List your company
ResearchFormerly WizWand

Text-based Visual Question Answering on TextVQA (Accuracy)

88.5AccuracyQ-Mask
77.715280.515183.31586.1149Sep 19, 2024Jan 4, 2025Apr 21, 2025Aug 6, 2025Nov 21, 2025Mar 8, 2026Jun 23, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Q-MaskSize=3B2026.0388.5——
Qwen2.5-VL-32B + VPPOZero-shot=true2026.0286.2——
ToolTreeBackbone Model=GPT-4o2026.0385.43——
DualPDBackbone=Qwen-2.5-VL-7B, Decoding Strategy=DualPD2026.0185.34——
Qwen2.5-VL-32B + AT-RL (Ours)Zero-shot=true2026.0285.3——
Qwen2.5VLModel Scale=7B, Date=2025.02, Tokenization Paradigm=Understanding-only Models2026.0284.9——
Qwen2.5-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=8B2025.1084.9——
Qwen2.5-VLSize=7B2026.0384.9——
DualPDBackbone=Qwen-2-VL-7B, Decoding Strategy=DualPD2026.0184.84——
Qwen2.5-VL-7BToken Pruning Rate=0%2026.0484.8——
Qwen2.5-VL-32B InstructZero-shot=true2026.0284.6——
LLaVA-OVLLM=Qwen2-7B, Number of Parameters=~7B2026.0284.5——
Qwen2.5-VL-72B InstructZero-shot=true2026.0284.3——
ManzanoModel Scale=30B, Date=2025.09, Tokenization Paradigm=Hybrid Tokenization2026.0284.3——
Qwen2-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=8B2025.1084.3——
InternVL3.5Size=78B2026.0384.3——
DoLABackbone=Qwen-2.5-VL-7B, Decoding Strategy=DoLA2026.0184.19——
DoLABackbone=Qwen-2-VL-7B, Decoding Strategy=DoLA2026.0183.99——
Qwen-2.5-VL-7BBackbone=Qwen-2.5-VL-7B, Decoding Strategy=Standard2026.0183.72——
Latent DenoisingArchitecture=Qwen-2.5-VL2026.0483.7——
Qwen2.5-VLSize=72B2026.0383.5——
Qwen-2-VL-7BBackbone=Qwen-2-VL-7B, Decoding Strategy=Standard2026.0183.34——
VanillaBackbone=Qwen2-VL-7B, Flops Ratio Reduction=0%2026.0483.2——
MolmoSize=72B2026.0383.1——
DenseModel=Qwen2.5-VL-7B, Bits=FP162026.0382.9——
MBQModel=Qwen2.5-VL-7B, Bits=W4A162026.0382.9——
MBQModel=Qwen2.5-VL-7B, Bits=W8A82026.0382.9——
DeFactoBackbone=Qwen2.5-VL-7B2025.0982.9——
InternVL3.5Size=38B2026.0382.7——
SQModel=Qwen2.5-VL-7B, Bits=W8A82026.0382.6——
MASQuantModel=Qwen2.5-VL-7B, Bits=W8A82026.0382.6——
Q-MaskSize=2B2026.0382.6——
FP16Model=Qwen2-VL-72B, Bitwidth=FP162024.1282.5—78.1
MBQModel=Qwen2-VL-72B, Bitwidth=W3A162024.1282.5—77.6
MASQuantModel=Qwen2.5-VL-7B, Bits=W4A162026.0382.5——
RTNModel=Qwen2.5-VL-7B, Bits=W8A82026.0382.5——
AWQModel=Qwen2-VL-72B, Bitwidth=W3A162024.1282.4—77.5
ToolTreeBackbone Model=GPT-4o-mini2026.0382.26——
Full ModelBackbone=Qwen3-VL-8B-Instruct, Pruning Ratio=0%2026.0282.24——
AWQModel=Qwen2.5-VL-7B, Bits=W4A162026.0382.2——
Qwen2-VL-7B-InstructMitigation=None (baseline)2025.0782.2——
SENTINELBase Model=Qwen2-VL-7B-Instruct2025.0782.2——
FP16Model=InternVL2-26B, Bitwidth=FP162024.1282.1—74.6
RTNModel=Qwen2.5-VL-7B, Bits=W4A162026.0382.1——
Qwen3-VLSize=8B2026.0382.1——
BaselineArchitecture=Qwen-2.5-VL2026.0481.8——
Qwen2VL-7BModel size=7B2025.0481.7——
VanillaBackbone=Qwen3-VL-4B-FP8, Flops Ratio Reduction=0%, Precision=FP82026.0481.7——
GPTQModel=Qwen2-VL-72B, Bitwidth=W3A162024.1281.6—76.6
KelixModel Scale=8B, Tokenization Paradigm=Discrete Tokenization2026.0281.4——
