Visual Question Answering on DocVQA (ANLS)
95.01ANLSOScaR
Evaluation Results
| Method | Links | |
|---|---|---|
| OScaRBackbone=Qwen3-VL-8B, Bit-width=INT2, Group Size=1282026.05 | 95.01 | |
| 16bitBackbone=Qwen3-VL-8B, Bit-width=16-bit2026.05 | 94.93 | |
| OTTBackbone=Qwen3-VL-8B, Bit-width=INT2, Group Size=1282026.05 | 94.86 | |
| KIVIBackbone=Qwen3-VL-8B, Bit-width=INT2, Group Size=1282026.05 | 94.7 | |
| TurboQuant+Backbone=Qwen3-VL-8B, Bit-width=2.5-bit2026.05 | 94.51 | |
| 16bitBackbone=Qwen3-VL-4B, Bit-width=16-bit2026.05 | 94.23 | |
| QuaRotBackbone=Qwen3-VL-8B, Bit-width=INT2, Group Size=1282026.05 | 93.93 | |
| OScaRBackbone=Qwen3-VL-4B, Bit-width=INT2, Group Size=1282026.05 | 93.85 | |
| InternVL-2.5-26B*Latency=0.74542025.12 | 93.78 | |
| TurboQuant+Backbone=Qwen3-VL-4B, Bit-width=2.5-bit2026.05 | 93.54 | |
| KIVIBackbone=Qwen3-VL-4B, Bit-width=INT2, Group Size=1282026.05 | 93.49 | |
| OTTBackbone=Qwen3-VL-4B, Bit-width=INT2, Group Size=1282026.05 | 93.24 | |
| RotateKVBackbone=Qwen3-VL-8B, Bit-width=INT2, Group Size=1282026.05 | 92.76 | |
| LUVCBackbone=InternVL-2.5-26B*, Latency=0.4302, Speedup=1.732025.12 | 92.7 | |
| InternVL-2.5-8B*Latency=0.29852025.12 | 91.96 | |
| LUVCBackbone=InternVL-2.5-8B*, Latency=0.1857, Speedup=1.612025.12 | 91.48 | |
| QuaRotBackbone=Qwen3-VL-4B, Bit-width=INT2, Group Size=1282026.05 | 89.89 | |
| ViQBase LLM=Qwen2.5-7B, Size=1.3B, AnyRes=true, Discrete=true2026.06 | 88.9 | |
| VTWBackbone=InternVL-2.5-26B*, Latency=0.4637, Speedup=1.612025.12 | 88.78 | |
| PaLI-3OCR pipeline input=true, Resolution=1064x10642023.10 | 88.6 | |
| ERNIE-ViL 2.0OCR pipeline input=true, Resolution=1064x10642023.10 | 88.4 | |
| PaLI-3OCR pipeline input=false, Resolution=1064x10642023.10 | 87.6 | |
| TILTInput type=Pipelined, Additional labeled data=true2022.10 | 87.1 | |
| PACTBackbone=InternVL-2.5-26B*, Latency=0.4690, Speedup=1.592025.12 | 85.79 | |
| InternViT-2.5-6BBase LLM=Qwen2.5-7B, Size=6.0B, AnyRes=false, Discrete=false2026.06 | 85.7 | |
| UDOPInput type=Pipelined2022.10 | 84.7 | |
| PACTBackbone=InternVL-2.5-8B*, Latency=0.1980, Speedup=1.512025.12 | 84.32 | |
| ViQBase LLM=Qwen2.5-1.5B, Size=1.3B, AnyRes=true, Discrete=true2026.06 | 84.2 | |
| LayoutLMv3Input type=Pipelined2022.10 | 83.4 | |
| InternViT-2.5Base LLM=Qwen2.5-1.5B, Size=0.3B, AnyRes=false, Discrete=false2026.06 | 81.7 | |
| T5 + 2D + UInput type=Pipelined2022.10 | 81 | |
| VTWBackbone=InternVL-2.5-8B*, Latency=0.1804, Speedup=1.652025.12 | 80.59 | |
| LeoPT=595 K, SFT=1 M2025.01 | 80.1 | |
| InternViT-2.5-6BBase LLM=Qwen2.5-1.5B, Size=6.0B, AnyRes=false, Discrete=false2026.06 | 80.1 | |
| AIMv2-HBase LLM=Qwen2.5-7B, Size=0.7B, AnyRes=false, Discrete=false2026.06 | 80.1 | |
| PaLI-XOCR pipeline input=false, Resolution=1064x10642023.10 | 80 | |
| PALI-X-55BModel type=Specialist SOTAs, Configuration=Single-task fine-tuning, without OCR Pipeline, Parameters=55B2023.08 | 80 | |
