List your company

Document Question Answering on DocVQA

97.87ANLSBaseline
4.914829.047453.1877.3126Oct 6, 2022May 15, 2023Dec 22, 2023Jul 30, 2024Mar 8, 2025Oct 15, 2025May 25, 2026

Evaluation results

MethodLinks
BaselineBackbone=LLaVA-OneVision-1.5-8B-Instruct2026.0297.87—
VisionSelectorRetain Tokens=25%2026.0294.08—
IDPrunerRetain Tokens=25%2026.0291.82—
VanillaBackbone=Qwen2.5-VL-7B, Token Retention Rate=100%2025.1291—
ERNIE-LayoutlargeFine-tuning set=train+dev, Context=leaderboard2022.1088.41—
TILTlargeFine-tuning set=-2022.1087.05—
LayoutLMv2largeFine-tuning set=train + dev2022.1085.290.082
ERNIE-LayoutlargeFine-tuning set=train+dev2022.1084.860.1534
DARTRetain Tokens=25%2026.0284.77—
StructuralLMlargeFine-tuning set=-2022.1083.49—
LayoutLMv2largeFine-tuning set=train2022.1083.480.0639
ERNIE-LayoutlargeFine-tuning set=train2022.1083.210.1369
DivPruneRetain Tokens=25%2026.0281.53—
D²PrunerBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1281.14—
UniLMv2largeFine-tuning set=train2022.1077.09—
SCOPERetain Tokens=25%2026.0275.68—
DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1273.88—
XDocPre-training Steps=1M, Pre-train (P)=true, Pre-train (D)=true, Pre-train (W)=true2022.1072.7—
XDocPre-training Steps=500K, Pre-train (P)=true, Pre-train (D)=true, Pre-train (W)=true2022.1072.6—
LayoutLMlargeFine-tuning set=train2022.1072.59—
VisionSelectorRetain Tokens=10%2026.0271.31—
FastVBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1271.26—
IDPrunerRetain Tokens=10%2026.0270.43—
ROBERTalargeFine-tuning set=train2022.1069.52—
XDocPre-training Steps=100K, Pre-train (D)=true, Pre-train (W)=true2022.1069.5—
XDocPre-training Steps=100K, Pre-train (D)=true2022.1069.4—
LayoutLMPre-train (D)=true2022.1069.2—
XDocPre-training Steps=100K, Pre-train (P)=true, Pre-train (D)=true2022.1069.1—
XDocPre-training Steps=100K, Pre-train (P)=true, Pre-train (D)=true, Pre-train (W)=true2022.1068.3—
BERTlargeFine-tuning set=train2022.1067.68—
UNITBase LVLM=Llava-Next, Resolution Strategy=high-resolution grid slicing, Resolution=4482024.0965.5—
XDocPre-training Steps=100K, Pre-train (P)=true, Pre-train (W)=true2022.1064.7—
XDocPre-training Steps=100K, Pre-train (P)=true2022.1064.5—
XDocPre-training Steps=100K, Pre-train (W)=true2022.1063.6—
SigLIPBase LVLM=Llava-Next, Resolution Strategy=high-resolution grid slicing, Resolution=3842024.0962—
VisPrunerRetain Tokens=25%2026.0261—
D²PrunerBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1257.54—
CLIP-LBase LVLM=Llava-Next, Resolution Strategy=high-resolution grid slicing, Resolution=3362024.0957.2—
DivPruneRetain Tokens=10%2026.0256.57—
FastVRetain Tokens=25%2026.0256.53—
DARTBackbone=Qwen2.5-VL-7B, Token Retention Rate=25%2025.1252.99—
DARTRetain Tokens=10%2026.0252.77—
DivPruneBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1249.34—
SCOPERetain Tokens=10%2026.0247.65—
FastVBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1236.85—
VisPrunerRetain Tokens=10%2026.0235.18—
DARTBackbone=Qwen2.5-VL-7B, Token Retention Rate=10%2025.1230.1—
FastVRetain Tokens=10%2026.0229.77—
VisionZipRetain Tokens=25%2026.0223.69—
HiPruneRetain Tokens=25%2026.0221.01—
VisionZipRetain Tokens=10%2026.0213.2—
HiPruneRetain Tokens=10%2026.0212.86—
MMaDA+VRCDL=384, FR=0.52026.0512.38—
MMaDA+VRCDL=192, FR=0.52026.0511.88—
MMaDA+VRCDL=384, FR=0.252026.0511.74—
MMaDAL=384, FR=0.52026.0511.64—
MMaDA+VRCDL=192, FR=0.252026.0511.38—
MMaDAL=384, FR=0.252026.0511.23—
MMaDAL=192, FR=0.52026.0511.19—
MMaDAL=192, FR=0.252026.0510.18—
MMaDA+VRCDL=384, FR=0.1252026.059.58—
MMaDAL=384, FR=0.1252026.059.27—
MMaDA+VRCDL=192, FR=0.1252026.059.09—
MMaDAL=192, FR=0.1252026.058.49—