List your company
ResearchFormerly WizWand

Visual Question Answering on InfoVQA

89.3AccuracyAutoMoT
54.35663.42872.581.572Dec 20, 2024Mar 22, 2025Jun 22, 2025Sep 22, 2025Dec 23, 2025Mar 25, 2026Jun 25, 2026
Updated 4mo ago

Evaluation results

MethodLinks
AutoMoT2026.0389.3——
VISEModel Scale=32B2026.0688.43——
VisionZero-CLEVRModel Scale=32B2026.0688.32——
VisionZero-ChartModel Scale=32B2026.0688.19——
iReasonerModel Scale=32B2026.0688.12——
EvoLMMModel Scale=32B2026.0688.05——
VisionZero-RWModel Scale=32B2026.0687.94——
VisPlayModel Scale=32B2026.0687.89——
BaseModel Scale=32B2026.0687.77——
Qwen2.5-VL-32B + AT-RL (Ours)Zero-shot=true2026.0287.1——
Penguin-VLModel Size=8B2026.0386.8——
Qwen3-VL 8B-RModel Scale=8B, Strategy=Reasoning2026.0486.7——
Qwen3-VL 8B + 4B-RMain Model=8B, Source Model=4B-R, Strategy=Reasoning Transfer2026.0484.1——
Qwen3-VLModel Size=8B2026.0383.1——
Qwen3-VL 4B-RModel Scale=4B, Strategy=Reasoning2026.0483.1——
VISEModel Scale=8B2026.0682.83——
Qwen3-VL 8B-SModel Scale=8B, Strategy=Simple2026.0482.8——
Qwen3-VL 8B + 2B-RMain Model=8B, Source Model=2B-R, Strategy=Reasoning Transfer2026.0482.8——
VisionZero-CLEVRModel Scale=8B2026.0682.44——
VisionZero-ChartModel Scale=8B2026.0682.31——
Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Token Retention Ratio=100%2025.1282.3——
iReasonerModel Scale=8B2026.0682.15——
EvoLMMModel Scale=8B2026.0682.02——
Direct Fine-tuning (r=1)Optimization steps=+700 steps2026.0281.61——
Direct Fine-tuning (r=1)Optimization steps=+0 steps2026.0281.56——
In-Squeeze (128 ... -> 1)Optimization mode=Standard2026.0281.45——
In-Squeeze (128 ... -> 1)Optimization mode=Min steps2026.0281.45——
VISEModel Scale=4B2026.0681.45——
VisionZero-RWModel Scale=8B2026.0681.44——
Cont-Squeeze (128 -> 1)Optimization steps=+200 steps2026.0281.36——
VisPlayModel Scale=8B2026.0681.36——
BaseModel Scale=8B2026.0681.23——
Direct Fine-tuning (r=1)Optimization steps=+200 steps2026.0281.07——
Cont-Squeeze (128 -> 1)Optimization steps=+700 steps2026.0280.92——
BaselineBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=1.00×, Channel multiplier=1.00×, KV Cache budget multiplier=1.00×2026.0580.12——
Qwen3-VL 4B-SModel Scale=4B, Strategy=Simple2026.0480——
Qwen3-VL 4B + 2B-RMain Model=4B, Source Model=2B-R, Strategy=Reasoning Transfer2026.0479.7——
InternVL-3.5Model Size=8B2026.0379.1——
iReasonerModel Scale=4B2026.0678.76——
VisionZero-ChartModel Scale=4B2026.0678.74——
EvoLMMModel Scale=4B2026.0678.58——
VisionZero-CLEVRModel Scale=4B2026.0678.28——
VisPlayModel Scale=4B2026.0677.96——
VisionZero-RWModel Scale=4B2026.0677.94——
Penguin-VLParameters=2B2026.0377.8——
BaseModel Scale=4B2026.0677.73——
OpenAI GPT-4oZero-shot=true2026.0277——
Qwen2.5-VL-72B InstructZero-shot=true2026.0276.9——
BaseRatio=100%, Backbone=Qwen2.5-VL-3B2026.0575.95——
ReCogDrive2026.0375.8——
Gemini 2.0 FlashZero-shot=true2026.0275.6——
CropVLMReward=Accuracy, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1175.6——
CropVLMReward=LL, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1175.6——
Qwen 2.5 VLInput Resolution=1792x17922025.1175.55——
DivPrune + RandomBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1274.9——
IVC-PruneRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0574.75——
QwenVL2.5-3BModel size=3B2026.0574.22——
BalCapRL-3BBackbone=QwenVL2.5-3B2026.0574.17——
DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1273.9——
ERASERatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0573.34——
CapRL-3BBackbone=QwenVL2.5-3B2026.0573.17——
DivPrune+VTWBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1273——
Claude 3.5 SonnetZero-shot=true2026.0272.8——
IVC-PruneRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0572.58——
Qwen3-VLParameters=2B2026.0372.4——
ERASERatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0571.74——
VISEModel Scale=2B2026.0671.43——
OpenEMMA2026.0371.4——
Qwen3-VL 2B-RModel Scale=2B, Strategy=Reasoning2026.0471.4——
RubiCap-3BBackbone=QwenVL2.5-3B2026.0571.26——
VisionZero-CLEVRModel Scale=2B2026.0670.93——
iReasonerModel Scale=2B2026.0670.82——
InternVL3.5Parameters=2B2026.0370.8——
EvoLMMModel Scale=2B2026.0670.69——
VisionZip + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.45×, Channel multiplier=0.25×, KV Cache budget multiplier=0.28×2026.0570.51——
VisPlayModel Scale=2B2026.0669.96——
ShareGPT5V-mini-SFTTraining=SFT2026.0569.94——
VisionZero-ChartModel Scale=2B2026.0669.61——
VisionZero-RWModel Scale=2B2026.0669.58——
ERASERatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.0569.25——
BaseModel Scale=2B2026.0669.02——
VisionZip + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.45×, Channel multiplier=0.25×, KV Cache budget multiplier=0.28×2026.0567.13——
Cont-Squeeze (128 -> 1)Optimization steps=+0 steps2026.0267.01——
DART + RandomBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1265.9——
DART+VTWBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1265.8——
Eagle2number of parameters=1.5B2025.0765.8——
FastV + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.40×, Channel multiplier=0.25×, KV Cache budget multiplier=0.25×2026.0565.79——
Qwen2-VL-2B#Params=2.2B2024.1265.5——
DARTBackbone=Qwen2.5-VL-7B, Token Retention Ratio=50%2025.1264.4——
IVC-PruneRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.0563.72——
InstructVLA-Generalistnumber of parameters=1.5B2025.0763.7——
InstructVLA-Generalistnumber of parameters=1.5B, robot state=true2025.0763.7——
FastV + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.40×, Channel multiplier=0.25×, KV Cache budget multiplier=0.25×2026.0563.01——
Qwen2-VLnumber of parameters=1.5B2025.0761.4——
Llama 4 ScoutZero-shot=true2026.0260.1——
VisionZip (token only)Backbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.28×, Channel multiplier=1.00×, KV Cache budget multiplier=0.28×2026.0558.46——
InternVL2#Params=2.2B2024.1257.7——
Gemma 3 4B ITZero-shot=true2026.0256.2——
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=2048, Cropping Strategy=CropVLM2025.1155.95——
HoVLE#Params=2.6B, Resolution=HD2024.1255.7——