High-Resolution Visual Understanding on HR-Bench-4K (Pass@1)
0.833Pass@1SFT + AXPO
Evaluation results
| Method | Links | |
|---|---|---|
| SFT + AXPOBase Model=Qwen3-VL-8B-Thinking2026.05 | 0.833 | |
| PyVision-RLBase Model=Qwen2.5-VL-7B2026.05 | 0.781 | |
| DeepEyes-v2Base Model=Qwen2.5-VL-7B2026.05 | 0.779 | |
| Mini-o3Base Model=Qwen2.5-VL-7B2026.05 | 0.775 | |
| ThymeBase Model=Qwen2.5-VL-7B2026.05 | 0.77 | |
| DeepEyesBase Model=Qwen2.5-VL-7B2026.05 | 0.751 | |
| PixelReasonerBase Model=Qwen2.5-VL-7B2026.05 | 0.74 | |
| Qwen3-VL-8B-Thinking (Agent)Base Model=Qwen3-VL-8B-Thinking2026.05 | 0.728 | |
| Qwen2.5-VL-7B-InstructBase Model=Qwen2.5-VL-7B2026.05 | 0.716 |