Image Understanding on GQA, MME, POPE, VQA^T, MMB, SQA
62.2GQA ScoreVanilla
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| VanillaBackbone=Qwen2-VL-7B, Pruning Status=w/o Pruning, Token Reduction Rate=0.0%2025.05 | 62.2 | 2,317 | 86.1 | 82.1 | 80.5 | 84.7 | 100 | |
| MoB (with η)Backbone=Qwen2-VL-7B, Token Reduction Rate=66.7%2025.05 | 61.8 | 2,268 | 84.7 | 81.1 | 79.5 | 82.3 | 98.4 | |
| DARTBackbone=Qwen2-VL-7B, Token Reduction Rate=66.7%2025.05 | 60.2 | 2,245 | 83.9 | 80.5 | 78.9 | 81.4 | 97 | |
| MoB (with η)Backbone=Qwen2-VL-7B, Token Reduction Rate=77.8%2025.05 | 59.4 | 2,203 | 82.8 | 75.8 | 78.1 | 80.4 | 95.2 | |
| DARTBackbone=Qwen2-VL-7B, Token Reduction Rate=77.8%2025.05 | 58.5 | 2,175 | 82.1 | 75.3 | 77.3 | 79.6 | 94.3 | |
| FastVBackbone=Qwen2-VL-7B, Token Reduction Rate=66.7%2025.05 | 58 | 2,130 | 82.1 | 77.3 | 76.1 | 80 | 94 | |
| FastVBackbone=Qwen2-VL-7B, Token Reduction Rate=77.8%2025.05 | 56.7 | 2,031 | 79.2 | 72 | 74.1 | 78.3 | 91 | |
| MoB (with η)Backbone=Qwen2-VL-7B, Token Reduction Rate=88.9%2025.05 | 56.5 | 2,094 | 78.5 | 62.7 | 72.8 | 78.4 | 88.6 | |
| DARTBackbone=Qwen2-VL-7B, Token Reduction Rate=88.9%2025.05 | 55.5 | 2,052 | 77.9 | 61.8 | 72 | 77.6 | 87.4 | |
| FastVBackbone=Qwen2-VL-7B, Token Reduction Rate=88.9%2025.05 | 51.9 | 1,962 | 76.1 | 60.3 | 70.1 | 75.8 | 84.4 |