=Intern-VL22025.09 | 83.5 | 82.1 | — | — | |
| DeGFModel=Qwen-VL2025.12 | 83.47 | 82.98 | 84.49 | — | |
| DoLaBackbone=Qwen2.5-VL2026.05 | 83.47 | 81.51 | — | — | |
| VTIModel=LLaVA-v1.52026.04 | 83.43 | 82.54 | 87.23 | 78.33 | |
| VTIModel=LLaVA-v1.52026.04 | 83.43 | 82.54 | — | — | |
| mPLUG-Owl2 + IGBase Model=mPLUG-Owl2, Setting=Adversarial, Gradient Method=Integrated Gradients (SIG implementation)2025.09 | 83.4 | 82.9 | — | — | |
| BaseBackbone=Qwen2.5-VL2026.05 | 83.37 | 81.57 | — | — | |
| PNDModel=InternVL2-2B2026.04 | 83.33 | 82.19 | 88.22 | 76.93 | |
| VCDBase Model=LLaVA-1.5-7B2026.05 | 83.33 | 82.02 | — | — | |
| PAIBackbone=Qwen-VL2026.06 | 83.33 | 81.64 | — | — | |
| HDDModel=LLaVA-1.52025.12 | 83.3 | 83.4 | — | — | |
| ICTBackbone=LLaVA-v1.52026.05 | 83.3 | 82.54 | — | — | |
| MESAModel=LLaVA-v1.52026.04 | 83.23 | 83.56 | 81.98 | 85.2 | |
| MESAModel=LLaVA-v1.52026.04 | 83.23 | 83.56 | — | — | |
| M3IDModel=Qwen-VL2025.05 | 83.2 | 80.8 | 90.4 | 73.1 | |
| GreedyModel=LLaVA-NeXT2025.12 | 83.2 | 81 | — | — | |
| VCDBase Model=Intern-VL22025.09 | 83.2 | 82.2 | — | — | |
| AGLAModel=LLaVA1.5-7B2026.04 | 83.13 | 82.21 | 86.97 | 77.95 | |
| baseModel=LLaVA-OneVision (Qwen2-7B)2024.05 | 83 | 82.4 | — | — | |
| VanillaBase Model=LLaVA-1.5-7B2026.05 | 82.97 | 81.26 | — | — | |
| VCDModel=LLaVA-NeXT2025.12 | 82.9 | 81.5 | — | — | |
| PNDModel=InstructBLIP-13B2026.04 | 82.83 | 82.34 | 84.75 | 80.06 | |
| RITUALModel=Qwen-VL2025.12 | 82.8 | 82.71 | 83.15 | — | |
| CAIBackbone=LLaVA-1.52026.06 | 82.77 | 82.73 | — | — | |
| SIDModel=LLaVA-1.52025.12 | 82.7 | 81.9 | — | — | |
| SDCDModel=Qwen-VL, Decoding=SDCD2026.01 | 82.67 | 81.08 | 89.26 | 74.27 | |
| GACDBase Model=InstructBLIP2025.09 | 82.5 | 82.1 | — | — | |
| PNDModel=QwenVL-7B2026.04 | 82.46 | 79.9 | 93.55 | 69.73 | |
| MoDModel=LLaVA-NEXT2025.05 | 82.4 | 81 | 88.3 | 74.8 | |
| Regular DecodingModel=Qwen-VL, Decoding=Regular2026.01 | 82.4 | 80.59 | 89.84 | 73.07 | |
| HDDModel=InstructBLIP2025.12 | 82.3 | 82.5 | — | — | |
| PNDModel=InstructBLIP-7B2026.04 | 82.2 | 81.98 | 82.99 | 81 | |
| samplingModel=Qwen-VL2025.05 | 82.1 | 80.2 | 90 | 72.3 | |
| M3IDModel=Qwen-VL2025.12 | 82.03 | 82.19 | 81.47 | — | |
| CAIBackbone=InstructBLIP2026.06 | 81.83 | 81.56 | — | — |