Zero-shot Language Reasoning on BoolQ, PIQA, HellaSwag, WinoGrande, ARC-e, ARC-c, OBQA, and MTQA
82.11BoolQ AccuracyDense
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| DenseLayer/Ratio=0/0%, Base Model=LLaMA3.1-8B2026.04 | 82.11 | 80.25 | 60.01 | 73.88 | 81.48 | 51.45 | 33.4 | 39.6 | 62.77 | 100 | |
| ShortGPTLayer/Ratio=6/16.3%, Base Model=LLaMA3.1-8B2026.04 | 71.99 | 72.52 | 49.98 | 68.51 | 63.93 | 39.59 | 25.4 | 30.08 | 52.75 | 84.04 | |
| Ours (MSSD)Layer/Ratio=8/21.7%, Base Model=LLaMA3.1-8B2026.04 | 71.8 | 67.85 | 43.84 | 68.98 | 56.57 | 35.32 | 21.2 | 26.5 | 49.01 | 78.08 | |
| Ours (MASD)Layer/Ratio=8/21.7%, Base Model=LLaMA3.1-8B2026.04 | 71.8 | 67.85 | 43.84 | 68.98 | 56.57 | 35.32 | 21.2 | 26.5 | 49.01 | 78.08 | |
| Ours (MASD)Layer/Ratio=4/10.9%, Base Model=LLaMA3.1-8B2026.04 | 70.76 | 74.92 | 53.73 | 71.43 | 75.08 | 42.66 | 28.6 | 32.33 | 56.19 | 89.51 | |
| ShortGPTLayer/Ratio=4/10.9%, Base Model=LLaMA3.1-8B2026.04 | 70.55 | 75.73 | 55.04 | 70.32 | 72.22 | 44.2 | 31.2 | 32.6 | 56.48 | 89.98 | |
| Ours (MSSD)Layer/Ratio=6/16.3%, Base Model=LLaMA3.1-8B2026.04 | 70.37 | 71.82 | 49.4 | 70.32 | 66.96 | 39.51 | 26.6 | 29.15 | 53.02 | 84.47 | |
| Ours (MSSD)Layer/Ratio=4/10.9%, Base Model=LLaMA3.1-8B2026.04 | 69.79 | 74.65 | 53.54 | 72.22 | 74.96 | 43.86 | 30.2 | 33.74 | 56.62 | 90.2 | |
| Ours (MASD)Layer/Ratio=6/16.3%, Base Model=LLaMA3.1-8B2026.04 | 69.33 | 72.03 | 49.18 | 69.3 | 67.34 | 38.99 | 26.4 | 29.38 | 52.74 | 84.02 | |
| Ours (MSSD)Layer/Ratio=10/27.2%, Base Model=LLaMA3.1-8B2026.04 | 64.77 | 63.87 | 36.93 | 59.83 | 48.36 | 33.19 | 21 | 18.96 | 43.36 | 69.08 | |
| Ours (MASD)Layer/Ratio=10/27.2%, Base Model=LLaMA3.1-8B2026.04 | 64.77 | 63.87 | 36.93 | 59.83 | 48.36 | 33.19 | 21 | 18.96 | 43.36 | 69.08 | |
| Ours (MASD)Layer/Ratio=12/32.6%, Base Model=LLaMA3.1-8B2026.04 | 63.09 | 62.35 | 33.71 | 56.04 | 42.85 | 31.06 | 21.4 | 19.2 | 41.21 | 65.65 | |
| EntroDropLayer/Ratio=10/27.2%, Base Model=LLaMA3.1-8B2026.04 | 62.23 | 65.67 | 38.17 | 62.27 | 46.21 | 26.71 | 23 | 22.24 | 43.31 | 69 | |
| EntroDropLayer/Ratio=4/10.9%, Base Model=LLaMA3.1-8B2026.04 | 62.2 | 72.58 | 40.91 | 64.64 | 61.28 | 38.82 | 29.4 | 33.07 | 50.36 | 80.23 | |
| EntroDropLayer/Ratio=8/21.7%, Base Model=LLaMA3.1-8B2026.04 | 62.11 | 68.77 | 39.88 | 65.98 | 53.7 | 33.79 | 24 | 25.39 | 46.7 | 74.4 | |
| EntroDropLayer/Ratio=6/16.3%, Base Model=LLaMA3.1-8B2026.04 | 61.35 | 70.57 | 41.33 | 65.82 | 57.58 | 35.84 | 29.4 | 26.63 | 48.56 | 77.36 | |
| ShortGPTLayer/Ratio=12/32.6%, Base Model=LLaMA3.1-8B2026.04 | 55.96 | 60.12 | 30.62 | 55.09 | 40.03 | 27.56 | 17.8 | 22.28 | 38.68 | 61.62 | |
| ShortGPTLayer/Ratio=10/27.2%, Base Model=LLaMA3.1-8B2026.04 | 52.17 | 60.94 | 27.64 | 57.62 | 44.15 | 28.75 | 18.2 | 24.22 | 39.21 | 62.47 | |
| EntroDropLayer/Ratio=12/32.6%, Base Model=LLaMA3.1-8B2026.04 | 51.56 | 62.13 | 33.42 | 58.8 | 42.51 | 23.46 | 20.4 | 21.91 | 39.27 | 62.56 | |
| Ours (MSSD)Layer/Ratio=12/32.6%, Base Model=LLaMA3.1-8B2026.04 | 45.81 | 62.89 | 34.12 | 61.01 | 40.45 | 26.96 | 17.6 | 21.31 | 38.77 | 61.76 | |
| ShortGPTLayer/Ratio=8/21.7%, Base Model=LLaMA3.1-8B2026.04 | 37.55 | 60.5 | 28.21 | 54.3 | 42.85 | 27.3 | 18 | 25.76 | 36.81 | 58.64 |