MBQModel=InternVL2-26B, Bitwidth=W3A162024.1281.1—73.8
WandaBackbone=Qwen3-VL-8B-Instruct, Pruning Ratio=30%2026.0281.08——
AWQModel=InternVL2-26B, Bitwidth=W3A162024.1281—73.5
Qwen3-VL + DISCO (10K)Backbone=Qwen3-VL, Alignment Strategy=DISCO (10K)2026.0280.83——
POPBackbone=Qwen3-VL-8B-Instruct, Pruning Ratio=33.3%2026.0280.73——
RTNModel=InternVL2-26B, Bitwidth=W3A162024.1280.6—73.3
VQ-RAEModel Scale=7B, Date=2025.11, Tokenization Paradigm=Hybrid Tokenization2026.0280.6——
Qwen3-VLSize=4B2026.0380.6——
MBQModel=Qwen2-VL-72B, Bitwidth=W4A82024.1280.5—75.8
SpaRE-7BModel size=7B2025.0480.5——
OpenAI GPT-4oZero-shot=true2026.0280.5——
Qwen3-VLBackbone=Qwen3-VL2026.0280.34——
InternVL3Input=Tile-wise, RoPE=1D-RoPE, Backbone=Dense, Parameter Scale=8B2025.1080.2——
Qwen3-VLToken Ratio=Full Tokens2026.0680.2——
GPTQModel=InternVL2-26B, Bitwidth=W3A162024.1280.1—72.3
DualPDBackbone=LLaVA-1.6-13B, Decoding Strategy=DualPD2026.0180.1——
MBQModel=InternVL2-26B, Bitwidth=W4A82024.1280—72.7
BagelModel Scale=14B, Date=2025.05, Tokenization Paradigm=Continuous Tokenization2026.0280——
Qwen3-VL + Textual (10K)Backbone=Qwen3-VL, Alignment Strategy=Textual (10K)2026.0279.91——
RTNModel=Qwen2-VL-72B, Bitwidth=W3A162024.1279.7—75
Qwen2.5VLModel Scale=3B, Date=2025.02, Tokenization Paradigm=Understanding-only Models2026.0279.7——
Qwen2-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=2B2025.1079.7——
CASTRatio=1%, Sampling Strategy=CAST, Backbone=Qwen2-VL-2B2026.0579.68——
RTNModel=InternVL2-26B, Bitwidth=W4A82024.1279.6—72.7
GLM-4.1VSize=9B2026.0379.6——
DeSAPToken Pruning Rate=66.7%2026.0479.5——
CASTRatio=10%, Sampling Strategy=CAST, Backbone=Qwen2-VL-2B2026.0579.43——
CASTRatio=5%, Sampling Strategy=CAST, Backbone=Qwen2-VL-2B2026.0579.4——
FP16Model=LLaVA-onevision-72B, Bitwidth=FP162024.1279.3—74.3
Qwen2.5-VL 3B (reported)Res.=Native2025.1279.3——
Qwen2.5-VLInput=Any Res., RoPE=M-RoPE, Backbone=Dense, Parameter Scale=2B2025.1079.3——
TokenVL-8BSize=8B2026.0379.3——
Qwen2.5-VLSize=3B2026.0379.3——
Qwen3-VLSize=2B2026.0379.3——
SpaRE-2BModel size=2B2025.0479.2——
Qwen2.5-VL 3B (reproduced)Res.=≤ 896²2025.1279.2——
RandomRatio=10%, Sampling Strategy=Random, Backbone=Qwen2-VL-2B2026.0579.15——
InternVL2.5Input=Tile-wise, RoPE=1D-RoPE, Backbone=Dense, Parameter Scale=8B2025.1079.1——
Qwen2.5-VLBackbone=Qwen2.5-VL-7B2025.0979.1——
HoloVToken Pruning Rate=66.7%2026.0478.9——
RandomRatio=5%, Sampling Strategy=Random, Backbone=Qwen2-VL-2B2026.0578.61——
MBQModel=LLaVA-onevision-72B, Bitwidth=W3A162024.1278.5—73.6
InternVL3.5Size=20B2026.0378.5——
SENTINELBase Model=Qwen2-VL-2B-Instruct2025.0778.5——
Oryx-1.5Size=32B2024.0978.3——
Qwen2-VL-2B-InstructMitigation=None (baseline)2025.0778.3——
LLaVA-1.6-13BBackbone=LLaVA-1.6-13B, Decoding Strategy=Greedy2026.0178.2——
DoLABackbone=LLaVA-1.6-13B, Decoding Strategy=DoLA2026.0178.2——
InternVL3.5Size=8B2026.0378.2——
RandomRatio=1%, Sampling Strategy=Random, Backbone=Qwen2-VL-2B2026.0578.13——