List your company
ResearchFormerly WizWand

Visual Question Answering on AI2D

91.81AccuracyGR3D-8B
78.508481.961785.41588.8683Aug 24, 2023Feb 13, 2024Aug 5, 2024Jan 26, 2025Jul 19, 2025Jan 9, 2026Jul 2, 2026
Updated 3mo ago

Evaluation results

MethodLinks
GR3D-8BTraining Stage=Stage 12026.0591.81
GR3D-8BTraining Stage=Stage 22026.0591.54
NVILA-Lite-8B2026.0591.01
Gemini 2.5 Pro2026.0488.4
OpenVLThinkerV2Parameters=8B2026.0487.5
BaselineModel=Llama Nemotron Nano 12B v2 VL2026.0187.3
VISEModel Scale=32B2026.0687.24
VisionZero-ChartModel Scale=32B2026.0687.16
VisionZero-CLEVRModel Scale=32B2026.0687.1
iReasonerModel Scale=32B2026.0687.08
EvoLMMModel Scale=32B2026.0687.02
PTQModel=Llama Nemotron Nano 12B v2 VL2026.0186.8
VisPlayModel Scale=32B2026.0686.71
QADModel=Llama Nemotron Nano 12B v2 VL2026.0186.7
VisionZero-RWModel Scale=32B2026.0686.69
BaseModel Scale=32B2026.0686.63
QATModel=Llama Nemotron Nano 12B v2 VL2026.0186.5
Qwen3-VL-8BRetained Tokens=1024 Tokens (100%)2026.0786.3
EMOVAModel Size=72B2024.0985.8
InternVL3Input=Tile-wise, RoPE=1D-RoPE, Backbone=Dense, Parameter Scale=8B2025.1085.2
OneThinker-8BParameters=8B2026.0485.2
VanillaSource=Qwen’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=0%2026.0685.2
Qwen3-VL GRPORL Strategy=GRPO2026.0485.1
Qwen3-VL GDPORL Strategy=GDPO2026.0485.1
GPT-4o2026.0484.9
VisionZero2026.0484.8
GPT-4o2024.0984.6
InternVL2.5Input=Tile-wise, RoPE=1D-RoPE, Backbone=Dense, Parameter Scale=8B2025.1084.5
MM-Eureka-7BParameters=7B2026.0484.1
VISEModel Scale=8B2026.0684.1
VisionZero-ChartModel Scale=8B2026.0684.04
Qwen3-VL-4BModel=Qwen3-VL, Parameter Count=4B2026.0484
VisionZero-CLEVRModel Scale=8B2026.0683.98
iReasonerModel Scale=8B2026.0683.96
Qwen2.5-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=8B2025.1083.9
EvoLMMModel Scale=8B2026.0683.89
PTPModel=InternVL2-8B, Pruning Method=PTP, Pruning Ratio (r)=0.5, Balancing Weight (alpha)=0.52025.0983.8
FLARE-X 8B# Vis tok.=14002025.0483.6
VisPlayModel Scale=8B2026.0683.45
Qwen2.5VL 7B# Vis tok.=14002025.0483.4
VisionZero-RWModel Scale=8B2026.0683.37
BaseModel Scale=8B2026.0683.31
NEOInput=Any Res., RoPE=Native-RoPE, Backbone=Dense, Parameter Scale=8B2025.1083.1
MoVALLM=Hermes-Yi-34B, Params=38B, Model Type=Specialist2024.0483
Qwen2-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=8B2025.1083
Encoder-BasedInput=Tile-wise, RoPE=1D-RoPE, Backbone=Dense, Parameter Scale=8B2025.1082.9
VanillaSource=Qwen’25, Token Reduction=0%, Backbone=Qwen2.5-VL-7B2026.0682.6
LLaVA-OneVision-7BModel=LLaVA-OneVision, Parameter Count=7B2026.0482.51
FP16 BaselineModel=InternVL2-8B, Bitwidth=FP162026.0382.42
