Long Video Reasoning on WorldSense
52.5Overall AccuracyFull Tokens
Evaluation Results
| Method | Links | |
|---|---|---|
| Full TokensModel=Qwen3-Omni-30B-A3B, Retained Ratio=100%, Frames=322026.05 | 52.5 | |
| OMAC (Ours)Model=Qwen3-Omni-30B-A3B, Retained Ratio=30%, Frames=322026.05 | 48 | |
| OmniZipModel=Qwen3-Omni-30B-A3B, Retained Ratio=30%, Frames=322026.05 | 47.4 | |
| OmniSIFT*Model=Qwen2.5-Omni-7B, Retained Ratio=30%, Frames=32, Training Data=100k+2026.05 | 44.4 | |
| O-MARC*Model=Qwen2.5-Omni-7B, Retained Ratio=30%, Frames=32, Training Data=100k+2026.05 | 44 | |
| Full TokensModel=Qwen2.5-Omni-7B, Retained Ratio=100%, Frames=322026.05 | 43.6 | |
| O-MARCModel=Qwen2.5-Omni-7B, Retained Ratio=30%, Frames=32, Training Data=3k+2026.05 | 43.4 | |
| OMAC (Ours)Model=Qwen2.5-Omni-7B, Retained Ratio=30%, Frames=322026.05 | 42.4 | |
| Full TokensModel=Qwen2.5-Omni-3B, Retained Ratio=100%, Frames=322026.05 | 42.2 | |
| O-MARCModel=Qwen2.5-Omni-3B, Retained Ratio=30%, Frames=32, Training Data=3k+2026.05 | 42.1 | |
| OMAC (Ours)Model=Qwen2.5-Omni-3B, Retained Ratio=30%, Frames=322026.05 | 41.2 | |
| OmniZipModel=Qwen2.5-Omni-3B, Retained Ratio=30%, Frames=322026.05 | 39.7 | |
| OmniZipModel=Qwen2.5-Omni-7B, Retained Ratio=30%, Frames=322026.05 | 39.6 |