Overall Performance on Average across all benchmarks
70.3Average ScoreAR-OPD
Evaluation Results
| Method | Links | |
|---|---|---|
| AR-OPD2026.06 | 70.3 | |
| Full OPD2026.06 | 67.9 | |
| Partial OPD2026.06 | 65.2 | |
| SFT2026.06 | 62.4 | |
| LaVerVisual Encoder Model=Qwen-ViT2025.12 | 59.94 | |
| Base2026.06 | 58.2 | |
| BaselineVisual Encoder Model=Qwen-ViT2025.12 | 57.96 | |
| LaVerVisual Encoder Model=SigLIP22025.12 | 57.87 | |
| LaVerVisual Encoder Model=AIMv22025.12 | 55.86 | |
| BaselineVisual Encoder Model=SigLIP22025.12 | 55.72 | |
| BaselineVisual Encoder Model=AIMv22025.12 | 54.7 | |
| LaVerVisual Encoder Model=CLIP2025.12 | 53.24 | |
| DSRMidtraining Schedule=DSR2026.05 | 51.46 | |
| BaselineVisual Encoder Model=CLIP2025.12 | 50.58 | |
| WSDMidtraining Schedule=WSD2026.05 | 49.76 | |
| LaVerVisual Encoder Model=DINOv22025.12 | 49.23 | |
| Unified 15B MoEMidtraining Corpus=1k→8k2026.05 | 49.07 | |
| Cosine-decayMidtraining Schedule=Cos2026.05 | 48.48 | |
| BaselineVisual Encoder Model=DINOv22025.12 | 47.92 | |
| Unified 15B MoEMidtraining Corpus=4k→8k2026.05 | 47.71 | |
| Unified 15B MoEMidtraining Corpus=4k2026.05 | 46.68 | |
| LaVerVisual Encoder Model=MLP + Qwen 2.52025.12 | 37.74 | |
| BaselineVisual Encoder Model=MLP + Qwen 2.52025.12 | 36.37 | |
| Unified 15B MoEPretraining Corpus=4k→8k2026.05 | 33.79 | |
| Unified 15B MoEPretraining Corpus=1k→8k2026.05 | 33.24 | |
| Unified 15B MoEPretraining Corpus=4k2026.05 | 32.6 |