Long Video Understanding on MLVU v1.0 (test)
67.77MLVU ScoreFLoC
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| FLoCModel=Qwen2-VL-7B, Ratio=2^-3, Tokens=4320, Frames=2562025.10 | 67.77 | — | — | |
| FLoCModel=Qwen2-VL-7B, Ratio=2^-4, Tokens=2160, Frames=2562025.10 | 66.94 | — | — | |
| Full (No Compression)Model=Qwen2-VL-7B, Ratio=100%, Tokens=34560, Frames=2562025.10 | 64.69 | — | — | |
| FLoCModel=Qwen2-VL-7B, Ratio=2^-5, Tokens=1080, Frames=2562025.10 | 64.23 | — | — | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-3, Tokens=4320, Frames=2562025.10 | 62.3 | — | — | |
| Full (No Compression)Model=Qwen2-VL-2B, Ratio=100%, Tokens=34560, Frames=2562025.10 | 62.25 | — | — | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-4, Tokens=2160, Frames=2562025.10 | 60.74 | — | — | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-5, Tokens=1080, Frames=2562025.10 | 59.31 | — | — | |
| Full CacheBackbone=InternVL3.5-8B, KV Budget=Full2026.04 | 51.28 | 67.46 | — | |
| SnapKVBackbone=InternVL3.5-8B, KV Budget=10242026.04 | 51 | 66.68 | 0.78 | |
| KeyDiffBackbone=InternVL3.5-8B, KV Budget=20482026.04 | 50.4 | 65.23 | 2.23 | |
| KeyDiffBackbone=InternVL3.5-8B, KV Budget=10242026.04 | 50.4 | 62.76 | 4.7 | |
| SnapKVBackbone=InternVL3.5-8B, KV Budget=40962026.04 | 50 | 66.5 | 0.96 | |
| Full CacheBackbone=InternVL3.5-14B, KV Budget=Full2026.04 | 49.67 | 68.95 | — | |
| SnapKVBackbone=InternVL3.5-8B, KV Budget=20482026.04 | 49.61 | 66.5 | 0.96 | |
| KeyDiffBackbone=InternVL3.5-8B, KV Budget=40962026.04 | 49.6 | 64.03 | 3.43 | |
| Full CacheBackbone=Qwen2.5-VL-7B, KV Budget=Full2026.04 | 48.8 | 65.52 | — | |
| Full CacheBackbone=Qwen2.5-VL-32B, KV Budget=Full2026.04 | 48.4 | 70.99 | — | |
| SnapKVBackbone=InternVL3.5-14B, KV Budget=10242026.04 | 47.61 | 63.26 | 7.73 | |
| SnapKVBackbone=Qwen2.5-VL-32B, KV Budget=10242026.04 | 47.61 | 63.26 | 7.73 | |
| KeyDiffBackbone=Qwen2.5-VL-7B, KV Budget=40962026.04 | 47.41 | 64.47 | 1.05 | |
| KeyDiffBackbone=Qwen2.5-VL-7B, KV Budget=20482026.04 | 46.41 | 60.62 | 4.9 | |
| KeyDiffBackbone=InternVL3.5-14B, KV Budget=10242026.04 | 46.41 | 60.01 | 8.84 | |
| SnapKVBackbone=Qwen2.5-VL-7B, KV Budget=10242026.04 | 46.02 | 54.46 | 11.06 | |
| KeyDiffBackbone=Qwen2.5-VL-32B, KV Budget=10242026.04 | 45.82 | 59.96 | 11.03 | |
| SnapKVBackbone=Qwen2.5-VL-7B, KV Budget=20482026.04 | 45.42 | 61.31 | 4.21 | |
| SnapKVBackbone=Qwen2.5-VL-7B, KV Budget=40962026.04 | 44.82 | 64.28 | 1.24 | |
| KeyDiffBackbone=Qwen2.5-VL-7B, KV Budget=10242026.04 | 43.43 | 55.82 | 9.7 |