Multimodal Evaluation on MMBench CN
82.8AccuracyQwen2.5-VL-7B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5-VL-7BRetain Tokens=1296, Retain Ratio (%)=100%2026.07 | 82.8 | 84 | |
| Qwen2.5-VL-7BRatio=100%, Backbone=Qwen2.5-VL-7B2026.05 | 82.37 | — | |
| F3ARatio=40%, Backbone=Qwen2.5-VL-7B2026.05 | 81.94 | — | |
| F3ARatio=60%, Backbone=Qwen2.5-VL-7B2026.05 | 81.82 | — | |
| VisionZipRatio=60%, Backbone=Qwen2.5-VL-7B2026.05 | 81.71 | — | |
| CDPrunerRatio=60%, Backbone=Qwen2.5-VL-7B2026.05 | 81.55 | — | |
| DivPruneRatio=60%, Backbone=Qwen2.5-VL-7B2026.05 | 81.38 | — | |
| EADPRetain Tokens=512, Retain Ratio (%)=↓ 60.5%2026.07 | 80.8 | 78 | |
| FastVRatio=60%, Backbone=Qwen2.5-VL-7B2026.05 | 80.72 | — | |
| VisionZipRatio=40%, Backbone=Qwen2.5-VL-7B2026.05 | 80.72 | — | |
| DivPruneRatio=40%, Backbone=Qwen2.5-VL-7B2026.05 | 80.56 | — | |
| DivPruneRetain Tokens=512, Retain Ratio (%)=↓ 60.5%2026.07 | 80.4 | 78.1 | |
| CDPrunerRatio=40%, Backbone=Qwen2.5-VL-7B2026.05 | 80.31 | — | |
| F3ARatio=20%, Backbone=Qwen2.5-VL-7B2026.05 | 80.24 | — | |
| HiPruneRetain Tokens=512, Retain Ratio (%)=↓ 60.5%2026.07 | 80.1 | 75.9 | |
| FastVRatio=40%, Backbone=Qwen2.5-VL-7B2026.05 | 80.06 | — | |
| HiPruneRetain Tokens=256, Retain Ratio (%)=↓ 80.2%2026.07 | 79 | 69.4 | |
| EADPRetain Tokens=256, Retain Ratio (%)=↓ 80.2%2026.07 | 78.5 | 74.3 | |
| CDPrunerRetain Tokens=512, Retain Ratio (%)=↓ 60.5%2026.07 | 78.3 | 71.6 | |
| DivPruneRetain Tokens=256, Retain Ratio (%)=↓ 80.2%2026.07 | 78.3 | 73.6 | |
| CDPrunerRatio=20%, Backbone=Qwen2.5-VL-7B2026.05 | 77.84 | — | |
| VisionZipRatio=20%, Backbone=Qwen2.5-VL-7B2026.05 | 77.59 | — | |
| DivPruneRatio=20%, Backbone=Qwen2.5-VL-7B2026.05 | 77.43 | — | |
| DivPruneRetain Tokens=128, Retain Ratio (%)=↓ 90.1%2026.07 | 77.3 | 66.4 | |
| FastVRatio=20%, Backbone=Qwen2.5-VL-7B2026.05 | 76.6 | — | |
| EADPRetain Tokens=128, Retain Ratio (%)=↓ 90.1%2026.07 | 76.4 | 68.4 | |
| HiPruneRetain Tokens=128, Retain Ratio (%)=↓ 90.1%2026.07 | 76.2 | 62.3 | |
| CDPrunerRetain Tokens=256, Retain Ratio (%)=↓ 80.2%2026.07 | 74.7 | 65 | |
| mPlug-Owl3LLM=Qwen-8B2024.11 | 74.3 | — | |
| SliMELLM=Llama3-8B2024.11 | 71.8 | — | |
| CDPrunerRetain Tokens=128, Retain Ratio (%)=↓ 90.1%2026.07 | 71.7 | 58.2 | |
| FocusLLaVALLM=Llama3-8B2024.11 | 70.3 | — | |
| LLaVA-NeXTLLM=Llama3-8B, implementation=ours2024.11 | 70.1 | — | |
| LLaVA-1.5LLM=Llama3-8B2024.11 | 67.7 | — | |
| APTBase Model=LLaVA-1.5-13B, Img/s=2.72, Speedup=+23%2025.10 | 63.7 | — | |
| LLaVA-1.5-13BImg/s=2.22, Speedup=-2025.10 | 63.4 | — | |
| FSRBackbone=LLaVA-NeXT-13B, Tokens Retained=960, Pruning Ratio=66.7%2026.02 | 63.3 | — | |
| FSRBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 63.3 | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=960, Pruning Ratio=66.7%2026.02 | 63.1 | — | |
| FSRBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 63.1 | — | |
| DivPruneBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 62.9 | — | |
| VisionZipBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 62.6 | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 62.6 | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=960, Pruning Ratio=66.7%2026.02 | 62.5 | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 62.5 | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 62.5 | — | |
| VisPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 62.5 | — | |
| ResizingBase Model=LLaVA-1.5-13B, Img/s=2.72, Speedup=+23%2025.10 | 62.5 | — | |
| VisionZipBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 62.3 | — | |
| DivPruneBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 61.9 | — | |
| CDPrunerBackbone=LLaVA-NeXT-13B, Tokens Retained=320, Pruning Ratio=88.9%2026.02 | 61.9 | — | |
| LLaVA-NeXT-13BBackbone=LLaVA-NeXT-13B, Tokens Retained=2880, Pruning Ratio=0%2026.02 | 61.2 | — | |
| RandomBase Model=LLaVA-1.5-13B, Img/s=2.79, Speedup=+26%2025.10 | 60.8 | — | |
| LLaVA-NeXTLLM=Vicuna-7B2024.11 | 60.6 | — | |
| VanillaBackbone=LLaVA-NeXT-7B, Token Retention=2880, Reduction Ratio=100%2026.02 | 60.6 | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 60.6 | — | |
| LLaVoltaStages=Three, Scheme=compr. deeper, #Tokens=10597, Compression Ratio (CR)=174%, TFLOPS=8.26, Latency (ms)=68.5, Train Time=12.8h2024.06 | 59.9 | — | |
| LLaVoltaStages=Four, Scheme=deeper then wider, #Tokens=10863, Compression Ratio (CR)=170%, TFLOPS=8.26, Latency (ms)=68.5, Train Time=12.8h2024.06 | 59.9 | — | |
| FastVBackbone=LLaVA-NeXT-13B, Tokens Retained=640, Pruning Ratio=77.8%2026.02 | 59.9 | — | |
| LLaVoltaStages=Single, Scheme=no compression, #Tokens=18432, Compression Ratio (CR)=-, TFLOPS=8.26, Latency (ms)=68.5, Train Time=15.3h2024.06 | 59.4 | — | |
| LLaVoltaStages=Four, Scheme=wider then deeper, #Tokens=11088, Compression Ratio (CR)=166%, TFLOPS=8.26, Latency (ms)=68.5, Train Time=12.9h2024.06 | 59.3 | — | |
| ApETBackbone=LLaVA-NeXT-7B, Token Retention=640, Reduction Ratio=77.8%2026.02 | 59.3 | — | |
| LLaVoltaStages=Three, Scheme=compr. wider, #Tokens=10407, Compression Ratio (CR)=177%, TFLOPS=8.26, Latency (ms)=68.5, Train Time=12.8h2024.06 | 59.1 | — | |
| LLaVoltaStages=Three, Scheme=last stage compression, #Tokens=7848, Compression Ratio (CR)=235%, TFLOPS=5.47, Latency (ms)=52.2, Train Time=12.4h2024.06 | 59.1 | — | |
