Multimodal Understanding on Multimodal Evaluation Suite (MME, TextVQA, DocVQA, AI2D, MMMU, MMBench)
1,687.9MME ScoreQwen2.5-VL-7B
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7BVisual-token budget=1296 (100%), Base Model=Qwen2.5-VL-7B2026.07 | 1,687.9 | 77.3 | 94.3 | 83.2 | 50.6 | 83.1 | 80.5 | 100 | |
| DivPruneVisual-token budget=256 (↓ 80.2%), Base Model=Qwen2.5-VL-7B, Venue=CVPR’252026.07 | 1,671.9 | 70 | 67.9 | 77.8 | 48.6 | 80 | 77.2 | 91.9 | |
| AnchorPruneVisual-token budget=128 (↓ 90.1%), Base Model=Qwen2.5-VL-7B2026.07 | 1,663.9 | 68.4 | 51 | 77.7 | 46.8 | 79.6 | 75.7 | 88.1 | |
| AnchorPruneVisual-token budget=256 (↓ 80.2%), Base Model=Qwen2.5-VL-7B2026.07 | 1,657.5 | 72.6 | 72.9 | 78.8 | 48.6 | 80.8 | 78.2 | 93.5 | |
| DivPruneVisual-token budget=128 (↓ 90.1%), Base Model=Qwen2.5-VL-7B, Venue=CVPR’252026.07 | 1,642.4 | 66.3 | 49.8 | 75.2 | 47.3 | 76.9 | 75.4 | 86.6 | |
| FastVVisual-token budget=256 (↓ 80.2%), Base Model=Qwen2.5-VL-7B, Venue=ECCV’242026.07 | 1,638.9 | 74.9 | 60 | 78.3 | 49.3 | 79.8 | 77.3 | 91.6 | |
| CDPrunerVisual-token budget=256 (↓ 80.2%), Base Model=Qwen2.5-VL-7B, Venue=NeurIPS’252026.07 | 1,621.9 | 63.2 | 52.7 | 78 | 47.7 | 78 | 77 | 87.3 | |
| CDPrunerVisual-token budget=128 (↓ 90.1%), Base Model=Qwen2.5-VL-7B, Venue=NeurIPS’252026.07 | 1,572.1 | 57.7 | 36.3 | 76.9 | 45.9 | 77.9 | 74.7 | 82.3 | |
| AnchorPruneVisual-token budget=64 (↓ 95.1%), Base Model=Qwen2.5-VL-7B2026.07 | 1,563.2 | 61.5 | 31.3 | 74.2 | 45.2 | 75.7 | 73.2 | 80.8 | |
| FastVVisual-token budget=128 (↓ 90.1%), Base Model=Qwen2.5-VL-7B, Venue=ECCV’242026.07 | 1,554.8 | 70.7 | 40 | 72.9 | 47.1 | 76.3 | 72.1 | 84 | |
| DivPruneVisual-token budget=64 (↓ 95.1%), Base Model=Qwen2.5-VL-7B, Venue=CVPR’252026.07 | 1,544.1 | 60.7 | 31.6 | 70.7 | 46.9 | 74.6 | 71.7 | 80 | |
| CDPrunerVisual-token budget=64 (↓ 95.1%), Base Model=Qwen2.5-VL-7B, Venue=NeurIPS’252026.07 | 1,431.7 | 51.6 | 22.8 | 73.6 | 44.7 | 73.9 | 72.9 | 76 | |
| FastVVisual-token budget=64 (↓ 95.1%), Base Model=Qwen2.5-VL-7B, Venue=ECCV’242026.07 | 1,381.4 | 63.7 | 27.7 | 68.4 | 45.4 | 67.5 | 64.7 | 75.3 |