Rolling-Forcing on LongVBench
84.19VBench ScoreTempCache-Qunt
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| TempCache-QuntCategory=KV compression2026.02 | 84.19 | 24.26 | 0.653 | 0.143 | 16.2 | 91.4 | 6.9 | |
| TempCache-LSHCategory=KV compression2026.02 | 84.17 | 24.13 | 0.651 | 0.149 | 16.8 | 90.2 | 6.8 | |
| FlowCacheCategory=KV compression2026.02 | 84.15 | 22.15 | 0.634 | 0.222 | 82.9 | 86.9 | 2.3 | |
| TeaCacheCategory=KV compression2026.02 | 84.11 | 16.12 | 0.315 | 0.523 | 93.2 | 84.6 | 1.1 | |
| Dense (FlashAttention 3)Category=Dense2026.02 | 84.08 | — | — | — | 100 | 100 | 1 | |
| All Ours-LSHCategory=Full2026.02 | 84.02 | 25.71 | 0.681 | 0.147 | — | — | 10.7 | |
| All Ours-QuantCategory=Full2026.02 | 83.99 | 25.73 | 0.678 | 0.147 | — | — | 10.8 | |
| AnnSA-QuantCategory=Sparse Self-Attention2026.02 | 83.29 | 25.77 | 0.689 | 0.141 | 28 | 92.6 | 5.2 | |
| AnnSA-LSHCategory=Sparse Self-Attention2026.02 | 83.25 | 25.73 | 0.688 | 0.142 | 27.6 | 92.4 | 5.1 | |
| AnnCA-LSHCategory=Sprase Cross-Attention2026.02 | 83.23 | 25.68 | 0.679 | 0.155 | 33.1 | 94.2 | 2.2 | |
| AnnCA-QuantCategory=Sprase Cross-Attention2026.02 | 82.89 | 24.11 | 0.646 | 0.148 | 29.5 | 91.1 | 2.3 | |
| RadialAttnCategory=Sparse Self-Attention2026.02 | 61.51 | 16.87 | 0.289 | 0.702 | 81.6 | 58.3 | 2.8 | |
| FlowCache+RadialAttnCategory=Full2026.02 | 45.15 | 16.98 | 0.294 | 0.687 | — | — | 4.4 | |
| SVG2Category=Sparse Self-Attention2026.02 | 34.66 | 14.29 | 0.226 | 0.736 | 68.8 | 38.9 | 0.2 | |
| SVG1Category=Sparse Self-Attention2026.02 | 33.15 | 14.22 | 0.201 | 0.744 | 76.1 | 30.2 | 0.1 |