InternVL2-8BModel=InternVL2-8B, Pruning Method=None, Pruning Ratio (r)=1.02025.0982.4
RADIOv2.5-H + TilingVision Encoder=RADIOv2.5-H (ours) + Tiling, Resolution=Up to 7 × 768², Compression=ToMe r=2108, Tokens/im=~ 12332024.1282.4
Qwen3-VL-Instruct-8BParameters=8B2026.0482.3
VISEModel Scale=4B2026.0682.16
GSearchModel=InternVL2-8B, Pruning Method=GSearch2025.0982.1
Vision-G12026.0482.1
VTWModel=InternVL2-8B, Pruning Method=VTW2025.0981.8
OpenVLThinker-7BParameters=7B2026.0481.8
VisionZero-ChartModel Scale=4B2026.0681.76
EMOVAModel Size=7B2024.0981.7
Qwen2.5-VL 3B (reported)Res.=Native2025.1281.6
Qwen2.5-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=2B2025.1081.6
Qwen3-VL-4B-Instruct2026.0281.6
LLaVA-OneVision 7B# Vis tok.=24002025.0481.6
FLARE-L 8B# Vis tok.=6302025.0481.6
GeminiProVision2024.0681.4
Qwen2.5VL 3B# Vis tok.=14002025.0481.4
FP16 BaselineModel=LLaVA-onevision-7B, Bitwidth=FP162026.0381.31
iReasonerModel Scale=4B2026.0681.22
PALI-X-55BModel type=Specialist SOTAs, Configuration=Single-task fine-tuning, without OCR Pipeline, Parameters=55B2023.0881.2
PALI-X-55BParams=55B, Model Type=Specialist2024.0481.2
FLARE-X 3B# Vis tok.=14002025.0481.2
EvoLMMModel Scale=4B2026.0681.13
MG-LLAVALLM=Yi1.5-34B, Parameters=34.4B, Resolution=336(768)2024.0681.1
FastVModel=InternVL2-8B, Pruning Method=FastV, Pruning Ratio (r)=0.52025.0981.1
Molmo 7B# Vis tok.=12002025.0481
VisionZero-CLEVRModel Scale=4B2026.0680.81
VL-Rethinker-7BParameters=7B2026.0480.8
SpB2.0-VL-5BModel=SpB2.0-VL, Parameter Count=5B2026.0480.73
RTNModel=InternVL2-8B, Bitwidth=W3A162026.0380.51
Qwen2.5-VL-3BModel=Qwen2.5-VL, Parameter Count=3B2026.0480.51
PDropModel=InternVL2-8B, Pruning Method=PDrop2025.0980.4
VisPlayModel Scale=4B2026.0680.34
VisionZero-RWModel Scale=4B2026.0680.31
Gemini Pro 1.52024.0980.3
FP16 BaselineModel=Qwen2-VL-7B, Bitwidth=FP162026.0380.12
Qwen2.5-VL 3B (reproduced)Res.=≤ 896²2025.1280.1
NEOInput=Any Res., RoPE=Native-RoPE, Backbone=Dense, Parameter Scale=2B2025.1080.1
BaseModel Scale=4B2026.0680.1
QIGModel=InternVL2-8B, Bitwidth=W3A162026.0379.73
MBQModel=InternVL2-8B, Bitwidth=W3A162026.0379.66
QIGModel=InternVL2-8B, Bitwidth=W4A82026.0379.63
AWQModel=InternVL2-8B, Bitwidth=W3A162026.0379.47
MBQModel=InternVL2-8B, Bitwidth=W4A82026.0379.47
FLARE-L 3B# Vis tok.=6302025.0479.4
VITAModel Size=1.52024.0979.3
SigLIP SO400M + TilingVision Encoder=SigLIP SO400M [52] + Tiling, Resolution=Up to 13 × 384², Compression=2 × 2 Unshuffle, Tokens/im=~ 19282024.1279.3
Eagle2number of parameters=1.5B2025.0779.3
QIGModel=LLaVA-onevision-7B, Bitwidth=W3A162026.0379.11
InstructVLA-Generalistnumber of parameters=1.5B2025.0779.1
RTNModel=InternVL2-8B, Bitwidth=W4A82026.0379.02