Object Hallucination Evaluation on POPE Adversarial
89.8AccuracySFT + VIGIL
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| SFT + VIGILBase Model=Qwen2.5-VL-72B2026.06 | 89.8 | — | — | — | — | — | — | — | — | |
| LogicCheckGPTBase Model=LLaVA-1.52026.04 | 89.33 | — | — | — | 89.4 | — | — | — | — | |
| SelfCheckBase Model=LLaVA-1.52026.04 | 88.67 | — | — | — | 88.27 | — | — | — | — | |
| R-CoVBase Model=LLaVA-1.52026.04 | 88.67 | — | — | — | 88.36 | — | — | — | — | |
| ICONS2025.02 | 87.5 | — | — | — | — | — | — | — | — | |
| SFT + DA-DPOBase Model=Qwen2.5-VL-72B2026.06 | 87.4 | — | — | — | — | — | — | — | — | |
| R-CoVBase Model=Qwen2.5-VL2026.04 | 87.33 | — | — | — | 86.13 | — | — | — | — | |
| LogicCheckGPTBase Model=Qwen2.5-VL2026.04 | 87 | — | — | — | 85.5 | — | — | — | — | |
| CRG(ours)Model=Qwen2.5-VL-7B-Instruct2026.05 | 86.98 | — | — | — | 87.07 | — | — | — | — | |
| SFT + VIGILBase Model=Qwen2.5-VL-7B2026.06 | 86.9 | — | — | — | — | — | — | — | — | |
| SFT + HA-DPOBase Model=Qwen2.5-VL-72B2026.06 | 86.2 | — | — | — | — | — | — | — | — | |
| COINCIDE2025.02 | 86.1 | — | — | — | — | — | — | — | — | |
| SpecEyesBase Model=Thyme, Aggregation Strategy=mean2026.03 | 85.89 | — | — | — | — | 1.78 | — | — | — | |
| SpecEyesBase Model=Thyme, Aggregation Strategy=bottom2026.03 | 85.89 | — | — | — | — | 1.81 | — | — | — | |
| SpecEyesBase Model=Thyme, Aggregation Strategy=min2026.03 | 85.87 | — | — | — | — | 1.77 | — | — | — | |
| VTIModel=Qwen2.5-VL-7B-Instruct2026.05 | 85.78 | — | — | — | 85.14 | — | — | — | — | |
| SpecEyesBase Model=Thyme, Aggregation Strategy=log2026.03 | 85.76 | — | — | — | — | 1.68 | — | — | — | |
| ACEBase Model=LLaVA-NeXT2026.05 | 85.68 | — | — | — | 82.89 | — | — | — | — | |
| VanillaBase Model=Qwen2.5-VL2026.04 | 85.67 | — | — | — | 83.65 | — | — | — | — | |
| EL2N2025.02 | 85.6 | — | — | — | — | — | — | — | — | |
| TIVE2025.02 | 85.6 | — | — | — | — | — | — | — | — | |
| PRISM2025.02 | 85.5 | — | — | — | — | — | — | — | — | |
| LUREBase Model=LLaVA-1.52026.04 | 85.33 | — | — | — | 86.25 | — | — | — | — | |
| Qwen2.5-VL-7BModel size=7B2026.05 | 85.33 | — | — | — | — | — | — | — | — | |
| DataTailor2025.02 | 85.3 | — | — | — | — | — | — | — | — | |
| SpecEyesBase Model=DeepEyes, Aggregation Strategy=mean2026.03 | 85.13 | — | — | — | — | 2.06 | — | — | — | |
| SpecEyesBase Model=DeepEyes, Aggregation Strategy=bottom2026.03 | 85.13 | — | — | — | — | 2.08 | — | — | — | |
| SpecEyesBase Model=DeepEyes, Aggregation Strategy=min2026.03 | 85.13 | — | — | — | — | 2.13 | — | — | — | |
| PAIModel=Qwen2.5-VL-7B-Instruct2026.05 | 85.12 | — | — | — | 84.77 | — | — | — | — | |
| OPERAModel=Qwen2.5-VL-7B-Instruct2026.05 | 84.93 | — | — | — | 85.17 | — | — | — | — | |
| VGAModel=Qwen2.5-VL2025.11 | 84.86 | — | — | — | 85.09 | — | — | — | — | |
| VanillaModel=Qwen2.5-VL2025.11 | 84.82 | — | — | — | 84.57 | — | — | — | — | |
| VAFModel=Qwen2.5-VL2025.11 | 84.8 | — | — | — | 84.74 | — | — | — | — | |
