ScreenSpot
Agents · 2024-01-17
ScreenSpot is a GUI grounding benchmark of about 1,200 instructions over mobile, desktop and web screenshots; a prediction counts when the predicted click point falls inside the target element. ScreenSpot-Pro is a separate, harder benchmark.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen2.5 VL 32B Instruct | 88.5 |
| Qwen2.5-VL-72B-Instruct | 87.1 |
| Qwen2.5-VL-7B-Instruct | 84.7 |
| Qwen2.5-VL-3B-Instruct | 55.5 |