General Utility on Utility Benchmark (test)
67.5AccuracyGemma2-9B
Evaluation Results
| Method | Links | |
|---|---|---|
| Gemma2-9BDefense Method=None2026.02 | 67.5 | |
| LATBase Model=Gemma2-9B2026.02 | 67.2 | |
| Fail-Closed AlignmentBase Model=Gemma2-9B2026.02 | 67.1 | |
| CATBase Model=Gemma2-9B2026.02 | 66.9 | |
| DeepRefusalBase Model=Gemma2-9B2026.02 | 66.2 | |
| CATBase Model=Llama3-8B2026.02 | 62.1 | |
| Llama3-8BDefense Method=None2026.02 | 61.2 | |
| Gemma2-2BDefense Method=None2026.02 | 61.2 | |
| Fail-Closed AlignmentBase Model=Llama3-8B2026.02 | 61 | |
| LATBase Model=Gemma2-2B2026.02 | 61 | |
| DeepRefusalBase Model=Llama3-8B2026.02 | 60.8 | |
| Fail-Closed AlignmentBase Model=Gemma2-2B2026.02 | 60.3 | |
| DeepRefusalBase Model=Gemma2-2B2026.02 | 60.1 | |
| CATBase Model=Gemma2-2B2026.02 | 59.9 | |
| CATBase Model=Llama2-7B2026.02 | 59.2 | |
| Augmented RTBase Model=Llama3-8B2026.02 | 59.1 | |
| LATBase Model=Llama3-8B2026.02 | 59.1 | |
| DeepRefusalBase Model=Llama2-7B2026.02 | 58.6 | |
| Llama2-7BDefense Method=None2026.02 | 58.5 | |
| Fail-Closed AlignmentBase Model=Llama2-7B2026.02 | 57.9 | |
| Augmented RTBase Model=Llama2-7B2026.02 | 57 | |
| LATBase Model=Llama2-7B2026.02 | 54.9 | |
| Augmented RTBase Model=Gemma2-9B2026.02 | 51.4 | |
| Augmented RTBase Model=Gemma2-2B2026.02 | 45.5 |