Zero-shot Common Sense Reasoning on PIQA, HellaSwag, WinoGrande, ARC-e, and ARC-c (test)
82.7PIQALLaMA-2 70B
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LLaMA-2 70BPruning Type=Original, Sparsity=0%2026.02 | 82.7 | 83.84 | 77.98 | 80.98 | 57.34 | 76.57 | |
| DenseSparsity=0%, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 80.47 | 79.37 | 72.22 | 77.48 | 49.23 | 71.75 | |
| LLaMA-2 13BPruning Type=Original, Sparsity=0%2026.02 | 80.47 | 79.39 | 72.22 | 77.48 | 49.23 | 71.76 | |
| FlattenGPTBase Model=LLaMA-2 70B, Pruning Type=Depth, Sparsity=19.84%2026.02 | 80.36 | 81.42 | 77.35 | 77.48 | 53.07 | 73.94 | |
| WandaBase Model=LLaMA-2 70B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 80.3 | 79.22 | 74.66 | 76.35 | 51.19 | 72.34 | |
| SLEBBase Model=LLaMA-2 70B, Pruning Type=Depth, Sparsity=19.84%2026.02 | 80.14 | 77.21 | 72.93 | 75.38 | 48.38 | 70.81 | |
| SpareGPTBase Model=LLaMA-2 70B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 80.03 | 76.09 | 76.56 | 76.94 | 49.74 | 71.87 | |
| Qwen-1.5 14BPruning Type=Original, Sparsity=0%2026.02 | 79.87 | 79.41 | 70.56 | 68.48 | 47.01 | 69.07 | |
| DensePrune Rate=0%, Backbone=LLaMA-3-8B2024.06 | 79.71 | 60.19 | 72.61 | 80.09 | 50.34 | 68.59 | |
| Qwen-1.5 7BPruning Type=Original, Sparsity=0%2026.02 | 79.22 | 76.92 | 66.46 | 62.16 | 42.66 | 65.48 | |
| LLaMA-2 7BPruning Type=Original, Sparsity=0%2026.02 | 79.11 | 75.99 | 69.06 | 74.58 | 46.25 | 69 | |
| Baichuan-2 13BPruning Type=Original, Sparsity=0%2026.02 | 78.84 | 75.23 | 70.4 | 74.07 | 47.7 | 69.25 | |
| UniCuCoSparsity=50%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 77.91 | 56.02 | 73.4 | 76.94 | 43.26 | 65.51 | |
| UniformSparsity=50%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 77.86 | 55.98 | 73.09 | 76.98 | 43.17 | 65.42 | |
| EvoPressSparsity=50%, Latency=139m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 77.69 | 56.49 | 71.9 | 76.85 | 42.75 | 65.14 | |
| Baichuan-2 7BPruning Type=Original, Sparsity=0%2026.02 | 77.48 | 72.18 | 68.27 | 72.98 | 42.75 | 66.73 | |
| OWLSparsity=50%, Latency=114m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 77.26 | 56.2 | 72.85 | 77.23 | 45.22 | 65.75 | |
| BlockPrunerBase Model=LLaMA-2 13B, Pruning Type=Depth, Sparsity=25.12%2026.02 | 76.93 | 72.2 | 66.3 | 65.82 | 41.38 | 64.53 | |
| LLaMA3.2Size=3B, #Tokens=-, Zero-shot protocol=true2026.01 | 76.8 | 55.2 | 69.1 | 74.5 | 42.2 | 63.6 | |
| SliceGPTBase Model=LLaMA-2 70B, Pruning Type=Width, Sparsity=20%2026.02 | 76.61 | 72.98 | 74.92 | 80.51 | 55.2 | 72.34 | |
| FlattenGPTBase Model=LLaMA-2 13B, Pruning Type=Depth, Sparsity=24.37%2026.02 | 76.33 | 73.44 | 71.11 | 72.1 | 44.54 | 67.5 | |
| ShortGPTBase Model=LLaMA-2 70B, Pruning Type=Depth, Sparsity=19.84%2026.02 | 76.02 | 78.87 | 71.96 | 76.02 | 52.95 | 71.68 | |
| HESTIASize=3B, #Tokens=10B, Zero-shot protocol=true2026.01 | 75.7 | 49.1 | 62.2 | 74.1 | 39.3 | 60.1 | |
