ScreenSpot-v2
Multimodal · 2024-10-30
ScreenSpot-v2, released with OS-Atlas, is a corrected version of the ScreenSpot GUI-grounding benchmark: given a mobile, desktop or web screenshot and a natural-language instruction, the model must point at the target element, and a click inside its bounding box counts as correct. Scores report grounding accuracy as reported by each release. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi-VL-A3B-Instruct | 92.8 |
| Kimi-VL-A3B-Thinking-2506 | 91.4 |
| Qwen2.5-VL-7B-Instruct | 84.2 |
| Nemotron Nano 12B V2 VL BF16 | 42.8 |
| GPT-4o Mini | 6.9 |