Visual Reasoning on MathVerse
61.29AccuracyVanilla
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| VanillaBackbone=Qwen3-VL-4B, Token Retention Ratio=Full Tokens, DSTP=false2026.04 | 61.29 | — | 100 | — | |
| Image-ExpertTraining Strategy=Image-Specific Expert2026.04 | 59.84 | — | — | — | |
| Image-ExpertTraining Strategy=Image-Expert2026.04 | 59.84 | — | — | — | |
| Mixed RLVRTraining Strategy=Mixed RLVR2026.04 | 59.82 | — | — | — | |
| OPD T→VTraining Strategy=Static On-policy Policy Distillation (Text to Image)2026.04 | 59.71 | — | — | — | |
| OPD V→TTraining Strategy=Static On-policy Policy Distillation (Image to Text)2026.04 | 59.5 | — | — | — | |
| CoPDTraining Strategy=CoPD2026.04 | 59.34 | — | — | — | |
| MOPDTraining Strategy=MOPD2026.04 | 58.93 | — | — | — | |
| CoPDTraining Strategy=Co-Evolving Policy Distillation2026.04 | 58.88 | — | — | — | |
| Text-ExpertTraining Strategy=Text-Specific Expert2026.04 | 58.41 | — | — | — | |
| Text-ExpertTraining Strategy=Text-Expert2026.04 | 58.41 | — | — | — | |
| Video-ExpertTraining Strategy=Video-Expert2026.04 | 58.16 | — | — | — | |
| BaseTraining Strategy=Base Model2026.04 | 57.82 | — | — | — | |
| BaseTraining Strategy=Base2026.04 | 57.82 | — | — | — | |
| Mixed RLVRTraining Strategy=Joint optimization on combined budget2026.04 | 57.81 | — | — | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 52.54 | 85.7 | 82.9 | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 50.76 | 82.8 | 81.04 | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 50.25 | 82 | 83.8 | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 46.19 | 75.4 | 73.34 | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 42.76 | 69.8 | 71.48 | — | |
| VanillaBackbone=InternVL3.5-8B, Token Retention Ratio=Full Tokens, DSTP=false2026.04 | 42 | — | 100 | — | |
| DSTPBackbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 40.46 | 66 | 68.74 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 40.22 | 95.8 | 92.44 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 39.59 | 94.3 | 90.09 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=FastV (ECCV’24), DSTP=true2026.04 | 38.94 | 92.7 | 87 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=33.3%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 37.56 | 89.4 | 87.83 | — | |
| VisionZip (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 36.8 | 60 | 60.42 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=DivPrune (CVPR’25), DSTP=true2026.04 | 36.54 | 87 | 82.98 | — | |
| DSTPBackbone=InternVL3.5-8B, Token Retention Ratio=22.2%, Pruning Method=VisionZip (CVPR’25), DSTP=true2026.04 | 34.13 | 81.3 | 81.01 | — | |
| DivPrune (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 33.9 | 55.3 | 55.24 | — | |
| FastV (ECCV’24)Backbone=Qwen3-VL-4B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 32.23 | 52.6 | 55.68 | — | |
| DivPrune (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 26.59 | 63.3 | 63.55 | — | |
| DivPrune (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 26.14 | 42.6 | 47.4 | — | |
| VisionZip (CVPR’25)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 25.98 | 42.4 | 46.4 | — | |
| FastV (ECCV’24)Backbone=Qwen3-VL-4B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 25.88 | 42.2 | 45.18 | — | |
| VisionZip (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 24.84 | 59.1 | 59.54 | — | |
| DivPrune (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 23.64 | 56.3 | 57.61 | — | |
| FastV (ECCV’24)Backbone=InternVL3.5-8B, Token Retention Ratio=33.3%, DSTP=false2026.04 | 21.57 | 51.4 | 56.66 | — | |
| VisionZip (CVPR’25)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 20.55 | 48.9 | 51.04 | — | |
| FastV (ECCV’24)Backbone=InternVL3.5-8B, Token Retention Ratio=22.2%, DSTP=false2026.04 | 20.06 | 47.8 | 49.42 | — | |
| DRIFTModel Category=Reasoning Fine-tuning Methods2025.10 | — | — | — | 43.9 | |
| InternLM-XComposer2.5Model Category=Open-source Models2025.10 | — | — | — | 16.2 | |
| InternVL2-8BModel Category=Open-source Models2025.10 | — | — | — | 20.4 | |
| InternVL2.5-8BModel Category=Open-source Models2025.10 | — | — | — | 22.8 | |
| InternVL3-8BModel Category=Open-source Models2025.10 | — | — | — | 33.9 | |
| Kimi-VL-16BModel Category=Open-source Models2025.10 | — | — | — | 34.1 | |
| LLaVA-OneVision-7BModel Category=Open-source Models2025.10 | — | — | — | 17.6 | |
| OpenVLThinker-7BModel Category=Reasoning Fine-tuning Methods2025.10 | — | — | — | 38.1 | |
| QvQ-72B-PreviewModel Category=Open-source Models2025.10 | — | — | — | 48.2 | |
| Qwen2-VL-7BModel Category=Open-source Models2025.10 | — | — | — | 25.4 | |
| Qwen2.5-VL-7BModel Category=Open-source Models, Evaluation Source=Reproduced by authors2025.10 | — | — | — | 41.4 | |
| R1-Onevision-7BModel Category=Reasoning Fine-tuning Methods2025.10 | — | — | — | 40 | |
| R1-VL-7BModel Category=Reasoning Fine-tuning Methods2025.10 | — | — | — | 40 | |
| SFT BaselineModel Category=Reasoning Fine-tuning Methods2025.10 | — | — | — | 42 |