Question Answering on ScienceQA, CommonsenseQA, and OpenbookQA
93.12ScienceQA AccuracyFull-Parameter
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Full-ParameterBackbone=Llama2-7B, Fine-tuning Strategy=Full-Parameter2026.03 | 93.12 | 77.48 | 80.4 | 83.67 | |
| MoLABackbone=Llama2-7B, Adapter Granularity=layer-wise2026.03 | 91.91 | 77.89 | 82.8 | 84.2 | |
| AdaFuseBackbone=Llama2-7B2026.03 | 91.39 | 79.03 | 80.4 | 83.6 | |
| LoRABackbone=Llama2-7B, Fine-tuning Strategy=LoRA2026.03 | 91.01 | 75.51 | 77 | 81.17 | |
| MoRALBackbone=Llama2-7B, Adapter Granularity=layer-wise2026.03 | 90.74 | 76.41 | 76.6 | 81.25 | |
| PESCBackbone=Llama2-7B, Adapter Granularity=block-wise2026.03 | 90.02 | 76 | 78.4 | 81.47 | |
| Llama2-7B (base)Backbone=Llama2-7B2026.03 | 53.19 | 47.82 | 45.8 | 48.94 |