| EvoPressSparsity=60%, Latency=138m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 75.52 | 50.43 | 70.74 | 73.11 | 39.25 | 61.81 | |
| SpareGPTBase Model=LLaMA-2 13B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 75.46 | 65.52 | 68.51 | 66.04 | 39.76 | 63.06 | |
| OWLSparsity=60%, Latency=98m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 75.35 | 51.08 | 69.85 | 72.69 | 39.85 | 61.76 | |
| BlockPrunerBase Model=Qwen-1.5 14B, Pruning Type=Depth, Sparsity=23.72%2026.02 | 75.24 | 66.92 | 61.48 | 59.51 | 39.08 | 60.45 | |
| UniCuCoSparsity=60%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 74.7 | 49.92 | 70.56 | 73.11 | 39.33 | 61.52 | |
| UniformSparsity=60%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 74.59 | 48.83 | 69.85 | 72.05 | 38.14 | 60.69 | |
| SpectraSize=3.9B, #Tokens=100B, Zero-shot protocol=true2026.01 | 74.4 | 48.3 | 63.1 | 66 | 31.9 | 56.7 | |
| BlockPrunerBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.99%2026.02 | 74.21 | 65.87 | 62.43 | 61.07 | 37.29 | 60.17 | |
| LLaMA3.2Size=1B, #Tokens=-, Zero-shot protocol=true2026.01 | 74.2 | 47.7 | 60.3 | 65.4 | 31.3 | 55.8 | |
| FlattenGPTBase Model=Qwen-1.5 14B, Pruning Type=Depth, Sparsity=22.25%2026.02 | 74.1 | 68.57 | 65.59 | 65.03 | 40.78 | 62.81 | |
| WandaBase Model=LLaMA-2 13B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 73.94 | 63.09 | 67.01 | 64.31 | 37.8 | 61.23 | |
| TequilaSize=3B, #Tokens=10B, Zero-shot protocol=true2026.01 | 73.9 | 46.4 | 62.7 | 70.2 | 34.6 | 57.6 | |
| RMBase Model=LLaMA-2 13B, Pruning Type=Depth, Sparsity=24.37%2026.02 | 73.72 | 66.8 | 66.61 | 66.12 | 41.98 | 63.05 | |
| SLEBBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.02%2026.02 | 73.07 | 62.47 | 58.96 | 56.48 | 33.02 | 56.8 | |
| FlattenGPTBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.02%2026.02 | 72.74 | 68.45 | 66.54 | 63.43 | 41.3 | 62.49 | |
| ShortGPTBase Model=LLaMA-2 13B, Pruning Type=Depth, Sparsity=24.37%2026.02 | 72.74 | 67.8 | 70.8 | 60.35 | 41.3 | 62.6 | |
| HESTIASize=1B, #Tokens=10B, Zero-shot protocol=true2026.01 | 72.6 | 41.8 | 57.7 | 66.5 | 35 | 54.7 | |
| SliceGPTBase Model=LLaMA-2 7B, Pruning Type=Width, Sparsity=21.45%2026.02 | 72.42 | 56.04 | 59.91 | 63.64 | 37.12 | 57.83 | |
| LaCoBase Model=LLaMA-2 13B, Pruning Type=Depth, Sparsity=24.37%2026.02 | 72.42 | 60.44 | 59.27 | 54.34 | 34.56 | 56.21 | |
| PG pruningPrune Rate=30%, Backbone=LLaMA-3-8B2024.06 | 72.25 | 43.56 | 59.04 | 59.85 | 29.44 | 52.83 | |
| SpareGPTBase Model=LLaMA-2 7B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 72.14 | 58.93 | 64.96 | 60.9 | 34.22 | 58.23 | |
| EvoPressSparsity=70%, Latency=138m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 71.76 | 44.31 | 67.25 | 69.19 | 33.87 | 57.28 | |
| BlockPrunerBase Model=Qwen-1.5 7B, Pruning Type=Depth, Sparsity=21.83%2026.02 | 71.71 | 59.31 | 55.56 | 53.7 | 33.28 | 54.71 | |
