Classification on Zero-shot Evaluation Suite (BoolQ, PIQA, HellaSwag, WinoGrande, ARC-e, ARC-c, OBQA)
69.99Average Accuracy (Zero-Shot Suite)Dense
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| DenseSparsity=0%, Backbone=LLaMA3-8B2026.03 | 69.99 | 81.38 | 80.79 | 79.16 | 72.61 | 77.74 | 53.24 | 45 | |
| ROSESparsity=30%, Backbone=LLaMA3-8B2026.03 | 69.57 | 82.23 | 80.58 | 78.56 | 74.51 | 76.43 | 50.51 | 44.2 | |
| SparseGPTSparsity=30%, Backbone=LLaMA3-8B2026.03 | 69.43 | 82.14 | 80.09 | 78.54 | 74.03 | 76.6 | 50.43 | 44.2 | |
| DenseRatio=0%, Backbone=LLaMA2-13B2026.02 | 69.31 | — | — | — | — | — | — | — | |
| LLaMA-7BRatio=0%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 68.59 | 76.5 | 79.8 | 76.1 | 70.1 | 72.8 | 47.6 | 57.2 | |
| ROSESparsity=40%, Backbone=LLaMA3-8B2026.03 | 68.53 | 82.14 | 79 | 76.62 | 72.61 | 74.37 | 50.34 | 44.6 | |
| SparseGPTSparsity=40%, Backbone=LLaMA3-8B2026.03 | 68.51 | 82.42 | 79.43 | 76.71 | 73.24 | 74.87 | 50.09 | 42.8 | |
| DenseRatio=0%, Backbone=Vicuna-7B2026.02 | 66.85 | — | — | — | — | — | — | — | |
| DenseRatio=0%, Backbone=LLaMA2-7B2026.02 | 66.79 | — | — | — | — | — | — | — | |
| ROSESparsity=50%, Backbone=LLaMA3-8B2026.03 | 65.58 | 78.87 | 77.2 | 73.19 | 72.45 | 70.92 | 45.65 | 40.8 | |
| SparseGPTSparsity=50%, Backbone=LLaMA3-8B2026.03 | 65.18 | 78.41 | 77.64 | 72.92 | 72.85 | 68.73 | 44.11 | 41.6 | |
| GradMAPRatio=12.50%, Backbone=LLaMA2-13B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=261.652026.02 | 64.39 | — | — | — | — | — | — | — | |
| GradMAPRatio=12.50%, Backbone=LLaMA2-13B, Pruning Stage=Stage 1 (layer selection), Pruning Time=185.062026.02 | 64.34 | — | — | — | — | — | — | — | |
| ShortGPTRatio=12.50%, Backbone=LLaMA2-13B, Pruning Time=1682.422026.02 | 64.14 | — | — | — | — | — | — | — | |
| GradMAPRatio=12.50%, Backbone=Vicuna-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=122.792026.02 | 63.52 | — | — | — | — | — | — | — | |
| LLaMA-7BPruning Ratio=0%, Calibration Dataset=C42023.05 | 63.25 | 73.18 | 78.35 | 72.99 | 67.01 | 67.45 | 41.38 | 42.4 | |
| LLaMA-7B*Ratio=0%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 63.25 | 73.18 | 78.35 | 72.99 | 67.01 | 67.45 | 41.38 | 42.4 | |
| ShortGPTRatio=12.50%, Backbone=Vicuna-7B, Pruning Time=523.342026.02 | 63.21 | — | — | — | — | — | — | — | |
| GradMAPRatio=12.50%, Backbone=Vicuna-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=71.772026.02 | 63.21 | — | — | — | — | — | — | — | |
| SLEBRatio=12.50%, Backbone=Vicuna-7B, Pruning Time=506.112026.02 | 62.79 | — | — | — | — | — | — | — | |
| GradMAPRatio=12.50%, Backbone=LLaMA2-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=71.402026.02 | 62.54 | — | — | — | — | — | — | — | |
| GradMAPRatio=12.50%, Backbone=LLaMA2-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=122.202026.02 | 62.54 | — | — | — | — | — | — | — | |
| HyWIARatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 62.44 | 68.53 | 77.8 | 70.58 | 67.49 | 70.24 | 40.44 | 42 | |
| ShortGPTRatio=12.50%, Backbone=LLaMA2-7B, Pruning Time=519.332026.02 | 62.32 | — | — | — | — | — | — | — | |
| FLAPRatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 62.08 | 69.63 | 76.82 | 71.2 | 68.35 | 69.91 | 39.25 | 39.4 | |
| SLEBRatio=12.50%, Backbone=LLaMA2-7B, Pruning Time=503.982026.02 | 60.87 | — | — | — | — | — | — | — | |
| SLEBRatio=12.50%, Backbone=LLaMA2-13B, Pruning Time=1638.512026.02 | 60.6 | — | — | — | — | — | — | — | |
| LoRAPrunePruning Ratio=20%2023.05 | 60.05 | 65.62 | 79.31 | 70 | 62.76 | 65.87 | 37.69 | 39.14 | |
| LoRAPruneRatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 60.05 | 65.62 | 79.31 | 70 | 62.76 | 65.87 | 37.69 | 39.14 | |
| MKARatio=12.50%, Backbone=LLaMA2-13B, Pruning Time=381.042026.02 | 59.82 | — | — | — | — | — | — | — | |
| GradMAPRatio=25.00%, Backbone=LLaMA2-13B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=368.302026.02 | 59.62 | — | — | — | — | — | — | — | |
| CompressoRatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 59.51 | 79.08 | 75.46 | 53.44 | 67.8 | 68.64 | 37.97 | 34.2 | |
| LoRAPrune-8bitPruning Ratio=20%, Quantization=8-bit2023.05 | 59.46 | 65.37 | 76.65 | 69.41 | 63.78 | 65.45 | 36.12 | 39.5 | |
| GradMAPRatio=25.00%, Backbone=LLaMA2-13B, Pruning Stage=Stage 1 (layer selection), Pruning Time=299.302026.02 | 59.33 | — | — | — | — | — | — | — | |
| LLM-PrunerPruning Ratio=20%2023.05 | 59.23 | 64.62 | 77.2 | 68.8 | 63.14 | 64.31 | 36.77 | 39.8 | |
| ShortGPTRatio=25.00%, Backbone=LLaMA2-13B, Pruning Time=2552.412026.02 | 58.97 | — | — | — | — | — | — | — | |
| SparseGPTSparsity=60%, Backbone=LLaMA3-8B2026.03 | 58.61 | 75.78 | 72.52 | 61.79 | 68.43 | 59.72 | 34.64 | 37.4 | |
| GradMAPRatio=25.00%, Backbone=LLaMA2-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=169.422026.02 | 58.59 | — | — | — | — | — | — | — | |
| MKARatio=12.50%, Backbone=Vicuna-7B, Pruning Time=181.012026.02 | 58.31 | — | — | — | — | — | — | — | |
| GradMAPRatio=25.00%, Backbone=Vicuna-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=194.132026.02 | 58.06 | — | — | — | — | — | — | — | |
| MKARatio=12.50%, Backbone=LLaMA2-7B, Pruning Time=179.552026.02 | 57.8 | — | — | — | — | — | — | — | |
| SLEBRatio=25.00%, Backbone=LLaMA2-13B, Pruning Time=2464.182026.02 | 57.79 | — | — | — | — | — | — | — | |
| SLEBRatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 57.6 | 65 | 75 | 65.7 | 57.9 | 67.06 | 36.6 | 35.8 | |
| ROSESparsity=60%, Backbone=LLaMA3-8B2026.03 | 57.52 | 76.39 | 72.2 | 61.61 | 66.69 | 57.58 | 33.19 | 35 | |
| SparseGPT*Ratio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 57.34 | 71.13 | 75.24 | 51.58 | 67.56 | 68.98 | 36.09 | 30.8 | |
| WANDAPruning Ratio=20%, Reproduction=true2023.05 | 57.23 | 65.75 | 74.7 | 64.52 | 59.35 | 60.65 | 36.26 | 39.4 | |
| WANDA*Ratio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 57.23 | 65.75 | 74.7 | 64.52 | 59.35 | 60.65 | 36.26 | 39.4 | |
