Visual Reasoning on MMMU-Pro
39.42Avg@8Vanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaBackbone=InternVL3.5-8B, Token Retention Ratio=Full Tokens, DSTP=false2026.04 | 39.42 | — | 100 | |
| VanillaBackbone=Qwen3-VL-4B, Token Retention Ratio=Full Tokens, DSTP=false2026.04 | 37.63 | — | 100 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 34.27 | 86.9 | 92.44 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 31.21 | 79.2 | 90.09 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 31.04 | 78.7 | 87 | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 30.52 | 81.1 | 82.9 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 28.26 | 71.7 | 87.83 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 28.21 | 71.6 | 82.98 | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 28.03 | 74.5 | 83.8 | |
| PEPO_GBackbone=Qwen2.5-VL-3B-Instruct, Training Data=ViRL39K2026.03 | 27.47 | — | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 27.39 | 72.8 | 81.04 | |
| PAPO_GBackbone=Qwen2.5-VL-3B-Instruct, Training Data=ViRL39K2026.03 | 27.11 | — | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 26.87 | 71.4 | 71.48 | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 25.85 | 65.6 | 81.01 | |
| GRPOBackbone=Qwen2.5-VL-3B-Instruct, Training Data=ViRL39K2026.03 | 25.66 | — | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 22.25 | 59.1 | 68.74 | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 21.66 | 57.6 | 73.34 | |
| DivPrune (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 19.42 | 49.3 | 63.55 | |
| FastV (ECCV’24)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 19.41 | 51.6 | 55.68 | |
| FastV (ECCV’24)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 18.96 | 48.1 | 56.66 | |
| VisionZip (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 16.71 | 42.4 | 59.54 | |
| VisionZip (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 16.36 | 43.5 | 60.42 | |
| DivPrune (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 14.1 | 35.8 | 57.61 | |
| DivPrune (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 13.08 | 34.8 | 55.24 | |
| VisionZip (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 13 | 34.5 | 46.4 | |
| VisionZip (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 12.02 | 30.5 | 51.04 | |
| FastV (ECCV’24)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 11.6 | 29.4 | 49.42 | |
| FastV (ECCV’24)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 11.5 | 30.6 | 45.18 | |
| DivPrune (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 11.21 | 29.8 | 47.4 |