List your company
ResearchFormerly WizWand

Visual Question Answering on OK-VQA (Accuracy)

84.7AccuracyVPD (55B)
52.87661.13869.477.662Aug 24, 2023Feb 9, 2024Jul 27, 2024Jan 12, 2025Jun 30, 2025Dec 16, 2025Jun 3, 2026
Updated 4mo ago

Evaluation results

MethodLinks
VPD (55B)LLM=PaLI2024.0284.7
CgRAGLLM/MLLM=InternVL3-8B2026.0577.8
KU-RAGLLM/MLLM=GPT-4o2026.0577.2
ProphetLLM=mPLUG2024.0276.6
Hyper-ICLBackbone=LLaVA-OneVision (Qwen2-7B)2026.0675.32
PromptcapLLM=GPT-32024.0273.2
HiFICLBackbone=LLaVA-OneVision (Qwen2-7B)2026.0373.12
Qwen2.5-VL-7BLLM/MLLM=-2026.0572.4
MimICBackbone=LLaVA-OneVision (Qwen2-7B)2026.0369.43
MimICBackbone=LLaVA-OneVision (Qwen2-7B)2026.0669.43
DeepMMSearch-R1-7BKnowledge Source=Google Search Google Lens, Evaluation Protocol=LLM-as-Judge2025.0867.8
HiFICLBackbone=LLaVA-v1.6 (Mistral-7B)2026.0366.88
Single Task SoTA2023.1266.8
8-shot ICLBackbone=LLaVA-OneVision (Qwen2-7B)2026.0366.59
8-shot ICLBackbone=LLaVA-OneVision (Qwen2-7B)2026.0666.59
Uni-MoE w/ MoE-Task4 (8E) + Aux lossModality=I,T,S,V, Training Task Strategy=MoE-Task4, Number of Experts=8, Auxiliary Balancing Loss=true2024.0566.4
Uni-MoE w/ MoE-Task4 (8E)Modality=I,T,S,V, Training Task Strategy=MoE-Task4, Number of Experts=82024.0566.2
Uni-MoE w/ MoE-Task4 (4E) + Aux lossModality=I,T,S,V, Training Task Strategy=MoE-Task4, Number of Experts=4, Auxiliary Balancing Loss=true2024.0566.13
PALI-X-55Bmode=specialist2023.0866.1
Uni-MoE w/ MoE-Task4 (4E)Modality=I,T,S,V, Training Task Strategy=MoE-Task4, Number of Experts=42024.0566.02
PMSR (Qwen3-VL-8B)Knowledge Source=Wikipedia, Backbone=Qwen3-VL-8B, Evaluation Protocol=LLM-as-Judge2025.0866
Emu2-Chattrained on task training set=true2023.1264.8
CogVLM# Param=17B2024.0864.8
MimICBackbone=LLaVA-v1.6 (Mistral-7B)2026.0364.62
PaLI-X-VPDModel Scale=55B2023.1264.6
PaLI-17B#Params=17B, Model Type=Specialist, Evaluation Protocol=Fine-tuned2023.1164.5
PaLI-X-InstructModel Scale=55B2023.1264.3
Raw dataMLLM=GPT-4V2026.0363.6
Zero-shotBackbone=LLaVA-v1.6 (Mistral-7B)2026.0363
Single-Modality-Expert-Task2Modality=I,T, Training Task Strategy=MoE-Task22024.0562.91
AdaMMSunsupervised=true2025.0362.11
MoE-LLaVA-2.7B×4-Top2Model Scale=2.7B, Number of Experts=4, Routing=Top22024.0562.1
Uni-MoE w/ MoE-Task2 (4E)Modality=I,T,S,V, Training Task Strategy=MoE-Task2, Number of Experts=42024.0562.1
LLaVA-1.5#Params=7.2B, Model Type=Generalist2023.1162
Qwen-VL-7BShots=42024.0662
Qwen-VL-7B + MTVProtocol=Multimodal Task Vectors2024.0662
Qwen2.5-VL-3BLLM/MLLM=-2026.0562
Uni-MoE w/ MoE-Task3 (4E)Modality=I,T,A,V, Training Task Strategy=MoE-Task3, Number of Experts=42024.0561.96
LLaVA-1.5-13BModality=I,T, Model Scale=13B2024.0561.93
Qwen-VL-7BShots=82024.0661.5
InfMLLM-7BLLM=Vicuna-7B2023.1161.33
DARE-Linearunsupervised=false2025.0361.07
