Visual Question Answering on In-Domain Multiple-Choice
92.9OCR-VQA AccuracyEXTRACT+THINK
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| EXTRACT+THINKPerception Module (P)=1.7B, Reasoning Module (R)=4.0B, #Vis. Data=2.4M2025.11 | 92.9 | 90.1 | 75.2 | 84.4 | 77.8 | 91.3 | 85.3 | |
| PrismCaptionerPerception Module (P)=7.0B, Reasoning Module (R)=70B, #Vis. Data=1.9M2025.11 | 91.5 | 77 | 68.1 | 79.9 | 67.5 | 85.8 | 78.3 | |
| EXTRACT+THINK†Perception Module (P)=1.7B, Reasoning Module (R)=4.0B, #Vis. Data=0.4M2025.11 | 91.5 | 84 | 71.3 | 84.6 | 77.8 | 86.9 | 82.7 | |
| EXTRACT+THINKPerception Module (P)=0.6B, Reasoning Module (R)=1.7B, #Vis. Data=2.4M2025.11 | 89.4 | 81.8 | 72.2 | 78 | 74.7 | 85.6 | 80.3 | |
| PrismCaptionerPerception Module (P)=1.8B, Reasoning Module (R)=70B, #Vis. Data=1.9M2025.11 | 89.2 | 72.7 | 64.6 | 77.8 | 66 | 82.3 | 75.4 | |
| CAPTION+THINKPerception Module (P)=1.7B, Reasoning Module (R)=4.0B, #Vis. Data=2.0M2025.11 | 89.2 | 84.8 | 68.9 | 80.5 | 72.1 | 84.3 | 80 | |
| EXTRACT+THINK†Perception Module (P)=0.6B, Reasoning Module (R)=1.7B, #Vis. Data=0.4M2025.11 | 86.9 | 79.8 | 69.9 | 76.6 | 72.5 | 82.1 | 78 | |
| CAPTION+THINKPerception Module (P)=0.6B, Reasoning Module (R)=1.7B, #Vis. Data=2.0M2025.11 | 84.9 | 80.6 | 60.6 | 74.7 | 66.2 | 83 | 75 | |
| InternVL2.5LLM Size=0.5B, #Vis. Data=64M2025.11 | 79.8 | 89.1 | 89.8 | 82 | 75.4 | 83 | 83.2 | |
| Baseline (from §3.3)Perception Module (P)=1.7B, Reasoning Module (R)=4.0B, #Vis. Data=1.0M2025.11 | 79.4 | 59.4 | 65 | 71.6 | 64.5 | 76.4 | 69.4 | |
| Baseline (from §3)LLM Size=1.7B, #Vis. Data=1.0M2025.11 | 73.4 | 83.4 | 76.2 | 77.8 | 74.3 | 75.8 | 76.8 | |
| SmolVLMLLM Size=1.7B, #Vis. Data=unk.2025.11 | 72.9 | 81.4 | 79.7 | 75.5 | 70.6 | 75.1 | 75.9 | |
| Baseline (from §3.3)Perception Module (P)=0.6B, Reasoning Module (R)=4.0B, #Vis. Data=1.0M2025.11 | 71.8 | 50.7 | 63 | 67.6 | 62.3 | 72.3 | 64.6 | |
| LLaVA-OneVisionLLM Size=0.5B, #Vis. Data=8.8M2025.11 | 69.5 | 77.2 | 55.7 | 75.7 | 73.6 | 74.7 | 71.1 | |
| Baseline (from §3)LLM Size=0.6B, #Vis. Data=1.0M2025.11 | 41.1 | 71.3 | 67.9 | 71.2 | 69.5 | 74.5 | 65.9 |