RefCOCO, RefCOCO+ and RefCOCOg eight-split average
Multimodal
Referring-expression grounding averaged over eight standard splits: RefCOCO val/testA/testB, RefCOCO+ val/testA/testB and RefCOCOg (umd) val/test. Each split scores box accuracy at IoU >= 0.5 (P@1), and the reported number is the unweighted mean of the eight split accuracies, a release-card convention rather than an aggregate defined by the dataset authors.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-VL-2B | 90.5 |
| LFM2.5-VL-3B | 87.9 |
| Qwen3.5-4B | 87.4 |
| Gemma 4 12B IT | 82.2 |
| GPT-5.6 Luna | 81.8 |
| Gemini 3.5 Flash-Lite | 81.6 |
| Qwen3.5-2B | 78.3 |
| Gemma 4 E2B IT | 68.0 |