Multimodal Understanding on Aggregate Suite (GQA, SQA-I, VQA-T, MME, VQAv2, MMB)
59.7GQA ScoreVanilla
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| VanillaBackbone=Qwen2.5-VL-7B, Retained Tokens=1280, Reduction Ratio=100%2026.04 | 59.7 | 77.4 | 76.7 | 2,324 | 83.9 | 83.8 | 100 | |
| LearnPrunerBackbone=Qwen2.5-VL-7B, Retained Tokens=426, Reduction Ratio=66.7%2026.04 | 59.3 | 77.4 | 75.2 | 2,339 | 82.4 | 83.2 | 99.3 | |
| LearnPrunerBackbone=Qwen2.5-VL-7B, Retained Tokens=284, Reduction Ratio=77.8%2026.04 | 59 | 77 | 74.4 | 2,358 | 81.7 | 82.2 | 98.7 | |
| FastVBackbone=Qwen2.5-VL-7B, Retained Tokens=426, Reduction Ratio=66.7%2026.04 | 58.4 | 79.1 | 75.7 | 2,330 | 81.4 | 81.4 | 98.9 | |
| FastVBackbone=Qwen2.5-VL-7B, Retained Tokens=284, Reduction Ratio=77.8%2026.04 | 53.6 | 77.8 | 72.7 | 2,228 | 77.8 | 79.1 | 94.7 | |
| LearnPrunerBackbone=Qwen2.5-VL-7B, Retained Tokens=142, Reduction Ratio=88.9%2026.04 | 51.9 | 76.1 | 70.8 | 2,292 | 79.3 | 81.8 | 94.1 | |
| FastVBackbone=Qwen2.5-VL-7B, Retained Tokens=142, Reduction Ratio=88.9%2026.04 | 42.8 | 70.5 | 54.8 | 1,565 | 57.2 | 56.2 | 71.4 |