| VISTABase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 84.8 | — | — | — | 83.39 | — | — | — | — | |
| SFT + SimPOBase Model=Qwen2.5-VL-72B2026.06 | 84.8 | — | — | — | — | — | — | — | — | |
| VCDModel=Qwen2.5-VL-7B-Instruct2026.05 | 84.78 | — | — | — | 84.63 | — | — | — | — | |
| TARACModel=Qwen2.5-VL2025.11 | 84.72 | — | — | — | 84.47 | — | — | — | — | |
| CRG(ours)Model=Qwen-VL-Chat2026.05 | 84.7 | — | — | — | 83.78 | — | — | — | — | |
| ACEBase Model=LLaVA-1.52026.05 | 84.65 | — | — | — | 82.59 | — | — | — | — | |
| SpecReasonBase Model=Thyme2026.03 | 84.62 | — | — | — | — | 0.46 | — | — | — | |
| PAIModel=Qwen2.5-VL2025.11 | 84.6 | — | — | — | 84.82 | — | — | — | — | |
| ESCModel=LLaVA [50]2026.07 | 84.53 | — | — | — | 84.15 | — | — | — | — | |
| SFT + DPOBase Model=Qwen2.5-VL-72B2026.06 | 84.5 | — | — | — | — | — | — | — | — | |
| GAP (LH+PCA+DA)Latent Head=true, PCA=true, Data Augmentation=true2026.05 | 84.49 | — | — | — | — | — | — | — | — | |
| CRG(ours)Model=LLaVA-1.5-7B2026.05 | 84.43 | — | — | — | 83.77 | — | — | — | — | |
| LH+DA (no PCA)Latent Head=true, Data Augmentation=true, PCA=false2026.05 | 84.36 | — | — | — | — | — | — | — | — | |
| CASTBackbone=LLaVA-1.5-7B2026.05 | 84.27 | — | — | — | — | — | — | — | — | |
| Full-Finetune2025.02 | 84.2 | — | — | — | — | — | — | — | — | |
| SFT + DA-DPOBase Model=Qwen2.5-VL-7B2026.06 | 84.2 | — | — | — | — | — | — | — | — | |
| Length2025.02 | 84.1 | — | — | — | — | — | — | — | — | |
| SIDBase Model=LLaVA-NeXT2026.05 | 84.08 | — | — | — | 82.92 | — | — | — | — | |
| PAIBase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 84.03 | — | — | — | 83.23 | — | — | — | — | |
| SpecEyesBase Model=DeepEyes, Aggregation Strategy=log2026.03 | 83.97 | — | — | — | — | 1.89 | — | — | — | |
| VCDBase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 83.86 | — | — | — | 82.89 | — | — | — | — | |
| Self-Filter2025.02 | 83.8 | — | — | — | — | — | — | — | — | |
| PTIBase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 83.77 | — | — | — | 83.63 | — | — | — | — | |
| R-CoVBase Model=mPLUG-Owl2026.04 | 83.67 | — | — | — | 83.93 | — | — | — | — | |
| Random2025.02 | 83.6 | — | — | — | — | — | — | — | — | |
| BaselineModel=LLaVA [50]2026.07 | 83.47 | — | — | — | 82.57 | — | — | — | — | |
| VTIBase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 83.46 | — | — | — | 83.48 | — | — | — | — | |
| PTIBase Model=Qwen-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 83.37 | — | — | — | 82.41 | — | — | — | — | |
| R-CoVBase Model=MiniGPT-42026.04 | 83.33 | — | — | — | 82.64 | — | — | — | — | |
| VanillaBase Model=LLaVA-1.52026.04 | 83.33 | — | — | — | 84.84 | — | — | — | — | |
| Perplexity2025.02 | 83.3 | — | — | — | — | — | — | — | — | |
| HALVABackbone=LLaVA-1.5-7B2026.05 | 83.2 | — | — | — | — | — | — | — | — | |
| PAIModel=LLaVA-1.5-7B2026.05 | 83.17 | — | — | — | 81.67 | — | — | — | — | |
