Audio Understanding on AudioMarathon
16.8SERFastV
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FastVBackbone=Qwen2.5-Omni-3B, Retain Ratio=10%2026.04 | 16.8 | 54.9 | 3.5 | 65.2 | 95.9 | 55.9 | 32.7 | 14.8 | 86.5 | 70.5 | 29.5 | |
| DARTBackbone=Phi-4-Multimodal, Retain Ratio=70%2026.04 | 16.8 | 67.6 | 57.2 | 54.5 | 46.1 | 31.8 | 28.6 | 27.1 | 91.6 | 93.4 | 6.6 | |
| DARTBackbone=Phi-4-Multimodal, Retain Ratio=10%2026.04 | 16.8 | 49.3 | 0 | 52 | 40.2 | 24.4 | 27.4 | 18.6 | 77.2 | 67.9 | 32.1 | |
| DARTBackbone=Qwen2.5-Omni-3B, Retain Ratio=10%2026.04 | 17.3 | 54.1 | 62.9 | 66.8 | 99.1 | 69.3 | 42.6 | 22.1 | 52.8 | 80.5 | 19.5 | |
| FrameBackbone=Phi-4-Multimodal, Retain Ratio=70%2026.04 | 17.7 | 64.4 | 63.4 | 31.4 | 32.6 | 29 | 31 | 27.4 | 92.4 | 86.4 | 13.6 | |
| FastVBackbone=Qwen2.5-Omni-3B, Retain Ratio=40%2026.04 | 18 | 63.8 | 39.2 | 60.5 | 97.5 | 57.8 | 28.6 | 17.1 | 95.3 | 79 | 21 | |
| DARTBackbone=Phi-4-Multimodal, Retain Ratio=40%2026.04 | 18 | 61.1 | 23.7 | 53.9 | 44.8 | 25.4 | 29.4 | 24.4 | 88 | 81.8 | 18.2 | |
| FastVBackbone=Phi-4-Multimodal, Retain Ratio=70%2026.04 | 18.3 | 64 | 43.9 | 33.2 | 40.6 | 29.6 | 29.1 | 25.7 | 92.9 | 83.6 | 16.4 | |
| VanillaBackbone=Phi-4-Multimodal, Retain Ratio=100%2026.04 | 18.4 | 69.3 | 92.7 | 55.1 | 46.7 | 23.4 | 27.3 | 26.6 | 91.1 | 100 | — | |
| RandomBackbone=Phi-4-Multimodal, Retain Ratio=70%2026.04 | 18.4 | 67.5 | 49.1 | 31.4 | 39.8 | 30.2 | 31 | 24.5 | 93.6 | 85.4 | 14.6 | |
| FastVBackbone=Qwen2.5-Omni-3B, Retain Ratio=70%2026.04 | 18.7 | 68.2 | 76.3 | 61.3 | 98.4 | 57.2 | 31.1 | 17.3 | 97.5 | 86.9 | 13.1 | |
| RandomBackbone=Phi-4-Multimodal, Retain Ratio=40%2026.04 | 18.7 | 61.6 | 7.9 | 30.3 | 27.4 | 30.6 | 29.4 | 20.5 | 91.2 | 70.5 | 29.5 | |
| RandomBackbone=Phi-4-Multimodal, Retain Ratio=10%2026.04 | 18.7 | 35.3 | 0 | 29.3 | 20.6 | 29.6 | 33.4 | 11.1 | 67.5 | 54.5 | 45.5 | |
| FrameBackbone=Qwen2.5-Omni-3B, Retain Ratio=10%2026.04 | 22.8 | 58.2 | 0 | 64.5 | 95 | 60.9 | 41.1 | 18.1 | 87 | 74 | 26 | |
| DARTBackbone=Qwen2.5-Omni-3B, Retain Ratio=40%2026.04 | 23.1 | 64.6 | 62.8 | 71.9 | 99.1 | 73.4 | 37.6 | 28.1 | 46 | 83.8 | 16.2 | |
| DARTBackbone=Qwen2.5-Omni-3B, Retain Ratio=70%2026.04 | 23.2 | 74.2 | 81.4 | 73.1 | 97.6 | 72.5 | 37.1 | 23 | 48.7 | 87.8 | 12.2 | |