| OryxViTBase LLM=Qwen2.5-7B, Size=0.4B, AnyRes=true, Discrete=false2026.06 | 79.8 | |
| Eagle-X2PT=595 K, SFT=1.8 M2025.01 | 77.7 | |
| Eagle-X3PT=595 K, SFT=1.8 M2025.01 | 77.7 | |
| DinoV2-gBase LLM=Qwen2.5-1.5B, Size=1.1B, AnyRes=false, Discrete=false2026.06 | 76.9 | |
| Pix2Struct-LargeInput type=Pixel-only, Scale=Large2022.10 | 76.6 | |
| Pix2Struct-Large (1.3B)Model type=Generalist Models, Parameters=1.3B2023.08 | 76.6 | |
| Pix2StructResolution=10242024.04 | 76.6 | |
| RotateKVBackbone=Qwen3-VL-4B, Bit-width=INT2, Group Size=1282026.05 | 76.12 | |
| SigLIP2-gBase LLM=Qwen2.5-7B, Size=1.1B, AnyRes=false, Discrete=false2026.06 | 75 | |
| AIMv2-HBase LLM=Qwen2.5-1.5B, Size=0.7B, AnyRes=false, Discrete=false2026.06 | 73.7 | |
| OryxViTBase LLM=Qwen2.5-1.5B, Size=0.4B, AnyRes=true, Discrete=false2026.06 | 73.5 | |
| Pix2Struct-BaseInput type=Pixel-only, Scale=Base2022.10 | 72.1 | |
| HRVDAResolution=15362024.04 | 72.1 | |
| SeRumResolution=1280, Fine-tuning=true2024.04 | 71.9 | |
| ROBERTaInput type=Pipelined2022.10 | 69.5 | |
| 16bitBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=16-bit2026.05 | 68.57 | |
| SigLIP2-gBase LLM=Qwen2.5-1.5B, Size=1.1B, AnyRes=false, Discrete=false2026.06 | 67.8 | |
| OScaRBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=INT2, Group Size=1282026.05 | 67.79 | |
| DonutInput type=Pixel-only2022.10 | 67.5 | |
| DonutResolution=1280, Fine-tuning=true2024.04 | 67.5 | |
| TurboQuant+Backbone=LLaVA-v1.6-vicuna-7B, Bit-width=2.5-bit2026.05 | 67.09 | |
| OTTBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=INT2, Group Size=1282026.05 | 66.11 | |
| QuaRotBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=INT2, Group Size=1282026.05 | 65.55 | |
| UReaderResolution=2242024.04 | 65.4 | |
| KIVIBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=INT2, Group Size=1282026.05 | 65.34 | |
| RotateKVBackbone=LLaVA-v1.6-vicuna-7B, Bit-width=INT2, Group Size=1282026.05 | 65.26 | |
| Qwen-VL (Qwen-7B)Model type=Generalist Models, Backbone=Qwen-7B2023.08 | 65.1 | |
| Qwen-VLResolution=448, Open-source evaluation=true2024.04 | 65.1 | |
| Qwen-VL-ChatModel type=Generalist Models2023.08 | 62.6 | |
| mPLUG-DocOwl (LLaMA-7B)Model type=Generalist Models, Backbone=LLaMA-7B2023.08 | 62.2 | |
| mPLUG-DocResolution=2242024.04 | 62.2 | |
| OAI-CLIP-LBase LLM=Qwen2.5-7B, Size=0.3B, AnyRes=false, Discrete=false2026.06 | 61.4 | |
| OAI-CLIP-LBase LLM=Qwen2.5-1.5B, Size=0.3B, AnyRes=false, Discrete=false2026.06 | 52.9 | |
| MoME-X3PT=NA, SFT=NA2025.01 | 50.8 | |
| SCVMBackbone=LLaVA-1.5-7B2026.02 | 21 | |
| LLaVA-1.5-7BBackbone=LLaVA-1.5-7B2026.02 | 17 | |
| QLIPBase LLM=Qwen2.5-1.5B, Size=0.3B, AnyRes=false, Discrete=true2026.06 | 12.2 | |
| UniTokBase LLM=Qwen2.5-1.5B, Size=0.3B, AnyRes=false, Discrete=true2026.06 | 11.6 |