Multimodal Understanding on GQA, MMB, MME, POPE, SQA, VQAv2, MMMU, SEEDI, and VizWiz (test val)
62.3GQA Accuracybaseline
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| baselineBackbone=Qwen2-VL 7B, Token retention ratio=100%2026.03 | 62.3 | 79.1 | 2,318 | 87.9 | 84.7 | 81.2 | 51.1 | 76.5 | 68.2 | 100 | |
| PRUNESIDBackbone=Qwen2-VL 7B, Token retention ratio=33.3%2026.03 | 60.5 | 72.1 | 2,146 | 87 | 81.8 | 79.8 | 48.3 | 75.5 | 66.1 | 96.1 | |
| PRUNESIDBackbone=Qwen2-VL 7B, Token retention ratio=22.2%2026.03 | 59.2 | 69.2 | 2,119 | 86 | 79.5 | 78.1 | 47.3 | 74.8 | 64.8 | 94.1 | |
| PACTBackbone=Qwen2-VL 7B, Token retention ratio=33.3%2026.03 | 59.1 | 75.8 | 2,068 | 85.6 | 80.5 | 78.2 | 48.2 | 74 | 67.9 | 95.3 | |
| PRUNESIDBackbone=Qwen2-VL 7B, Token retention ratio=11.1%2026.03 | 55.9 | 65 | 2,039 | 82.9 | 76.1 | 73.6 | 46.9 | 71.8 | 64.1 | 90.4 | |
| PACTBackbone=Qwen2-VL 7B, Token retention ratio=22.2%2026.03 | 55.8 | 72.2 | 2,012 | 82.4 | 78.3 | 77.6 | 48.8 | 71.7 | 67.4 | 92.8 | |
| PACTBackbone=Qwen2-VL 7B, Token retention ratio=11.1%2026.03 | 50.1 | 63.1 | 1,785 | 71.4 | 75 | 74.3 | 48.5 | 66 | 63.1 | 85.8 |