| GradMAPRatio=25.00%, Backbone=LLaMA2-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=123.422026.02 | 56.08 | — | — | — | — | — | — | — | |
| SLEBRatio=25.00%, Backbone=LLaMA2-7B, Pruning Time=801.582026.02 | 55.96 | — | — | — | — | — | — | — | |
| GradMAPRatio=25.00%, Backbone=Vicuna-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=148.632026.02 | 55.88 | — | — | — | — | — | — | — | |
| SLEBRatio=25.00%, Backbone=Vicuna-7B, Pruning Time=902.322026.02 | 54.88 | — | — | — | — | — | — | — | |
| MKARatio=25.00%, Backbone=LLaMA2-13B, Pruning Time=542.562026.02 | 54.26 | — | — | — | — | — | — | — | |
| ShortGPTRatio=25.00%, Backbone=LLaMA2-7B, Pruning Time=831.162026.02 | 54.19 | — | — | — | — | — | — | — | |
| MagnitudePruning Ratio=20%, Reproduction=true2023.05 | 53.59 | 61.89 | 70.81 | 58.34 | 56.87 | 54.87 | 34.02 | 38.4 | |
| MagnitudeRatio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 53.59 | 61.89 | 70.81 | 58.34 | 56.87 | 54.87 | 34.02 | 38.4 | |
| MKARatio=25.00%, Backbone=LLaMA2-7B, Pruning Time=182.582026.02 | 53.44 | — | — | — | — | — | — | — | |
| ShortGPTRatio=25.00%, Backbone=Vicuna-7B, Pruning Time=923.202026.02 | 52.98 | — | — | — | — | — | — | — | |
| MKARatio=25.00%, Backbone=Vicuna-7B, Pruning Time=199.162026.02 | 52.56 | — | — | — | — | — | — | — | |
| HyWIARatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 51.8 | 60.55 | 72.36 | 55.25 | 55.09 | 50.84 | 31.48 | 37 | |
| SparseGPT*Ratio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 51.24 | 64.52 | 69.9 | 43.29 | 64.95 | 61.86 | 30.37 | 23.8 | |
| GradMAPRatio=40.63%, Backbone=LLaMA2-13B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=522.062026.02 | 50.39 | — | — | — | — | — | — | — | |
| FLAPRatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 50.37 | 60.21 | 67.52 | 52.14 | 57.54 | 49.66 | 29.95 | 35.6 | |
| GradMAPRatio=40.63%, Backbone=LLaMA2-13B, Pruning Stage=Stage 1 (layer selection), Pruning Time=447.912026.02 | 50.11 | — | — | — | — | — | — | — | |
| LoRAPrunePruning Ratio=50%2023.05 | 49.71 | 61.88 | 71.53 | 47.86 | 55.01 | 45.13 | 31.62 | 34.98 | |
| LoRAPruneRatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 49.71 | 61.88 | 71.53 | 47.86 | 55.01 | 45.13 | 31.62 | 34.98 | |
| LoRAPrune-8bitPruning Ratio=50%, Quantization=8-bit2023.05 | 49.56 | 61.43 | 70.88 | 47.65 | 55.12 | 45.78 | 30.5 | 35.62 | |
| LoRAPrune-8bitRatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 49.56 | 61.43 | 70.88 | 47.65 | 55.12 | 45.78 | 30.5 | 35.62 | |
| ShortGPTRatio=40.63%, Backbone=LLaMA2-13B, Pruning Time=3530.802026.02 | 49.34 | — | — | — | — | — | — | — | |
| Element2Ratio=20%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 48.98 | 61.47 | 68.82 | 47.56 | 55.09 | 46.46 | 28.24 | 35.2 | |
| Element2Ratio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 48.98 | 61.47 | 68.82 | 47.56 | 55.09 | 46.46 | 28.24 | 35.2 | |
