Visual Grounding on ScreenSpot-Pro 1,097-sample subset
0.599AUCKV-Ground-8B†
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| KV-Ground-8B†n=1,097, Precision=bf16, Decoding strategy=greedy, Zoom steps=2-step, Crop ratio r=0.52026.04 | 0.599 | 0.123 | 4 | |
| Qwen3.5-27Bn=1,097, Quantization=AWQ-4bit, Decoding strategy=greedy, Zoom steps=2-step, Crop ratio r=0.52026.04 | 0.588 | 0.109 | 0.0003 |