Multimodal Long-Context Understanding on MileBench
51.5Object ExistenceFull Attention
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Full AttentionBackbone=LLaVA-1.5-7B2026.04 | 51.5 | 24 | 43.5 | 42 | 15.8 | 14.7 | 40.5 | 33.14 | — | |
| SinkRouterBackbone=LLaVA-1.5-7B2026.04 | 51.5 | 24 | 44 | 42 | 15.8 | 15.9 | 37 | 32.89 | -0.25 | |
| FastVBackbone=LLaVA-1.5-7B2026.04 | 51 | 26.5 | 43 | 41.5 | 16.2 | 15.3 | 41.5 | 33.57 | 0.43 | |
| LOOK-MBackbone=LLaVA-1.5-7B2026.04 | 50.5 | 24 | 41 | 38.5 | 17.1 | 11.9 | 9.5 | 27.5 | -5.64 | |
| FastVBackbone=LLaVA-1.5-13B2026.04 | 48.5 | 32.5 | 51.5 | 52.5 | 15.2 | 14.8 | 45 | 37.14 | 0.5 | |
| Full AttentionBackbone=LLaVA-1.5-13B2026.04 | 47.5 | 31.5 | 51 | 51.5 | 14.8 | 15.2 | 45 | 36.64 | — | |
| LOOK-MBackbone=LLaVA-1.5-13B2026.04 | 47.5 | 30 | 50.5 | 51.5 | 13.6 | 13.5 | 9.5 | 30.87 | -5.77 | |
| SinkRouterBackbone=LLaVA-1.5-13B2026.04 | 47.5 | 31.5 | 51 | 51.5 | 15.4 | 16.6 | 44 | 36.79 | 0.15 |