Video Understanding on MVBench, EgoSchema, NExT-QA, and VideoMME Aggregate
56.4Aggregate ScoreLLaVA-OneVision
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LLaVA-OneVisionRelative token budget=100%, Prefilling FLOPs (T)=40.8, FLOPs Ratio=100%2026.04 | 56.4 | 100 | |
| HieraVidRelative token budget=30%, Prefilling FLOPs (T)=10.7, FLOPs Ratio=26.3%2026.04 | 56.1 | 99.3 | |
| FramefusionRelative token budget=30%, Prefilling FLOPs (T)=10.6, FLOPs Ratio=25.9%2026.04 | 55.6 | 98.6 | |
| HieraVidRelative token budget=20%, Prefilling FLOPs (T)=7.5, FLOPs Ratio=18.5%2026.04 | 55.2 | 97.8 | |
| PruneVidRelative token budget=30%, Prefilling FLOPs (T)=10.6, FLOPs Ratio=25.9%2026.04 | 54.9 | 97.2 | |
| FramefusionRelative token budget=20%, Prefilling FLOPs (T)=6.9, FLOPs Ratio=16.9%2026.04 | 54.2 | 96.1 | |
| FastVRelative token budget=30%, Prefilling FLOPs (T)=12.7, FLOPs Ratio=31.2%2026.04 | 53.9 | 95.5 | |
| PruMergeRelative token budget=30%, Prefilling FLOPs (T)=10.6, FLOPs Ratio=25.9%2026.04 | 53.8 | 95.3 | |
| PruneVidRelative token budget=20%, Prefilling FLOPs (T)=6.9, FLOPs Ratio=16.9%2026.04 | 53.6 | 94.9 | |
| HieraVidRelative token budget=10%, Prefilling FLOPs (T)=3.8, FLOPs Ratio=9.3%2026.04 | 53.5 | 94.8 | |
| PruneVidRelative token budget=10%, Prefilling FLOPs (T)=4.3, FLOPs Ratio=10.6%2026.04 | 52.3 | 92.7 | |
| FramefusionRelative token budget=10%, Prefilling FLOPs (T)=3.4, FLOPs Ratio=8.4%2026.04 | 52 | 92.1 | |
| PruMergeRelative token budget=20%, Prefilling FLOPs (T)=6.9, FLOPs Ratio=16.9%2026.04 | 51.6 | 91.4 | |
| FastVRelative token budget=20%, Prefilling FLOPs (T)=9.3, FLOPs Ratio=22.8%2026.04 | 51.5 | 91.2 | |
| StreamingLLMRelative token budget=30%, Prefilling FLOPs (T)=30.6, FLOPs Ratio=75%2026.04 | 51.1 | 90.6 | |
| PruMergeRelative token budget=10%, Prefilling FLOPs (T)=3.4, FLOPs Ratio=8.3%2026.04 | 49.2 | 87.2 | |
| FastVRelative token budget=10%, Prefilling FLOPs (T)=6.037, FLOPs Ratio=14.8%2026.04 | 49.1 | 87 | |
| StreamingLLMRelative token budget=20%, Prefilling FLOPs (T)=29.5, FLOPs Ratio=72.3%2026.04 | 48.6 | 86.2 | |
| StreamingLLMRelative token budget=10%, Prefilling FLOPs (T)=26.9, FLOPs Ratio=65.9%2026.04 | 46.9 | 83.2 |