| LaCoBase Model=Qwen-1.5 14B, Pruning Type=Depth, Sparsity=22.25%2026.02 | 71.55 | 60.16 | 58.33 | 53.7 | 34.04 | 55.56 | |
| BitNetSize=3B, #Tokens=100B, Zero-shot protocol=true2026.01 | 71.5 | 42.9 | 59.3 | 61.4 | 28.3 | 52.7 | |
| BlockPrunerBase Model=Baichuan-2 13B, Pruning Type=Depth, Sparsity=24.19%2026.02 | 71.44 | 64.2 | 64.01 | 59.81 | 37.88 | 59.47 | |
| LLM-PrunerPrune Rate=30%, Backbone=LLaMA-3-8B2024.06 | 71.38 | 37.84 | 55.64 | 57.78 | 27.21 | 49.97 | |
| FlattenGPTBase Model=Baichuan-2 13B, Pruning Type=Depth, Sparsity=22.68%2026.02 | 71.22 | 65.27 | 68.19 | 58.75 | 37.03 | 60.09 | |
| TequilaSize=1B, #Tokens=10B, Zero-shot protocol=true2026.01 | 71 | 39.1 | 54.2 | 64.5 | 30.5 | 51.9 | |
| LaCoBase Model=Baichuan-2 13B, Pruning Type=Depth, Sparsity=22.68%2026.02 | 70.89 | 54 | 58.01 | 57.11 | 32.94 | 54.59 | |
| WandaBase Model=LLaMA-2 7B, Pruning Type=Width, Sparsity=2:4 (50%)2026.02 | 70.84 | 55.33 | 62.27 | 57.58 | 31.91 | 55.59 | |
| OWLSparsity=70%, Latency=121m, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 70.57 | 41.53 | 67.09 | 65.28 | 31.31 | 55.16 | |
| LaCoBase Model=Qwen-1.5 7B, Pruning Type=Depth, Sparsity=20.97%2026.02 | 70.4 | 56.35 | 58.64 | 46.89 | 32.85 | 53.03 | |
| ShortGPTBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.02%2026.02 | 70.24 | 62.63 | 65.9 | 56.06 | 36.09 | 58.18 | |
| BlockPrunerBase Model=Baichuan-2 7B, Pruning Type=Depth, Sparsity=22.45%2026.02 | 69.75 | 58.09 | 61.48 | 58.08 | 33.02 | 56.08 | |
| ShortGPTBase Model=Qwen-1.5 7B, Pruning Type=Depth, Sparsity=20.97%2026.02 | 69.53 | 58.87 | 62.12 | 43.6 | 32.17 | 53.26 | |
| FlattenGPTBase Model=Baichuan-2 7B, Pruning Type=Depth, Sparsity=21.57%2026.02 | 69.42 | 61.5 | 64.33 | 56.27 | 35.24 | 57.35 | |
| UniCuCoSparsity=70%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 69.31 | 39.71 | 64.64 | 65.66 | 31.23 | 54.11 | |
| ShortGPTBase Model=Baichuan-2 13B, Pruning Type=Depth, Sparsity=22.68%2026.02 | 69.31 | 61.71 | 68.27 | 56.52 | 36.69 | 58.5 | |
| SpectraSize=1.1B, #Tokens=100B, Zero-shot protocol=true2026.01 | 69.3 | 38.8 | 55.5 | 56.3 | 24.6 | 48.9 | |
| RMBase Model=Baichuan-2 13B, Pruning Type=Depth, Sparsity=22.68%2026.02 | 68.99 | 63.78 | 67.88 | 57.49 | 37.54 | 59.14 | |
| BitNetSize=1.3B, #Tokens=100B, Zero-shot protocol=true2026.01 | 68.8 | 37.7 | 55.8 | 54.9 | 24.2 | 48.3 | |
| SliceGPTBase Model=LLaMA-2 13B, Pruning Type=Width, Sparsity=25%2026.02 | 68.55 | 58.1 | 67.48 | 62.5 | 37.88 | 58.9 | |
| FlattenGPTBase Model=Qwen-1.5 7B, Pruning Type=Depth, Sparsity=20.97%2026.02 | 68.39 | 62.89 | 59.27 | 56.99 | 37.46 | 57 | |
| SliceGPTPrune Rate=30%, Backbone=LLaMA-3-8B2024.06 | 68.34 | 53.92 | 57.22 | 49.41 | 28.07 | 51.39 | |
| LaCoBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.02%2026.02 | 68.34 | 54.08 | 60.46 | 55.39 | 35.84 | 54.82 | |
