Computer Vision Reasoning on CV-Bench 3D
82.9AccuracyQwen2.5-VL 7B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen2.5-VL 7BType=Generalist Baseline2026.05 | 82.9 | |
| SpatialReasonerType=Specialist Reference2026.05 | 80.3 | |
| w/ Entropy-RegularizationEvaluation Protocol=RL Fine-tuning2026.05 | 78.8 | |
| Mini-o3Mode=Zero-shot, Type=Visual Agents2026.05 | 77.6 | |
| DeepEyesMode=Zero-shot, Type=Visual Agents2026.05 | 76.7 | |
| vanilla RFT (Base)Evaluation Protocol=RL Fine-tuning2026.05 | 76.7 | |
| BLIP-3-4BModel Scale=4B, Model Access Type=Open-source2025.03 | 75.4 | |
| w/ Tool-Encourage RewardEvaluation Protocol=RL Fine-tuning2026.05 | 74.5 | |
| AKI-4BModel Scale=4B, Model Access Type=Open-source2025.03 | 71.8 | |
| Phi-3-Vision-4BModel Scale=4B, Model Access Type=Open-source2025.03 | 68.2 | |
| VILA-1.5-3BModel Scale=3B, Model Access Type=Open-source2025.03 | 60.3 |