Zero-shot Evaluation on ARC-C, ARC-E, HellaSwag, PIQA, and WinoGrande (test val)
76Average AccuracyFP16 Baseline
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FP16 BaselinePruning Ratio=0%, Backbone=Nemotron-H-8B, Fine-tuned=false2025.12 | 76 | — | — | — | — | — | |
| FP16 BaselinePruning Ratio=0%, Backbone=Bamba-v2-9B, Fine-tuned=false2025.12 | 74.6 | — | — | — | — | — | |
| FP16 BaselinePruning Ratio=0%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 74 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Nemotron-H-8B, Fine-tuned=true2025.12 | 73 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Bamba-v2-9B, Fine-tuned=true2025.12 | 72.9 | — | — | — | — | — | |
| FP16 BaselinePruning Ratio=0%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 72.4 | — | — | — | — | — | |
| MoDeGPTPruning Ratio=15%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 72.4 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Llama-3.1-8B, Fine-tuned=true2025.12 | 71.9 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Bamba-v2-9B, Fine-tuned=false2025.12 | 70.9 | — | — | — | — | — | |
| FP16 BaselinePruning Ratio=0%, Backbone=Mamba2-8B, Fine-tuned=false2025.12 | 70.6 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 70.5 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Qwen-2.5-7B, Fine-tuned=true2025.12 | 70 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Bamba-v2-9B, Fine-tuned=true2025.12 | 69.7 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Llama-3.1-8B, Fine-tuned=true2025.12 | 69.6 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=15%, Backbone=Qwen-2.5-7B, Fine-tuned=true2025.12 | 69.5 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 69.1 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Nemotron-H-8B, Fine-tuned=false2025.12 | 68.9 | — | — | — | — | — | |
| FP16 BaselinePruning Ratio=0%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 68.8 | — | — | — | — | — | |
| DenseSparsity=0%, Backbone=Mistral-7B2025.05 | 68.72 | 48.9 | 79.6 | 60.9 | 80.3 | 73.9 | |
| UniQLPruning Ratio=25%, Backbone=Nemotron-H-8B, Fine-tuned=true2025.12 | 67.3 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Llama-2-7B, Fine-tuned=true2025.12 | 67.2 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 67 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=25%, Backbone=Qwen-2.5-7B, Fine-tuned=true2025.12 | 66.8 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 66.7 | — | — | — | — | — | |
| UniQLPruning Ratio=15%, Backbone=Mamba2-8B, Fine-tuned=true2025.12 | 66.4 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Bamba-v2-9B, Fine-tuned=false2025.12 | 66.4 | — | — | — | — | — | |
| MoDeGPTPruning Ratio=15%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 66.2 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Qwen-2.5-7B, Fine-tuned=true2025.12 | 65.8 | — | — | — | — | — | |
| EvoPressSparsity=50%, Latency=122m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 65.74 | 44.88 | 76.85 | 56.46 | 79.16 | 71.35 | |
| UniQLPruning Ratio=15%, Backbone=Mamba2-8B, Fine-tuned=false2025.12 | 65.6 | — | — | — | — | — | |
| UniCuCoSparsity=50%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 65.23 | 44.11 | 76.73 | 55.66 | 78.4 | 71.27 | |
| UniformSparsity=50%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 65.1 | 43.7 | 76.7 | 55.7 | 78.4 | 71 | |
| OWLSparsity=50%, Latency=40m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 65 | 43.9 | 76.9 | 55.4 | 78.5 | 70.3 | |
| MoDeGPTPruning Ratio=25%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 64.9 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Llama-2-7B, Fine-tuned=true2025.12 | 64.9 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=15%, Backbone=Llama-2-7B, Fine-tuned=true2025.12 | 64.7 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=15%, Backbone=Llama-3.1-8B, Fine-tuned=true2025.12 | 64.5 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 63.7 | — | — | — | — | — | |
| MoDeGPTPruning Ratio=25%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 63.4 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Mamba2-8B, Fine-tuned=true2025.12 | 62.7 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=15%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 62.6 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=25%, Backbone=Llama-2-7B, Fine-tuned=true2025.12 | 62.4 | — | — | — | — | — | |
| UniQLPruning Ratio=25%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 62.1 | — | — | — | — | — | |
| UniCuCoSparsity=60%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 61.57 | 39.33 | 72.9 | 49.91 | 75.79 | 69.93 | |
| EvoPressSparsity=60%, Latency=121m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 61.1 | 38.05 | 72.56 | 49.91 | 76.01 | 68.98 | |
| UniformSparsity=60%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 60.82 | 38 | 72.4 | 49.4 | 75 | 69.3 | |
| UniQLPruning Ratio=25%, Backbone=Nemotron-H-8B, Fine-tuned=false2025.12 | 60.6 | — | — | — | — | — | |
| OWLSparsity=60%, Latency=40m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 60.52 | 38.5 | 71.9 | 46.9 | 75.1 | 70.2 | |
| UniQLPruning Ratio=25%, Backbone=Mamba2-8B, Fine-tuned=false2025.12 | 59.8 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=25%, Backbone=Llama-3.1-8B, Fine-tuned=true2025.12 | 59.5 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=15%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 56.7 | — | — | — | — | — | |
| EvoPressSparsity=70%, Latency=120m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 56.48 | 33.45 | 67.13 | 43.91 | 72.63 | 65.27 | |
| SVD-LLMPruning Ratio=15%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 56.3 | — | — | — | — | — | |
| SVD-LLMPruning Ratio=25%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 53.2 | — | — | — | — | — | |
| UniCuCoSparsity=70%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 52.87 | 29.1 | 64.27 | 39.05 | 69.04 | 62.9 | |
| MoDeGPTPruning Ratio=15%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 52.1 | — | — | — | — | — | |
| OWLSparsity=70%, Latency=40m, Zero-shot=true, Backbone=Mistral-7B2025.05 | 51.92 | 27.9 | 62.6 | 38.6 | 67 | 63.5 | |
| SVD-LLMPruning Ratio=25%, Backbone=Llama-2-7B, Fine-tuned=false2025.12 | 50.8 | — | — | — | — | — | |
| UniformSparsity=70%, Latency=<1s, Zero-shot=true, Backbone=Mistral-7B2025.05 | 49.88 | 27.1 | 60.9 | 36.1 | 65.9 | 59.4 | |
| SVD-LLMPruning Ratio=25%, Backbone=Llama-3.1-8B, Fine-tuned=false2025.12 | 45.8 | — | — | — | — | — | |
| MoDeGPTPruning Ratio=25%, Backbone=Qwen-2.5-7B, Fine-tuned=false2025.12 | 40.8 | — | — | — | — | — |