Commonsense Reasoning on HellaSwag (Accuracy, Rank)
70.5AccuracyBaseline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaselineBackbone=LLaMA-3.1-8B-Instruct2026.05 | 70.5 | — | |
| SKOPBackbone=LLaMA-3.1-8B-Instruct2026.05 | 65.2 | 2.69 | |
| Angular SteerBackbone=LLaMA-3.1-8B-Instruct2026.05 | 59.7 | 4.75 | |
| SADIBackbone=LLaMA-3.1-8B-Instruct2026.05 | 58.8 | 3.38 | |
| CASTBackbone=LLaMA-3.1-8B-Instruct2026.05 | 55 | 5.63 | |
| LoRABackbone=LLaMA-3.1-8B-Instruct2026.05 | 52.6 | — | |
| ITIBackbone=LLaMA-3.1-8B-Instruct2026.05 | 45.2 | 4.75 | |
| DISCO-QBackbone=LLaMA-3.1-8B-Instruct2026.05 | 38.3 | 4.75 | |
| Comm SteerBackbone=LLaMA-3.1-8B-Instruct2026.05 | 35.5 | 4.63 | |
| CAABackbone=LLaMA-3.1-8B-Instruct2026.05 | 27.5 | 5.44 |