| LLM-PrunerPruning Ratio=50%2023.05 | 48.69 | 60.28 | 69.31 | 47.06 | 53.43 | 45.96 | 29.18 | 35.6 | |
| GradMAPRatio=40.63%, Backbone=Vicuna-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=248.342026.02 | 47.93 | — | — | — | — | — | — | — | |
| GradMAPRatio=40.63%, Backbone=Vicuna-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=198.802026.02 | 47.88 | — | — | — | — | — | — | — | |
| MKARatio=40.63%, Backbone=LLaMA2-13B, Pruning Time=564.522026.02 | 47.45 | — | — | — | — | — | — | — | |
| MKARatio=40.63%, Backbone=Vicuna-7B, Pruning Time=211.602026.02 | 47.19 | — | — | — | — | — | — | — | |
| SLEBRatio=40.63%, Backbone=LLaMA2-13B, Pruning Time=3411.582026.02 | 46.77 | — | — | — | — | — | — | — | |
| SLEBRatio=40.63%, Backbone=Vicuna-7B, Pruning Time=1204.442026.02 | 46.74 | — | — | — | — | — | — | — | |
| GradMAPRatio=40.63%, Backbone=LLaMA2-7B, Pruning Stage=Stage 2 (projection compensation), Pruning Time=239.842026.02 | 46.72 | — | — | — | — | — | — | — | |
| VectorRatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 46.61 | 62.11 | 64.96 | 40.52 | 51.54 | 46.38 | 28.33 | 32.4 | |
| GradMAPRatio=40.63%, Backbone=LLaMA2-7B, Pruning Stage=Stage 1 (layer selection), Pruning Time=190.602026.02 | 46.26 | — | — | — | — | — | — | — | |
| MKARatio=40.63%, Backbone=LLaMA2-7B, Pruning Time=197.082026.02 | 46.14 | — | — | — | — | — | — | — | |
| ROSESparsity=70%, Backbone=LLaMA3-8B2026.03 | 46.07 | 68.65 | 61.7 | 40.6 | 57.14 | 40.74 | 24.83 | 28.8 | |
| SparseGPTSparsity=70%, Backbone=LLaMA3-8B2026.03 | 46 | 68.78 | 61.15 | 41.12 | 55.88 | 40.11 | 25.34 | 29.6 | |
| WANDAPruning Ratio=50%, Reproduction=true2023.05 | 45.43 | 50.9 | 57.38 | 38.12 | 55.98 | 42.68 | 34.2 | 38.78 | |
| WANDA*Ratio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 45.43 | 50.9 | 57.38 | 38.12 | 55.98 | 42.68 | 34.2 | 38.78 | |
| ShortGPTRatio=40.63%, Backbone=LLaMA2-7B, Pruning Time=1244.632026.02 | 44.87 | — | — | — | — | — | — | — | |
| ShortGPTRatio=40.63%, Backbone=Vicuna-7B, Pruning Time=1247.182026.02 | 42.46 | — | — | — | — | — | — | — | |
| SLEBRatio=40.63%, Backbone=LLaMA2-7B, Pruning Time=1201.522026.02 | 42.4 | — | — | — | — | — | — | — | |
| MagnitudePruning Ratio=50%, Reproduction=true2023.05 | 40.42 | 47.4 | 54.36 | 33.49 | 53.1 | 37.88 | 26.6 | 30.12 | |
| MagnitudeRatio=50%, Zero-shot evaluation=true, Model Backbone=LLaMA-7B2024.03 | 40.42 | 47.4 | 54.36 | 33.49 | 53.1 | 37.88 | 26.6 | 30.12 | |
| ROSESparsity=80%, Backbone=LLaMA3-8B2026.03 | 38.27 | 56.67 | 53.48 | 28.36 | 50.59 | 30.01 | 21.76 | 27 | |
| SparseGPTSparsity=80%, Backbone=LLaMA3-8B2026.03 | 37.3 | 53.76 | 53.48 | 28.37 | 49.25 | 30.3 | 20.31 | 25.6 | |
| ROSESparsity=90%, Backbone=LLaMA3-8B2026.03 | 35.34 | 37.83 | 52.94 | 26.99 | 51.3 | 28.07 | 23.04 | 27.2 | |
| SparseGPTSparsity=90%, Backbone=LLaMA3-8B2026.03 | 34.58 | 38.07 | 51.52 | 27.07 | 48.62 | 28.49 | 23.29 | 25 |