Compositional Vision-Language Evaluation on VL-CheckList-VG
71.29VLC Object AccuracyOurs50
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ours50Data curation strategy=CPI-curated 50% subset2026.05 | 71.29 | 57.53 | 39.62 | 67.27 | |
| Full-Random50Data curation strategy=50% random subset2026.05 | 71.15 | 56.75 | 39.15 | 67.06 | |
| Align-only top50Data curation strategy=Global alignment-based 50% subset2026.05 | 70.78 | 57.49 | 38.62 | 66.73 | |
| FullData curation strategy=Full dataset2026.05 | 70.68 | 56.67 | 37.71 | 66.51 |