Visual Question Answering on GQA (Accuracy and Average performance retention)
65.34AccuracyLLaVA-NeXT 8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LLaVA-NeXT 8BBackbone=LLaVA-NeXT 8B, Sparsity Level=0%, Pruning Method=Dense, Pruning Type=None2026.03 | 65.34 | 100 | |
| VanillaBackbone=LLaVA-1.5-13B, Retained Tokens=5762026.06 | 63.3 | — | |
| ATV-PruningBackbone=LLaVA-NeXT 8B, Sparsity Level=50%, Pruning Method=ATV-Pruning, Pruning Type=unstructured, uniform pruning2026.03 | 63.28 | 94 | |
| TAMPBackbone=LLaVA-NeXT 8B, Sparsity Level=50%, Pruning Method=TAMP, Pruning Type=unstructured, uniform pruning2026.03 | 63.15 | 92.67 | |
| FP32Model=LLaVA-1.6-vicuna-13B, Quantization=FP322025.10 | 63.08 | — | |
| SparseGPTBackbone=LLaVA-NeXT 8B, Sparsity Level=50%, Pruning Method=SparseGPT, Pruning Type=unstructured, uniform pruning2026.03 | 62.86 | 91.74 | |
| WandaBackbone=LLaVA-NeXT 8B, Sparsity Level=50%, Pruning Method=Wanda, Pruning Type=unstructured, uniform pruning2026.03 | 62.61 | 88.36 | |
| Qwen2-VL 7BBackbone=Qwen2-VL 7B, Sparsity Level=0%, Pruning Method=Dense, Pruning Type=None2026.03 | 62.35 | 100 | |
| PriorTRBackbone=LLaVA-1.5-13B, Retained Tokens=1922026.06 | 62.3 | — | |
| VanillaSource=CVPR’24, Average Token=576, Reroute=false2026.06 | 61.9 | 100 | |
| PriorTRBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 61.3 | — | |
| Nüwa+RerouteSource=Ours, Average Token=192, Reroute=true2026.06 | 61.1 | 100.6 | |
| NüwaSource=ICLR’26, Average Token=192, Reroute=false2026.06 | 60.9 | 99.2 | |
| Nüwa+RerouteSource=Ours, Average Token=128, Reroute=true2026.06 | 60.4 | 99.8 | |
| NüwaSource=ICLR’26, Average Token=128, Reroute=false2026.06 | 60.2 | 98.4 | |
| PDropBackbone=LLaVA-1.5-13B, Retained Tokens=1922026.06 | 59.6 | — | |
| VisionZipSource=CVPR’25, Average Token=192, Reroute=false2026.06 | 59.3 | 97.9 | |
| FastVBackbone=LLaVA-1.5-13B, Retained Tokens=1922026.06 | 59.1 | — | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1922026.06 | 59.1 | — | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=1922026.06 | 58.7 | — | |
| PriorTRBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 58.7 | — | |
| SparseGPTBackbone=LLaVA-NeXT 8B, Sparsity Level=60%, Pruning Method=SparseGPT, Pruning Type=unstructured, uniform pruning2026.03 | 58.52 | 76.24 | |
| Nüwa+RerouteSource=Ours, Average Token=64, Reroute=true2026.06 | 58.5 | 96.7 | |
| NüwaSource=ICLR’26, Average Token=64, Reroute=false2026.06 | 58.3 | 95.8 | |
| ATV-PruningBackbone=Qwen2-VL 7B, Sparsity Level=60%, Pruning Method=ATV-Pruning, Pruning Type=unstructured, uniform pruning2026.03 | 58.13 | 85.65 | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 57.9 | — | |
| PruMergeSource=ICCV’25, Average Token=128, Reroute=false2026.06 | 57.8 | 94.7 | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 57.8 | — | |
| FastVBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 57.7 | — | |
| ATV-PruningBackbone=LLaVA-NeXT 8B, Sparsity Level=60%, Pruning Method=ATV-Pruning, Pruning Type=unstructured, uniform pruning2026.03 | 57.61 | 77.01 | |
| SparseVLMSource=ICML’25, Average Token=192, Reroute=false2026.06 | 57.6 | 95.4 | |
| VisionZipSource=CVPR’25, Average Token=128, Reroute=false2026.06 | 57.6 | 96.9 | |
| TAMPBackbone=Qwen2-VL 7B, Sparsity Level=60%, Pruning Method=TAMP, Pruning Type=unstructured, uniform pruning2026.03 | 57.47 | 83.79 | |
| PDropBackbone=LLaVA-1.5-13B, Retained Tokens=1282026.06 | 57.2 | — | |
| PDropSource=CVPR’25, Average Token=192, Reroute=false2026.06 | 57.1 | 95.7 | |
| SparseGPTBackbone=Qwen2-VL 7B, Sparsity Level=60%, Pruning Method=SparseGPT, Pruning Type=unstructured, uniform pruning2026.03 | 56.98 | 85.3 | |
| FP32Model=LLaVA-1.6-mistral-7B, Quantization=FP322025.10 | 56.59 | — | |
| PDropSource=CVPR’25, Average Token=128, Reroute=false2026.06 | 56 | 93.8 | |
| SparseVLMSource=ICML’25, Average Token=128, Reroute=false2026.06 | 56 | 92.9 | |
| VisPrunerBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 55.9 | — | |
| PruMergeSource=ICCV’25, Average Token=64, Reroute=false2026.06 | 55.4 | 92.1 | |
| VisionZipSource=CVPR’25, Average Token=64, Reroute=false2026.06 | 55.1 | 93.2 | |
| TAMPBackbone=LLaVA-NeXT 8B, Sparsity Level=60%, Pruning Method=TAMP, Pruning Type=unstructured, uniform pruning2026.03 | 54.62 | 73.75 | |
| WandaBackbone=Qwen2-VL 7B, Sparsity Level=60%, Pruning Method=Wanda, Pruning Type=unstructured, uniform pruning2026.03 | 54.37 | 77.78 | |
| SparseVLMSource=ICML’25, Average Token=64, Reroute=false2026.06 | 53.8 | 91.5 | |
| FastVBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 53.7 | — | |
| FastVSource=ECCV’24, Average Token=192, Reroute=false2026.06 | 52.7 | 90.1 | |
| RegCacheModel=LLaVA-1.6-vicuna-13B, Quantization=4-bit2025.10 | 51.4 | — | |
| PDropBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 50.7 | — | |
| SparseVLMBackbone=LLaVA-1.5-13B, Retained Tokens=642026.06 | 50.6 | — | |
| WandaBackbone=LLaVA-NeXT 8B, Sparsity Level=60%, Pruning Method=Wanda, Pruning Type=unstructured, uniform pruning2026.03 | 49.88 | 64.45 | |
| FastVSource=ECCV’24, Average Token=128, Reroute=false2026.06 | 49.6 | 85.2 | |
| RepQ-ViTModel=LLaVA-1.6-vicuna-13B, Quantization=4-bit2025.10 | 48.93 | — | |
| RegCacheModel=LLaVA-1.6-mistral-7B, Quantization=4-bit2025.10 | 46.21 | — | |
| FastVSource=ECCV’24, Average Token=64, Reroute=false2026.06 | 46.1 | 77.9 | |
| RepQ-ViTModel=LLaVA-1.6-mistral-7B, Quantization=4-bit2025.10 | 44.63 | — | |
| PDropSource=CVPR’25, Average Token=64, Reroute=false2026.06 | 41.9 | 70.3 | |
| NaïveModel=LLaVA-1.6-vicuna-13B, Quantization=4-bit2025.10 | 35.72 | — | |
| NaïveModel=LLaVA-1.6-mistral-7B, Quantization=4-bit2025.10 | 31.87 | — |