| LLaVoltaStages=Two, Scheme=compression, #Tokens=10062, Compression Ratio (CR)=183%, TFLOPS=8.26, Latency (ms)=68.5, Train Time=12.8h2024.06 | 59 | — | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Token Retention=640, Reduction Ratio=77.8%2026.02 | 58.5 | — | |
| PDropBackbone=LLaVA-NeXT-7B, Token Retention=640, Reduction Ratio=77.8%2026.02 | 58.5 | — | |
| LLaVA-1.5LLM=Vicuna-7B2024.11 | 58.3 | — | |
| Upper Bound2025.12 | 58.3 | — | |
| LLaVA-1.5-7BRetain Tokens=5762026.04 | 58.3 | — | |
| APTBase Model=LLaVA-1.5-7B, Img/s=4.51, Speedup=+22%2025.10 | 58.2 | — | |
| LLaVA-1.5-7BRetained Tokens=5762026.03 | 58.1 | — | |
| LLaVA-1.5-7BImg/s=3.70, Speedup=-2025.10 | 58.1 | — | |
| ResizingBase Model=LLaVA-1.5-7B, Img/s=4.51, Speedup=+22%2025.10 | 58.1 | — | |
| LLaVA-1.5-7BRetain Tokens=576, Pruning Ratio=0%2026.07 | 58.1 | — | |
| FlashVLMToken Budget=128 Tokens, Pruning Rate=77.8%2025.12 | 57.6 | — | |
| RandomBase Model=LLaVA-1.5-7B, Img/s=4.58, Speedup=+24%2025.10 | 57.6 | — | |
| LLaVA-1.5Inference Mode=Normal, Pruning Rate (p)=0%, Mode Isolator Employment=No2026.02 | 57.39 | — | |
| VisPrunerRetained Tokens=1282026.03 | 57.3 | — | |
| DualSpeedInference Mode=Pruning, Pruning Rate (p)=90%, Mode Isolator Employment=No2026.02 | 57.13 | — | |
| VisionZipRetained Tokens=1282026.03 | 57 | — | |
| NaivePruneInference Mode=Pruning, Pruning Rate (p)=90%, Mode Isolator Employment=No2026.02 | 56.96 | — | |
| SparseVLMRetained Tokens=1282026.03 | 56.9 | — | |
| DualSpeedInference Mode=Normal, Pruning Rate (p)=0%, Mode Isolator Employment=No2026.02 | 56.87 | — | |
| DualSpeedInference Mode=Pruning, Pruning Rate (p)=90%, Mode Isolator Employment=Yes2026.02 | 56.87 | — | |
| ApETBackbone=LLaVA-NeXT-7B, Token Retention=320, Reduction Ratio=88.9%2026.02 | 56.6 | — | |
| HoloVBackbone=LLaVA-NeXT-13B, Tokens Retained=960, Pruning Ratio=66.7%2026.02 | 56.4 | — | |
| HiPruneRetain Tokens=128, Pruning Ratio=77.8%2026.07 | 56.4 | — | |
| PDropBackbone=LLaVA-NeXT-7B, Token Retention=320, Reduction Ratio=88.9%2026.02 | 56.2 | — | |
| SparseVLMRetain Tokens=128, Pruning Ratio=77.8%2026.07 | 56.1 | — | |
| DARTRetain Tokens=128, Pruning Ratio=77.8%2026.07 | 56 | — | |
| SparseVLMBackbone=LLaVA-NeXT-7B, Token Retention=320, Reduction Ratio=88.9%2026.02 | 55.9 | — | |
| FastVRetained Tokens=1282026.03 | 55.9 | — | |
| VisionZipRetain Tokens=128, Pruning Ratio=77.8%2026.07 | 55.9 | — | |
| FlashVLMToken Budget=64 Tokens, Pruning Rate=88.9%2025.12 | 55.8 | — | |
| PruMerge+Retained Tokens=1282026.03 | 55.8 | — | |
| AgilePrunerRetained Tokens=642026.03 | 55.8 | — | |
| ID-SelectionImportance Criterion=unified score, Retain Tokens=642026.04 | 55.8 | — | |
| AgilePrunerRetained Tokens=1282026.03 | 55.5 | — | |
| VisionZipRetained Tokens=642026.03 | 55.4 | — |