InstructBlip 7BPre-training sample size=129M, Instruction tuning sample size=1.2M, Trained on other VQA datasets=true2023.1061
mPLUG-Owl2(base)2025.0360.98
Task Arithmeticunsupervised=false2025.0360.93
CogVLM2025.0360.82
mPLUG-Owl3# Param=8B2024.0860.1
MoE-LLaVA-1.6B×4-Top2Model Scale=1.6B, Number of Experts=4, Routing=Top22024.0559.9
MMSearch-R1-7BKnowledge Source=Google Search Google Lens, Evaluation Protocol=LLM-as-Judge2025.0859.9
augmented Tchebycheff objectiveMLLM=GPT-4V2026.0359.2
CVLMLLM=Qwen-VL2024.0258.91
CogVLMtrained on task training set=true2023.1258.9
Qwen-VLzero-shot=true, backbone=Qwen-7B2023.0858.6
QWen-VL-7BLLM=Qwen-7B2023.1158.6
Qwen-VLModel Size=9.6B2023.1258.6
Qwen-VLModel Scale=9.7B2023.1258.6
Qwen-VL-7BShots=02024.0658.6
Qwen-VL-7B‡Modality=I,T, Model Scale=7B2024.0558.6
LLaVA-665KRecipe=LLaVA-665K2025.0457.96
FlamingoModel Scale=80B, Zero-shot=true2023.1257.8
Single-Modality-Expert-Task6Modality=I,T,S, Training Task Strategy=MoE-Task62024.0557.77
mPLUG-Owl2#Params=8.2B, Model Type=Generalist2023.1157.7
mPLUG-Owl2Model Size=8.2B2023.1257.7
mPLUG-Owl2Model Scale=8.2B2023.1257.7
mPLUG-Owl2# Param=8B2024.0857.7
Uni-MoE w/ MoE-Task1 (4E)Modality=I,T,S, Training Task Strategy=MoE-Task1, Number of Experts=42024.0557.63
PaLI-3-VPDModel Scale=5B2023.1257.5
InstructBLIPLLM=Vicuna-7B2024.0257.36
CVLM (3M IKPairs, Objects=3)LLM=Vicuna-7B, Objects=32024.0257.17
Qwen-VLLLM=Qwen2024.0257.13
CVLM (3M IKPairs, Objects=5)LLM=Vicuna-7B, Objects=52024.0256.92
MiniGPT-v2-7BLLM=LLaMA2-7B2023.1156.9
CVLM (3M IKPairs, Objects=1)LLM=Vicuna-7B, Objects=12024.0256.9
Hyper-ICLBackbone=LLaVA-Interleave-7B2026.0656.77
CVLM (3M IKPairs, Objects=8)LLM=Vicuna-7B, Objects=82024.0256.71
PaLI-3-InstructModel Scale=5B2023.1256.7
Qwen-VL-Chatzero-shot=true2023.0856.6
Qwen-VL-Chat#Params=9.6B, Model Type=Generalist2023.1156.6
Qwen-VL-ChatModel Scale=9.7B2023.1256.6
Qwen-VL-13B-Chattrained on task training set=true2023.1256.6
Qwen-VL-Chat# Param=9B2024.0856.6
Surrogate model of SPActMLLM=GPT-4V2026.0356.5
DownsampleMLLM=GPT-4V2026.0356.2
Obf-WeakBlurMLLM=GPT-4V2026.0356.1
Single-Modality-Expert-Task5Modality=I,T,S, Training Task Strategy=MoE-Task52024.0556.01
8-shot ICLBackbone=LLaVA-v1.6 (Mistral-7B)2026.0356
SelectivePrivacyMLLM=GPT-4V2026.0355.9
CVLM (3M IKPairs) w/o FKALLM=Vicuna-7B2024.0255.7
Surrogate model of MPPARMLLM=GPT-4V2026.0355.6
PaLM-E-12B#Params=12B, Model Type=Generalist2023.1155.5
UIO-2XXLModel Size=6.8B2023.1255.5
H-GIVRModel=Qwen2.5vl:7b, Prompt=Ours2026.0255.43
Mantis-SigLIP# Param=8B2024.0855.4
Surrogate model of VITAMLLM=GPT-4V2026.0355.4
Obf-StrongBlurMLLM=GPT-4V2026.0355.3
LLaVA-1.5-7BModality=I,T, Model Scale=7B2024.0555.09
LLaVAPre-training sample size=558K, Instruction tuning sample size=158K2023.1054.4
CVLMLLM=Vicuna-7B2024.0254.3
Ties-Mergingunsupervised=false2025.0354.15
Self-ConModel=Qwen2.5vl:7b, Prompt=Self-Con2026.0254.1