Visual Grounding on TreeBench
24.6Error RateAP-GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AP-GRPOBase model=Qwen2.5-VL-7B, Training=Post-Training2026.02 | 24.6 | 75.4 | |
| MGPOBase model=Qwen2.5-VL-7B, Training=Post-Training2026.02 | 48.7 | 51.3 | |
| GRPOBase model=Qwen2.5-VL-7B, Training=Post-Training2026.02 | 49 | 51 | |
| Qwen2.5-VL-7B2026.02 | 49.8 | 50.2 | |
| LLaVA-OneVision-7B2026.02 | 61.7 | 38.3 | |
| InternVL3-8B2026.02 | 84.9 | 15.1 |