List your company
ResearchFormerly WizWand

Visual Question Answering on TextVQA (TextVQA Metric)

85.9TextVQA AccuracyAttWarp-Chain
62.70868.72974.7580.771Apr 14, 2025Jun 19, 2025Aug 25, 2025Oct 31, 2025Jan 6, 2026Mar 14, 2026May 20, 2026
Updated 5mo ago

Evaluation results

MethodLinks
AttWarp-ChainBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Adaptive Chains2025.1085.9
Qwen2.5-VL#Params=7B, Model Category=Vision Language Models, Co-training=false2026.0584.9
Qwen2.5-VLLLM=Qwen2.5-7B, Token Type=2D-Continuous dynamic2026.0584.9
AttWarpBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Rectilinear warping2025.1084.7
Qwen2-VL#Params=7B, Model Category=Vision Language Models, Co-training=false2026.0584.3
DeepSeek-VL2#Params=4B/27B, Model Category=Vision Language Models, Co-training=false2026.0584.2
UAM#Params=7B MoT, Model Category=Vision-Language-Action Models, Co-training=false2026.0584.2
AttWarp-DistillBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Efficient inference2025.1084.1
ViCropBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Add object crop2025.1083.8
Qwen2.5VL 7B# Vis tok.=14002025.0483.5
Qwen2.5-VL-7BToken Budget=100%2025.0883.1
BLIP3-oLLM=Qwen2.5VL-7B-Instruct, Token Type=2D-Continuous dynamic2026.0583.1
BaselineBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=1.00×, Channel multiplier=1.00×, KV Cache budget multiplier=1.00×2026.0582.92
FastV + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.40×, Channel multiplier=0.25×, KV Cache budget multiplier=0.25×2026.0582.38
Vanilla2026.0482.1
Qwen3-VL-4BModel=Qwen3-VL, Parameter Count=4B2026.0481.86
APIBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Alpha channel fade2025.1081.6
FastV (token only)Backbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.25×, Channel multiplier=1.00×, KV Cache budget multiplier=0.25×2026.0581.49
Base MLLMBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Cross-attention VL adapter & partially closed data2025.1081
VisionZip + RotateKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.45×, Channel multiplier=0.25×, KV Cache budget multiplier=0.28×2026.0580.84
DeepSeek-VL2-Tiny# Vis tok.=15002025.0480.7
Molmo 7B# Vis tok.=12002025.0480.4
UniVLRVariant=Ours2026.0580.4
CropVLMReward=Accuracy, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1180.12
CropVLMReward=LL, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1180.07
FLARE-X 8B# Vis tok.=14002025.0479.7
Qwen2-VL#Params=1.5B, Model Category=Vision Language Models, Co-training=false2026.0579.7
Qwen2.5VL 3B# Vis tok.=14002025.0479.3
SkiLaModel=SkiLa2026.0579.3
Qwen2.5-VL#Params=3B, Model Category=Vision Language Models, Co-training=false2026.0579.3
SLVR-7BReasoning Paradigm=Visual + Semantic Latent2026.0579.3
Qwen2.5-VL-3BModel=Qwen2.5-VL, Parameter Count=3B2026.0479.29
Qwen 2.5 VLInput Resolution=1792x17922025.1179.14
UniVLRStage=12026.0579.1
InternVL2.5LLM=InternLM2.5-7B, Token Type=2D-Continuous dynamic2026.0579.1
Qwen2.5-VL-7BReasoning Paradigm=Text-only2026.0579.1
BaseRatio=100%, Backbone=Qwen2.5-VL-3B2026.0579.06
ERASERatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0578.62
HiPruneBase Model=Qwen2.5-VL-7B, Token Budget=33.3%2025.0878.5
InfiniteVL-4BModel=InfiniteVL, Parameter Count=4B, Result Source=InfiniteVL (Tao et al., 2025) paper2026.0478.5
LLaVA-OneVision 7B# Vis tok.=24002025.0478.5
ERASERatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0577.95
HiPrune++Base Model=Qwen2.5-VL-7B, Token Budget=33.3%2025.0877.9
FastV + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.40×, Channel multiplier=0.25×, KV Cache budget multiplier=0.25×2026.0577.81
