Visual Patch Grounding on MedSG-Bench n=1000 (VPG)
30.81IoUQwen2.5-VL-7B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5-VL-7BPipeline=PDF→code (direct), max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 30.81 | 9.8 | |
| Qwen2.5-VL-7BPipeline=PDF→CT→code, max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 30.17 | 9.7 | |
| Qwen2.5-VL-7BPipeline=Reported2026.05 | 28.87 | 5.7 | |
| Qwen2.5-VL-72BPipeline=PDF→code (direct), max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 28.08 | 7 | |
| Qwen2.5-VL-3BPipeline=Reported2026.05 | 27.36 | 3.4 | |
| Qwen2.5-VL-3BPipeline=PDF→code (direct), max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 27.22 | 3.2 | |
| Qwen2.5-VL-32BPipeline=PDF→CT→code, max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 27.2 | 7.8 | |
| Qwen2.5-VL-32BPipeline=PDF→code (direct), max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 27.2 | 7.8 | |
| Qwen2.5-VL-32BPipeline=Reported2026.05 | 26.92 | 7.1 | |
| Qwen2.5-VL-3BPipeline=PDF→CT→code, max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 26.83 | 3.1 | |
| Qwen2.5-VL-72BPipeline=Reported2026.05 | 26.45 | 6.3 | |
| Qwen2.5-VL-72BPipeline=PDF→CT→code, max_new_tokens=128, zero-shot=true, precision=bf16, batch size=1, do_sample=False2026.05 | 14.39 | 2 |