| FastVBackbone=Phi-4-Multimodal, Retain Ratio=10%2026.04 | 23.4 | 43 | 0 | 27.6 | 30.1 | 29.3 | 24.9 | 17.3 | 82.9 | 61.7 | 38.3 | |
| FrameBackbone=Phi-4-Multimodal, Retain Ratio=40%2026.04 | 23.8 | 59 | 23.3 | 31.1 | 28.5 | 30 | 30.1 | 22.3 | 87.8 | 74.5 | 25.5 | |
| RandomBackbone=Qwen2.5-Omni-3B, Retain Ratio=10%2026.04 | 24 | 58.1 | 0 | 65.9 | 97.6 | 60 | 41.8 | 17.1 | 84.3 | 74.3 | 25.7 | |
| RandomBackbone=Qwen2.5-Omni-3B, Retain Ratio=40%2026.04 | 24.2 | 75.3 | 59.7 | 68.7 | 95.8 | 68.3 | 37.9 | 27.2 | 93.5 | 91.1 | 8.9 | |
| FrameBackbone=Phi-4-Multimodal, Retain Ratio=10%2026.04 | 24.6 | 36.8 | 0 | 28.4 | 25 | 28.1 | 30.1 | 12.1 | 66.6 | 55.9 | 44.1 | |
| HeadRouterBackbone=Phi-4-Multimodal, Retain Ratio=70%2026.04 | 24.7 | 67 | 51 | 53.5 | 42.5 | 29.4 | 29.2 | 32.8 | 94.5 | 94.2 | -5.8 | |
| VanillaBackbone=Qwen2.5-Omni-3B, Retain Ratio=100%2026.04 | 25.2 | 82.3 | 94.7 | 70.2 | 97.4 | 69.3 | 39.6 | 29.1 | 97.2 | 100 | — | |
| FastVBackbone=Qwen2.5-Omni-7B, Retain Ratio=10%2026.04 | 25.3 | 64.3 | 1.8 | 75.2 | 96.7 | 64.5 | 46.1 | 30.6 | 99.3 | 79.7 | 20.3 | |
| FrameBackbone=Qwen2.5-Omni-3B, Retain Ratio=40%2026.04 | 25.6 | 75.3 | 82.2 | 69 | 100 | 68.3 | 38.6 | 28.3 | 91 | 95.7 | 4.3 | |
| FastVBackbone=Qwen2.5-Omni-7B, Retain Ratio=40%2026.04 | 25.7 | 71.8 | 0 | 78.7 | 98.3 | 71 | 51.1 | 25.4 | 99.3 | 82.4 | 17.6 | |
| HeadRouterBackbone=Phi-4-Multimodal, Retain Ratio=10%2026.04 | 25.7 | 51.5 | 0.2 | 44.5 | 40.8 | 29.8 | 28.7 | 29.4 | 73.8 | 71.9 | -28.1 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Retain Ratio=70%2026.04 | 25.9 | 86.8 | 95.2 | 78 | 99.2 | 73.2 | 43.3 | 50.9 | 99.5 | 103 | 3 | |
| FastVBackbone=Phi-4-Multimodal, Retain Ratio=40%2026.04 | 26.1 | 52.8 | 0 | 32.5 | 28.2 | 30.3 | 28 | 22.2 | 89.4 | 68.7 | 31.3 | |
| VanillaBackbone=Qwen2.5-Omni-7B, Retain Ratio=100%2026.04 | 26.3 | 85.1 | 98.1 | 78.4 | 100 | 72.2 | 53.4 | 21.4 | 98 | 100 | — | |
| RandomBackbone=Qwen2.5-Omni-3B, Retain Ratio=70%2026.04 | 26.5 | 80.3 | 88.4 | 71.1 | 97.5 | 69.7 | 38.4 | 28.6 | 95.7 | 98.5 | 1.5 | |
| DARTBackbone=Qwen2.5-Omni-7B, Retain Ratio=70%2026.04 | 26.5 | 84 | 86.2 | 79.1 | 99.2 | 73.6 | 44.9 | 29 | 99.3 | 98.3 | 1.7 | |