| LaCoBase Model=Baichuan-2 7B, Pruning Type=Depth, Sparsity=21.57%2026.02 | 68.28 | 51.5 | 58.56 | 52.9 | 28.5 | 51.95 | |
| PG pruningPrune Rate=40%, Backbone=LLaMA-3-8B2024.06 | 67.63 | 37.36 | 56.91 | 50.67 | 24.91 | 47.5 | |
| RMBase Model=Qwen-1.5 7B, Pruning Type=Depth, Sparsity=20.97%2026.02 | 67.36 | 42 | 49.88 | 54.17 | 28.58 | 48.4 | |
| RMBase Model=Qwen-1.5 14B, Pruning Type=Depth, Sparsity=22.25%2026.02 | 67.08 | 42.08 | 53.28 | 50.72 | 29.01 | 48.43 | |
| LLM-PrunerPrune Rate=40%, Backbone=LLaMA-3-8B2024.06 | 66.26 | 31.9 | 54.06 | 49.74 | 22.52 | 44.9 | |
| BonsaiPrune Rate=30%, Backbone=LLaMA-3-8B2024.06 | 64.53 | 36.1 | 55.09 | 47.64 | 22.52 | 45.18 | |
| SEQSize=3B, #Tokens=10B, Zero-shot protocol=true2026.01 | 64.5 | 30.3 | 52.9 | 49.8 | 23.1 | 44.1 | |
| ShortGPTBase Model=Baichuan-2 7B, Pruning Type=Depth, Sparsity=21.57%2026.02 | 63.71 | 50.01 | 62.67 | 47.31 | 30.72 | 50.88 | |
| LLM-QATSize=3B, #Tokens=100B, Zero-shot protocol=true2026.01 | 62.7 | 43.4 | 50.6 | 44.5 | 30.7 | 46.4 | |
| PG pruningPrune Rate=50%, Backbone=LLaMA-3-8B2024.06 | 62.51 | 30.89 | 51.85 | 41.12 | 20.65 | 41.4 | |
| UniformSparsity=70%, Latency=<1s, Backbone=Llama-2-13B, Zero-shot evaluation=true2025.05 | 62.13 | 32.17 | 57.14 | 52.19 | 24.06 | 45.54 | |
| SliceGPTPrune Rate=40%, Backbone=LLaMA-3-8B2024.06 | 61.53 | 39.98 | 52.8 | 36.66 | 25.17 | 43.23 | |
| LLM-PrunerPrune Rate=50%, Backbone=LLaMA-3-8B2024.06 | 61.15 | 29.1 | 51.93 | 39.98 | 19.36 | 40.3 | |
| SEQSize=1B, #Tokens=10B, Zero-shot protocol=true2026.01 | 60.4 | 27.3 | 51 | 42.1 | 18 | 39.8 | |
| RMBase Model=Baichuan-2 7B, Pruning Type=Depth, Sparsity=21.57%2026.02 | 59.96 | 30.87 | 52.33 | 38.17 | 23.63 | 40.99 | |
| Wanda-spPrune Rate=30%, Backbone=LLaMA-3-8B2024.06 | 59.74 | 31.46 | 52.64 | 44.02 | 19.88 | 41.55 | |
| BonsaiPrune Rate=40%, Backbone=LLaMA-3-8B2024.06 | 58.81 | 29.43 | 48.93 | 33.21 | 18.15 | 37.71 | |
| ShortGPTBase Model=Qwen-1.5 14B, Pruning Type=Depth, Sparsity=22.25%2026.02 | 58.6 | 36.16 | 55.96 | 38.09 | 34.81 | 44.72 | |
| DLTSize=3B, #Tokens=10B, Zero-shot protocol=true2026.01 | 57.2 | 26 | 49.6 | 36.1 | 16.1 | 37 | |
| Wanda-spPrune Rate=40%, Backbone=LLaMA-3-8B2024.06 | 56.58 | 27.46 | 50.35 | 32.07 | 17.06 | 36.7 | |
| SliceGPTPrune Rate=50%, Backbone=LLaMA-3-8B2024.06 | 56.37 | 32.66 | 48.38 | 32.45 | 22.1 | 38.39 | |
| DLTSize=1B, #Tokens=10B, Zero-shot protocol=true2026.01 | 56.3 | 25.6 | 51.3 | 42.4 | 17.4 | 38.6 | |
| BonsaiPrune Rate=50%, Backbone=LLaMA-3-8B2024.06 | 55.66 | 26.94 | 50.51 | 30.64 | 17.83 | 36.32 | |
| Wanda-spPrune Rate=50%, Backbone=LLaMA-3-8B2024.06 | 55.39 | 27.07 | 49.72 | 29.59 | 18.26 | 36.01 | |
| LLM-QATSize=1B, #Tokens=100B, Zero-shot protocol=true2026.01 | 55.1 | 31.3 | 49.6 | 36 | 26.2 | 39.7 | |
| RMBase Model=LLaMA-2 7B, Pruning Type=Depth, Sparsity=21.02%2026.02 | 54.46 | 29.22 | 49.25 | 34.43 | 22.53 | 37.98 |