Qwen2.5-VL-3BToken Budget=100%2025.0877.8
Qwen2.5-VL-7BModel=Qwen2.5-VL-7B2026.0577.5
IVC-PruneRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0577.32
FGVP-maskBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Green mask overlay2025.1077.3
CDPrunerRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0577.25
FLARE-X 3B# Vis tok.=14002025.0477.2
Eagle 8B# Vis tok.=10242025.0477.1
VisionZip (token only)Backbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.28×, Channel multiplier=1.00×, KV Cache budget multiplier=0.28×2026.0576.84
MiniCPM-Llama3-V-2.5 8B# Vis tok.=4002025.0476.6
VanillaToken Retention Ratio=100%2026.0576.5
ERASERatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.0576.23
FLARE-L 8B# Vis tok.=6302025.0476.2
SPprunerToken Retention Ratio=70%2026.0576
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=1024, Cropping Strategy=CropVLM2025.1175.97
LLaVA-OneVision-7BModel=LLaVA-OneVision, Parameter Count=7B2026.0475.92
DARTToken Retention Ratio=70%2026.0575.9
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=2048, Cropping Strategy=CropVLM2025.1175.72
VisionZip + ThinKBackbone=QWEN2.5-VL-7B-INSTRUCT, Token multiplier=0.45×, Channel multiplier=0.25×, KV Cache budget multiplier=0.28×2026.0575.72
CDPrunerRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0575.61
LVRModel=LVR2026.0575.6
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=512, Cropping Strategy=CropVLM2025.1175.46
SPprunerToken Retention Ratio=50%2026.0575.3
LVRReasoning Paradigm=Visual Latent2026.0575.1
IVC-PruneRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0575.06
DASHViT retain tokens=50%, LLM retain tokens=50%2026.0474.7
DARTToken Retention Ratio=50%2026.0574.6
Florence-VL 8B# Vis tok.=5762025.0474.2
HiPruneBase Model=Qwen2.5-VL-7B, Token Budget=22.2%2025.0874.1
MiniCPM-V-2.0 3B# Vis tok.=4002025.0474.1
FLARE-L 3B# Vis tok.=6302025.0473.3
HiPrune++Base Model=Qwen2.5-VL-7B, Token Budget=22.2%2025.0873.2
SPprunerToken Retention Ratio=30%2026.0573
FGVP-blurBase MLLM=Qwen (Yang et al., 2024a), Key Technique=Blur background2025.1072.3
DARTRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0572.24
ILLUMELLM=Vicuna-7B, Token Type=2D-Continuous, Res.=2242026.0572.1
DARTToken Retention Ratio=30%2026.0572
SpB2.0-VL-5BModel=SpB2.0-VL, Parameter Count=5B2026.0471.72
Cambrain 8B# Vis tok.=5762025.0471.7
CDPrunerRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.0571.57
ChatVLA#Params=2B, Model Category=Vision-Language-Action Models, Co-training=true2026.0571.2
HiPruneBase Model=Qwen2.5-VL-3B, Token Budget=33.3%2025.0870.1
Visual-SR1Reasoning Paradigm=Text CoT2026.0570.1
HiPrune++Base Model=Qwen2.5-VL-3B, Token Budget=33.3%2025.0869.2
Florence-VL 3B# Vis tok.=5762025.0469.1
SmolVLM2-2.2BModel=SmolVLM2, Parameter Count=2.2B2026.0468.81
FastVToken Retention Ratio=70%2026.0568.4
IVC-PruneRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.0568.27
Qwen 2.5 VL 3BBase Model=Qwen 2.5 VL 3B, Resolution=-, Cropping Strategy=None2025.1168.2
FastVToken Retention Ratio=50%2026.0567.8
DARTRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.0566.58
CoVTModel=CoVT2026.0566.5
PruneSIDRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.0565.89
BaselineBackbone=LLAMA3-LLAVA-NEXT-8B, Token multiplier=1.00×, Channel multiplier=1.00×, KV Cache budget multiplier=1.00×2026.0565.4
FastVToken Retention Ratio=30%2026.0565.3
Emu3LLM=8B (from scratch), Token Type=2D-Discrete, Res.=5122026.0564.7
DART (ViT)ViT retain tokens=50%, LLM retain tokens=50%2026.0463.6