| RandomBackbone=Qwen2.5-Omni-7B, Retain Ratio=40%2026.04 | 26.5 | 81.8 | 69.3 | 77.2 | 99.2 | 72.5 | 44.9 | 28.3 | 99.3 | 94.7 | 5.3 | |
| FrameBackbone=Qwen2.5-Omni-3B, Retain Ratio=70%2026.04 | 26.8 | 80.9 | 92.2 | 70.5 | 98.5 | 70.2 | 36.4 | 31.4 | 96.7 | 99.9 | 0.1 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Retain Ratio=10%2026.04 | 26.9 | 64.2 | 0 | 72.8 | 99.2 | 68.1 | 46.1 | 49 | 99.1 | 83.1 | -16.9 | |
| HeadRouterBackbone=Phi-4-Multimodal, Retain Ratio=40%2026.04 | 26.9 | 61.8 | 14.6 | 52 | 44.2 | 29.6 | 29.8 | 29.1 | 91.3 | 84 | -16 | |
| RandomBackbone=Qwen2.5-Omni-7B, Retain Ratio=70%2026.04 | 27.4 | 87 | 91.5 | 78.7 | 98.3 | 72.5 | 44.4 | 30.3 | 99.5 | 99.4 | 0.6 | |
| FrameBackbone=Qwen2.5-Omni-7B, Retain Ratio=10%2026.04 | 27.4 | 66.1 | 0 | 73.2 | 99.2 | 66.2 | 43.3 | 24.8 | 99.3 | 78.9 | 21.1 | |
| FastVBackbone=Qwen2.5-Omni-7B, Retain Ratio=70%2026.04 | 27.6 | 83.4 | 55.3 | 77.6 | 99.2 | 73 | 47.2 | 27.4 | 99.4 | 93.3 | 6.7 | |
| FrameBackbone=Qwen2.5-Omni-7B, Retain Ratio=70%2026.04 | 27.6 | 87.1 | 94.9 | 79.5 | 99.2 | 71.7 | 41.6 | 29.2 | 99.5 | 99.6 | 0.4 | |
| FrameBackbone=Qwen2.5-Omni-7B, Retain Ratio=40%2026.04 | 27.6 | 83.8 | 82.2 | 78.4 | 99.2 | 71.3 | 39.3 | 25.9 | 99.3 | 95.9 | 4.1 | |
| HeadRouterBackbone=Qwen2.5-Omni-7B, Retain Ratio=40%2026.04 | 28 | 80.4 | 64.3 | 78.7 | 99.2 | 72.2 | 43.8 | 51 | 99.7 | 97.6 | -2.4 | |
| RandomBackbone=Qwen2.5-Omni-7B, Retain Ratio=10%2026.04 | 28 | 65.5 | 0 | 76.4 | 99.2 | 68.9 | 44.4 | 27.1 | 99.4 | 80.4 | 19.6 | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Retain Ratio=70%2026.04 | 28.8 | 82.9 | 96.8 | 73.6 | 100 | 69.4 | 34.3 | 30.6 | 99.4 | 101.8 | 1.8 | |
| DARTBackbone=Qwen2.5-Omni-7B, Retain Ratio=40%2026.04 | 29 | 76.2 | 46.6 | 76.4 | 99.2 | 73.5 | 47.2 | 24.3 | 97.9 | 90.2 | 9.8 | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Retain Ratio=10%2026.04 | 29.2 | 63 | 5.4 | 67.7 | 99.2 | 63.2 | 34.3 | 29.3 | 96.5 | 80.7 | -19.3 | |
| HeadRouterBackbone=Qwen2.5-Omni-3B, Retain Ratio=40%2026.04 | 29.6 | 81.8 | 88.8 | 72.8 | 100 | 67.8 | 30.9 | 30.2 | 99.6 | 99.4 | -0.6 | |
| DARTBackbone=Qwen2.5-Omni-7B, Retain Ratio=10%2026.04 | 31.6 | 65.5 | 0 | 71.3 | 98.3 | 65.3 | 41.6 | 20.5 | 73.4 | 73.8 | 26.2 |