Referring-expression grounding on RefCOCO+ Avg
85.8Accuracy @ IoU 0.5Visual Para-Thinker++ (3B)
Evaluation results
| Method | Links | |
|---|---|---|
| Visual Para-Thinker++ (3B)Backbone=Qwen2.5-VL-3B2026.06 | 85.8 | |
| Visual Para-Thinker (3B)Backbone=Qwen2.5-VL-3B2026.06 | 82.6 | |
| SequentialBackbone=Qwen2.5-VL-3B2026.06 | 79.7 | |
| Qwen2.5-VL-3BBackbone=Qwen2.5-VL-3B2026.06 | 76.7 |