Utility Evaluation on MMbench and DocVQA (test)
87.02MMbench ScoreFull Model
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Full ModelBackbone=Qwen2.5-VL, Sparsity=0%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 87.02 | 94.51 | 90.76 | |
| WandaBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 85.15 | 91.97 | 88.56 | |
| Wanda w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.020‱2025.05 | 85.01 | 92.13 | 88.57 | |
| SNIP w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.150‱2025.05 | 84.62 | 92.9 | 88.76 | |
| SNIPBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 84.55 | 92.93 | 88.74 | |
| SparseGPTBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy2025.05 | 83.88 | 90.64 | 87.26 | |
| SparseGPT w/ HSRBackbone=Qwen2.5-VL, Sparsity=50%, Evaluation Setting=zero-shot, Decoding Strategy=greedy, Restoration ratio=0.133‱2025.05 | 83.88 | 90.63 | 87.25 |