Video Understanding on MLVU (Score, Avg Score, Acc. %)
102.8AccuracySCORE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SCORERetention Ratio R=25%, Max Frames=1282026.03 | 102.8 | 70.1 | 58.9 | |
| SCORERetention Ratio R=40%, Max Frames=1282026.03 | 102.4 | 70.6 | 58.7 | |
| VidCom^2Retention Ratio R=40%, Max Frames=1282026.03 | 102.1 | 70.6 | 58.5 | |
| DyCokeRetention Ratio R=40%, Max Frames=1282026.03 | 101.7 | 70 | 58.2 | |
| VidCom^2Retention Ratio R=25%, Max Frames=1282026.03 | 101.6 | 68.7 | 58.2 | |
| FastVIDRetention Ratio R=40%, Max Frames=1282026.03 | 100.3 | 68.8 | 57.5 | |
| VanillaRetention Ratio R=100%, Max Frames=1282026.03 | 100 | 69.7 | 57.3 | |
| FastVIDRetention Ratio R=25%, Max Frames=1282026.03 | 99.8 | 68.4 | 57.1 | |
| SCORERetention Ratio R=10%, Max Frames=1282026.03 | 99.5 | 67.7 | 57 | |
| HoliTomRetention Ratio R=40%, Max Frames=1282026.03 | 99.3 | 69.9 | 56.9 | |
| HoliTomRetention Ratio R=25%, Max Frames=1282026.03 | 97.1 | 68.4 | 55.6 | |
| FastVIDRetention Ratio R=10%, Max Frames=1282026.03 | 95.9 | 65 | 54.9 | |
| VidCom^2Retention Ratio R=10%, Max Frames=1282026.03 | 95.3 | 63.4 | 54.6 | |
| DyCokeRetention Ratio R=25%, Max Frames=1282026.03 | 94.8 | 64.6 | 54.3 | |
| HoliTomRetention Ratio R=10%, Max Frames=1282026.03 | 92.3 | 65.2 | 52.8 | |
| QCALLM Size=30B, Frames=64, Backbone=Qwen3-VL-30B-A3B2026.07 | 77.1 | — | — | |
| InternVL3.5-38BLLM Size=38B, Frames=642026.07 | 77 | — | — | |
| QCALLM Size=8B, Frames=64, Backbone=Qwen3-VL-8B2026.07 | 75.7 | — | — | |
| Qwen2.5-VL-72BLLM Size=72B, Frames=1fps2026.07 | 74.6 | — | — | |
| LLaVA-Video-72BLLM Size=72B, Frames=642026.07 | 74.4 | — | — | |
| Kimi-VL-16B-A3BLLM Size=16B, Frames=642026.07 | 74.2 | — | — | |
| Qwen3-VL-30B-A3BLLM Size=30B, Frames=642026.07 | 72.8 | — | — | |
| Qwen3-VL-8BLLM Size=8B, Frames=642026.07 | 71 | — | — | |
| NVILALLM Size=7B, Frames=2562026.07 | 70.1 | — | — | |
| mPLUG-Owl3LLM Size=8B, Frames=1282026.07 | 70 | — | — | |
| ProtoKVBudget |M|=6K, Backbone=Qwen2-VL-7B2026.06 | 68.9 | — | — | |
| ApolloLLM Size=7B, Frames=2fps2026.07 | 68.7 | — | — | |
| InfiniPot-VBudget |M|=6K, Backbone=Qwen2-VL-7B2026.06 | 68.4 | — | — | |
| InfiniPot-VBudget |M|=3K, Backbone=Qwen2-VL-7B2026.06 | 67.2 | — | — | |
| ProtoKVBudget |M|=3K, Backbone=Qwen2-VL-7B2026.06 | 67 | — | — | |
| GPT-4oLLM Size=-, Frames=256 / 0.5fps2026.07 | 64.6 | — | — | |
| DyCokeBudget |M|=6K, Backbone=Qwen2-VL-7B2026.06 | 60.5 | — | — | |
| DyCokeBudget |M|=3K, Backbone=Qwen2-VL-7B2026.06 | 57.5 | — | — | |
| Chain-of-Time Reasoning (Ours)2026.04 | — | 76.2 | — | |
| InternVideo2.5-8BModel Size=8B2026.04 | — | 72.8 | — | |
| InternVL2.5-4BModel Size=4B2026.04 | — | 48.3 | — | |
| Qwen2.5-VL-3B-InstructModel Size=3B2026.04 | — | 68.2 | — | |
| Qwen3-VI-8B-InstructModel Size=8B2026.04 | — | 78.1 | — | |
| Qwen3-VL-4B-InstructModel Size=4B2026.04 | — | 75.3 | — | |
| SF-LLaVA-1.5-3BModel Size=3B2026.04 | — | 68.8 | — | |
| Video-R1-7BModel Size=7B2026.04 | — | 58.9 | — | |
| VideoLLaMA3-7BModel Size=7B2026.04 | — | 68.7 | — |