Model Steering on 10-task steering suite (excluding JailbreakBench) (test)
88.4AccuracyCLAS
Evaluation Results
| Method | Links | |
|---|---|---|
| CLASBase Model=Llama-3.1-70B2026.04 | 88.4 | |
| CLASBase Model=Qwen2.5-7B2026.04 | 86.3 | |
| ReFTBase Model=Llama-3.1-70B2026.04 | 84.9 | |
| CLASBase Model=Llama-3.1-8B2026.04 | 83.5 | |
| LoRABase Model=Llama-3.1-8B2026.04 | 82.9 | |
| LoRABase Model=Qwen2.5-7B2026.04 | 78.8 | |
| LoRABase Model=Llama-3.1-70B2026.04 | 77.9 | |
| ReFTBase Model=Qwen2.5-7B2026.04 | 76.8 | |
| LASBase Model=Llama-3.1-70B2026.04 | 75.1 | |
| LASBase Model=Llama-3.1-8B2026.04 | 74.6 | |
| ReFTBase Model=Llama-3.1-8B2026.04 | 73.7 | |
| ReFTBase Model=Llama-3.2-1B2026.04 | 65.1 | |
| LASBase Model=Qwen2.5-7B2026.04 | 64 | |
| LoRABase Model=Llama-3.2-1B2026.04 | 62.4 | |
| CLASBase Model=Llama-3.2-1B2026.04 | 58.3 | |
| LASBase Model=Llama-3.2-1B2026.04 | 52.5 |