Visual Question Answering on OCRBench
909ScoreQwen3-VL-32B-Instruct + SpatialBoost
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3-VL-32B-Instruct + SpatialBoostEncoder Enhancement=SpatialBoost2026.03 | 909 | — | — | — | |
| Qwen3-VL-32B-Instruct2026.03 | 895 | — | — | — | |
| InternVL3-38B + SpatialBoostEncoder Enhancement=SpatialBoost2026.03 | 894 | — | — | — | |
| InternVL3-38B2026.03 | 886 | — | — | — | |
| Baseline (No Recompute)Model=Qwen3-VL-8B, Recomputation budget (k)=02026.03 | 878 | — | — | — | |
| BaseRatio=100%, Model Backbone=Qwen3-VL-4B2026.05 | 874 | 81.95 | 100 | — | |
| Qwen3-VL-2BParameter Count=2B2025.12 | 858 | — | — | — | |
| CacheBlendModel=Qwen3-VL-8B, Recomputation budget (k)=22026.03 | 845 | — | — | — | |
| InfoFlow KVModel=Qwen3-VL-8B, Recomputation budget (k)=22026.03 | 842 | — | — | — | |
| No RecomputeModel=Qwen3-VL-8B, Recomputation budget (k)=22026.03 | 839 | — | — | — | |
| InternVL3.5-2BParameter Count=2B2025.12 | 836 | — | — | — | |
| EPICModel=Qwen3-VL-8B, Recomputation budget (k)=22026.03 | 836 | — | — | — | |
| InternVL3-2BParameter Count=2B2025.12 | 835 | — | — | — | |
| BaseRatio=100%, Backbone=Qwen2.5-VL-3B2026.05 | 824 | — | — | — | |
| Qwen2-VL-2BParameter Count=2B2025.12 | 809 | — | — | — | |
| InfoFlow KVModel=Qwen3-VL-8B, Recomputation budget (k)=42026.03 | 802 | — | — | — | |
| CacheBlendModel=Qwen3-VL-8B, Recomputation budget (k)=42026.03 | 786 | — | — | — | |
| EPICModel=Qwen3-VL-8B, Recomputation budget (k)=42026.03 | 785 | — | — | — | |
| ERASERatio=-65.11%, Model Backbone=Qwen3-VL-4B2026.05 | 784 | 78.98 | 96.38 | — | |
| No RecomputeModel=Qwen3-VL-8B, Recomputation budget (k)=42026.03 | 781 | — | — | — | |
| jina-vlm2025.12 | 778 | — | — | — | |
| IVC-PruneRatio=-65.11%, Model Backbone=Qwen3-VL-4B2026.05 | 749 | 76.81 | 93.74 | — | |
| ERASERatio=-72.77%, Model Backbone=Qwen3-VL-4B2026.05 | 745 | 77.1 | 94.09 | — | |
| DARTRatio=-65.11%, Model Backbone=Qwen3-VL-4B2026.05 | 712 | 69.49 | 84.8 | — | |
| ERASERatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.05 | 706 | — | — | — | |
| IVC-PruneRatio=-72.77%, Model Backbone=Qwen3-VL-4B2026.05 | 695 | 75.09 | 91.64 | — | |
| PruneSIDRatio=-65.11%, Model Backbone=Qwen3-VL-4B2026.05 | 660 | 66.97 | 81.72 | — | |
| DARTRatio=-72.77%, Model Backbone=Qwen3-VL-4B2026.05 | 656 | 63.72 | 77.75 | — | |
| CDPrunerRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.05 | 654 | — | — | — | |
| ERASERatio=-84.43%, Model Backbone=Qwen3-VL-4B2026.05 | 632 | 73.25 | 89.39 | — | |
| DARTRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.05 | 626 | — | — | — | |
| CDPrunerRatio=-65.11%, Model Backbone=Qwen3-VL-4B2026.05 | 619 | 60.84 | 74.24 | — | |
| IVC-PruneRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.05 | 617 | — | — | — | |
| ERASERatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.05 | 611 | — | — | — | |
| CDPrunerRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.05 | 591 | — | — | — | |
| PruneSIDRatio=-62.1%, Backbone=Qwen2.5-VL-3B2026.05 | 587 | — | — | — | |
| PruneSIDRatio=-72.77%, Model Backbone=Qwen3-VL-4B2026.05 | 583 | 63.72 | 77.75 | — | |
| CDPrunerRatio=-72.77%, Model Backbone=Qwen3-VL-4B2026.05 | 534 | 55.15 | 67.3 | — | |
| IVC-PruneRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.05 | 527 | — | — | — | |
| DARTRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.05 | 519 | — | — | — | |
| CDPrunerRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.05 | 515 | — | — | — | |
| DARTRatio=-84.43%, Model Backbone=Qwen3-VL-4B2026.05 | 510 | 50.73 | 61.91 | — | |
| IVC-PruneRatio=-84.43%, Model Backbone=Qwen3-VL-4B2026.05 | 494 | 66.15 | 80.73 | — | |
| PruneSIDRatio=-75.0%, Backbone=Qwen2.5-VL-3B2026.05 | 494 | — | — | — | |
| ERASERatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.05 | 493 | — | — | — | |
| PruneSIDRatio=-84.43%, Model Backbone=Qwen3-VL-4B2026.05 | 454 | 51.54 | 62.9 | — | |
| PruneSIDRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.05 | 426 | — | — | — | |
| DARTRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.05 | 413 | — | — | — | |
| CDPrunerRatio=-84.43%, Model Backbone=Qwen3-VL-4B2026.05 | 405 | 43.71 | 53.34 | — | |
| IVC-PruneRatio=-85.0%, Backbone=Qwen2.5-VL-3B2026.05 | 396 | — | — | — | |
| SeProD2026.05 | 85.3 | — | — | — | |
| Mini-o32026.05 | 83.8 | — | — | — | |
| Qwen2.5-VL2026.05 | 81.5 | — | — | — | |
| DeepSeek-OCR2026.07 | — | — | — | 430 | |
| Gemini 2.5 Pro2026.07 | — | — | — | 872 | |
| HunyuanOCR2026.07 | — | — | — | 860 | |
| HunyuanOCR-1.52026.07 | — | — | — | 861 | |
| Qwen3-VL-235B-A22B-InstructModel Size=235B-A22B2026.07 | — | — | — | 920 | |
| Qwen3-VL-2B-InstructModel Size=2B2026.07 | — | — | — | 858 | |
| Seed-1.6-Vision2026.07 | — | — | — | 881 |