Visual Perception on CV-Bench 2D 1.0 (test)
59.25AccuracyFull Attention
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Full AttentionBackbone=LLaVA-1.5-13B2026.04 | 59.25 | 60.69 | |
| SinkRouterBackbone=LLaVA-1.5-13B2026.04 | 58.97 | 60.54 | |
| SinkRouterBackbone=LLaVA-1.5-7B2026.04 | 47.15 | 53.82 | |
| Full AttentionBackbone=LLaVA-1.5-7B2026.04 | 46.87 | 54.05 |