Aerial image reasoning on HRVQA-VL 1.0 (test)
86.38Stage A AccuracyUAV-VL-R1
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| UAV-VL-R1Training Stage=Stage A only, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 86.38 | 48.37 | 46.59 | 60.44 | |
| UAV-VL-R1Training Stage=All stages, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 86.27 | 63.4 | 64.22 | 71.3 | |
| UAV-VL-R1Training Stage=Stage B only, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 64.01 | 62.71 | 25.82 | 50.84 | |
| UAV-VL-R1Training Stage=Stage C only, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 61.66 | 38.06 | 49.37 | 49.69 | |
| Qwen2-VL-72B-InstructEvaluation Mode=Zero-shot, Prompting=Structured2025.08 | 56.82 | 51.39 | 40.05 | 49.42 | |
| Qwen2-VL-7B-InstructEvaluation Mode=Zero-shot, Prompting=Structured2025.08 | 49.95 | 34.38 | 18.92 | 34.41 | |
| InstructBLIP-7BEvaluation Mode=Zero-shot, Prompting=Structured2025.08 | 38.77 | 23.21 | 30.5 | 30.82 | |
| Qwen2-VL-2B-InstructVersion=SFT, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 35.82 | 30.61 | 20.98 | 29.13 | |
| Qwen2-VL-2B-InstructVersion=Base, Evaluation Mode=Zero-shot, Prompting=Structured2025.08 | 35.16 | 21.96 | 14.76 | 23.96 |