Question Answering on Commonsense QA (ARC, HellaSwag, OBQA, RTE, CoPa, Race)
82.83ARC-E AccuracyDense
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| DenseModel=Qwen-3-14B, Pruning Ratio (Lp/Lt)=0/40, Inference Mode=Zero-shot2026.05 | 82.83 | 60.24 | 78.82 | 72.85 | 89.3 | 46.2 | 77.62 | 90 | 43.16 | 71.22 | |
| DenseModel=OLMo-2-7B, Pruning Ratio (Lp/Lt)=0/32, Inference Mode=Zero-shot2026.05 | 81.19 | 56.14 | 78.93 | 74.66 | 78.23 | 45.2 | 71.12 | 89 | 40.1 | 68.29 | |
| DenseModel=LLaMA-2-7B, Pruning Ratio (Lp/Lt)=0/32, Inference Mode=Zero-shot2026.05 | 74.49 | 46.25 | 75.99 | 68.9 | 77.71 | 44.2 | 62.82 | 87 | 39.62 | 64.11 | |
| Ghost LayerModel=OLMo-2-7B, Pruning Ratio (Lp/Lt)=7/32, Inference Mode=Zero-shot, Training-free=true2026.05 | 67.67 | 41.81 | 63.62 | 70.88 | 70.24 | 37.87 | 70.04 | 80 | 37.42 | 59.95 | |
| Ghost LayerModel=LLaMA-2-7B, Pruning Ratio (Lp/Lt)=7/32, Inference Mode=Zero-shot, Training-free=true2026.05 | 52.99 | 33.79 | 58.01 | 66.38 | 70.28 | 35.8 | 53.43 | 76 | 37.7 | 53.82 | |
| Ghost LayerModel=LLaMA-2-7B, Pruning Ratio (Lp/Lt)=11/32, Inference Mode=Zero-shot, Training-free=true2026.05 | 49.81 | 31.83 | 49.28 | 65.82 | 73.7 | 33.8 | 57.04 | 72 | 33.01 | 51.81 | |
| Ghost LayerModel=OLMo-2-7B, Pruning Ratio (Lp/Lt)=11/32, Inference Mode=Zero-shot, Training-free=true2026.05 | 46.51 | 33.87 | 45.45 | 67.25 | 75.99 | 33 | 77.26 | 68 | 30.81 | 53.13 | |
| Ghost LayerModel=Qwen-3-14B, Pruning Ratio (Lp/Lt)=15/40, Inference Mode=Zero-shot, Training-free=true2026.05 | 44.78 | 27.39 | 43.57 | 53.91 | 43.39 | 30.6 | 51.99 | 67 | 29.86 | 43.61 | |
| Ghost LayerModel=Qwen-3-14B, Pruning Ratio (Lp/Lt)=13/40, Inference Mode=Zero-shot, Training-free=true2026.05 | 35.65 | 28.92 | 33.91 | 60.46 | 72.84 | 29 | 48.38 | 66 | 28.04 | 44.8 |