List your company
ResearchFormerly WizWand

Text-based Visual Question Answering on TextVQA (Accuracy and Latency)

100Average ScoreVanilla
46.242460.198774.15588.1113Dec 22, 2025Dec 28, 2025Jan 4, 2026Jan 10, 2026Jan 17, 2026Jan 23, 2026Jan 30, 2026
Updated 5mo ago

Evaluation results

MethodLinks
VanillaBackbone=LLaVA-1.5-13B, Tokens=322026.0110059.42,394
VisionTrimBackbone=LLaVA-1.5-13B, Tokens=322026.0196.357.6704
VanillaBackbone=Qwen2.5-VL-7B, Token Retention Rate=100%2025.1276.23——
FastVBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1274.65——
D²PrunerBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1274.64——
VisionZipBackbone=LLaVA-1.5-13B, Tokens=322026.0171.453.61,096
D²PrunerBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1270.54——
DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1270.43——
FastVBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1266.4——
DARTBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1264.74——
S-MFTLanguage Model Backbone=Qwen2.5-7B, Vision Tower=SigLIP-ViT-Large2025.1264.2——
FFTLanguage Model Backbone=Qwen2.5-7B, Vision Tower=SigLIP-ViT-Large2025.1264——
LoRALanguage Model Backbone=Qwen2.5-7B, Vision Tower=SigLIP-ViT-Large2025.1263.7——
S-MFTLanguage Model Backbone=Vicuna-7B, Vision Tower=CLIP-ViT-Large2025.1258.8——
DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1258.79——
FFTLanguage Model Backbone=Vicuna-7B, Vision Tower=CLIP-ViT-Large2025.1258.4——
LoRALanguage Model Backbone=Vicuna-7B, Vision Tower=CLIP-ViT-Large2025.1258.3——
FFTLanguage Model Backbone=LLaMA2-7B, Vision Tower=CLIP-ViT-Large2025.1257.7——
S-MFTLanguage Model Backbone=LLaMA2-7B, Vision Tower=CLIP-ViT-Large2025.1257.6——
LoRALanguage Model Backbone=LLaMA2-7B, Vision Tower=CLIP-ViT-Large2025.1256.5——
DARTBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1248.31——
CLIP-ScoreBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—39.7—
CoIDOBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—43.2—
COINCIDEBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—33—
EL2NBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—42.4—
Full-DataBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=100%2026.05—38.3—
ICONSBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—49.9—
IFDBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—42—
OFABackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—55.7—
PreSelBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—39.7—
RandomBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—38.7—
Self-FilterBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—42.6—
TypiClustBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—37.7—
XMASBackbone=LLaVA-v1.5-7B, Fine-tuning Dataset=Vision-Flan-186K, Sampling Ratio=15%2026.05—42.4—