| VAFModel=LLaVA-Next2025.11 | 83.14 | — | — | — | 83 | — | — | — | — | |
| VTIModel=Qwen-VL-Chat2026.05 | 83.13 | — | — | — | 82.16 | — | — | — | — | |
| VCDBase Model=Qwen-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 83.1 | — | — | — | 82.04 | — | — | — | — | |
| VCDModel=Qwen-VL-Chat2026.05 | 83.1 | — | — | — | 82.04 | — | — | — | — | |
| SFT + SimPOBase Model=Qwen2.5-VL-7B2026.06 | 83.1 | — | — | — | — | — | — | — | — | |
| VGAModel=LLaVA-Next2025.11 | 83.06 | — | — | — | 83.22 | — | — | — | — | |
| SIDBase Model=LLaVA-1.52026.05 | 83.05 | — | — | — | 82.32 | — | — | — | — | |
| HADPOBackbone=LLaVA-1.5-7B2026.05 | 82.9 | — | — | — | — | — | — | — | — | |
| SFT + DPOBase Model=Qwen2.5-VL-7B2026.06 | 82.8 | — | — | — | — | — | — | — | — | |
| RegularModel=Qwen2.5-VL-7B-Instruct2026.05 | 82.79 | — | — | — | 83.15 | — | — | — | — | |
| VanillaBase Model=DeepSeek-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 82.73 | — | — | — | 81.97 | — | — | — | — | |
| OPERAModel=Qwen-VL-Chat2026.05 | 82.67 | — | — | — | 79.89 | — | — | — | — | |
| PAIModel=LLaVA-Next2025.11 | 82.64 | — | — | — | 81.82 | — | — | — | — | |
| VTIModel=LLaVA-1.5-7B2026.05 | 82.57 | — | — | — | 82.11 | — | — | — | — | |
| Qwen3-VL-2BMode=draft only2026.03 | 82.56 | — | — | — | — | 4.2 | — | — | — | |
| GraNd2025.02 | 82.5 | — | — | — | — | — | — | — | — | |
| VanillaModel=LLaVA-Next2025.11 | 82.31 | — | — | — | 81.4 | — | — | — | — | |
| PAIModel=Qwen-VL-Chat2026.05 | 82.19 | — | — | — | 82.06 | — | — | — | — | |
| SFT onlyBase Model=Qwen2.5-VL-72B2026.06 | 82.1 | — | — | — | — | — | — | — | — | |
| OPERABase Model=LLaVA-NeXT2026.05 | 81.94 | — | — | — | 81.89 | — | — | — | — | |
| PAIBase Model=Qwen-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 81.63 | — | — | — | 81.12 | — | — | — | — | |
| ThymeBase Model=Thyme2026.03 | 81.32 | — | — | — | — | 1 | — | — | — | |
| OPERAModel=LLaVA-1.5-7B2026.05 | 81.24 | — | — | — | 81.38 | — | — | — | — | |
| VCDBase Model=LLaVA-NeXT2026.05 | 81.09 | — | — | — | 81.03 | — | — | — | — | |
| TARACModel=LLaVA-Next2025.11 | 81.04 | — | — | — | 79.73 | — | — | — | — | |
| RegularModel=Qwen-VL-Chat2026.05 | 81.03 | — | — | — | 79.3 | — | — | — | — | |
| LogicCheckGPTBase Model=mPLUG-Owl2026.04 | 81 | — | — | — | 81.19 | — | — | — | — | |
| OPERABase Model=LLaVA-1.52026.05 | 80.98 | — | — | — | 81.35 | — | — | — | — | |
| VCDModel=LLaVA-1.5-7B2026.05 | 80.88 | — | — | — | 81.33 | — | — | — | — | |
| VISTABase Model=Qwen-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 80.8 | — | — | — | 79.46 | — | — | — | — | |
| LogicCheckGPTBase Model=MiniGPT-42026.04 | 80.67 | — | — | — | 78.99 | — | — | — | — | |
| SFT onlyBase Model=Qwen2.5-VL-7B2026.06 | 80.5 | — | — | — | — | — | — | — | — | |
| ESCModel=Qwen2 [92]2026.07 | 80.43 | — | — | — | 76.17 | — | — | — | — | |
| VanillaBase Model=Qwen-VL-Chat, Decoding Strategy=nucleus sampling, Max New Tokens=322026.04 | 80.26 | — | — | — | 79.76 | — | — | — | — |