Commonsense Reasoning on ARC-e, OBQA, SIQA, ARC-c, WinoG, PIQA, BoolQ, HellaS (LLaMA3-8B)
92.9ARC-e AccuracyEffiLoRA
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| EffiLoRABackbone=LLaMA3-8B, Configuration=Multiple B, Param.=0.53%2025.11 | 92.9 | 87 | 81.7 | 81.8 | 88.4 | 89.6 | 74.1 | 95.8 | 86.4 | |
| HydraLoRABackbone=LLaMA3-8B, Param.=0.93%2025.11 | 92.4 | 87 | 82.6 | 81.9 | 87.8 | 88 | 73.6 | 96.2 | 86.1 | |
| AdaLoRABackbone=LLaMA3-8B, Param.=0.35%2025.11 | 90.4 | 85 | 76.7 | 79.1 | 83.3 | 86.4 | 75.1 | 75.4 | 81.4 | |
| GraphMoEBackbone=LLaMA3-8B, Param.=5.90%2025.11 | 90.3 | 88.2 | 79.4 | 80.6 | 83.7 | 88.8 | 75.9 | 95.3 | 85.3 | |
| EffiLoRABackbone=LLaMA3-8B, Configuration=Single B, Param.=0.18%2025.11 | 89.8 | 86.6 | 80.5 | 79.9 | 84.4 | 88.3 | 72.7 | 94.7 | 84.6 | |
| LoKrBackbone=LLaMA3-8B, Param.=0.01%2025.11 | 89.2 | 81.8 | 78.7 | 76.7 | 82.1 | 81.6 | 65.1 | 92 | 80.9 | |
| NoRABackbone=LLaMA3-8B, Param.=0.09%2025.11 | 88.2 | 85 | 79.1 | 77.5 | 84.3 | 86.4 | 73.3 | 94.1 | 83.1 | |
| LoRAMoEBackbone=LLaMA3-8B, Param.=3.20%2025.11 | 87.8 | 85 | 74.8 | 79.5 | 83.4 | 87.1 | 72.4 | 94.8 | 83.5 | |
| ShareLoRABackbone=LLaMA3-8B, Param.=0.18%2025.11 | 87.5 | 83.1 | 80.2 | 75 | 84 | 85.5 | 71 | 96.1 | 82.8 | |
| MixLoRABackbone=LLaMA3-8B, Param.=3.00%2025.11 | 86.5 | 84.8 | 78.8 | 79.9 | 82.1 | 87.6 | 75 | 93.3 | 83.5 | |
| MoLABackbone=LLaMA3-8B, Param.=2.70%2025.11 | 86.4 | 84.4 | 76.4 | 77.9 | 83.3 | 86.7 | 74 | 93.9 | 82.9 | |
| LoRA-FABackbone=LLaMA3-8B, Param.=0.17%2025.11 | 86.1 | 81 | 79.5 | 73.4 | 83.8 | 84.2 | 69 | 93.4 | 81.3 | |
| LoRABackbone=LLaMA3-8B, Param.=0.35%2025.11 | 84.2 | 79 | 79.9 | 71.2 | 84.3 | 85.2 | 70.8 | 91.7 | 80.8 |