Controllability on HarmBench
87.5HarmBench ScoreAAS
Evaluation Results
| Method | Links | |
|---|---|---|
| AASModel=Qwen2.5-7B, Steering Configuration=Best steering2026.01 | 87.5 | |
| Selective SteeringModel=Qwen2.5-7B, Steering Configuration=Best steering2026.01 | 87.5 | |
| ActAddModel=Qwen2.5-7B, Steering Configuration=Best steering2026.01 | 86.54 | |
| SASModel=Qwen2.5-7B, Steering Configuration=Best steering2026.01 | 86.54 | |
| Selective SteeringModel=Llama-3.2-3B, Steering Configuration=Best steering2026.01 | 85.58 | |
| AASModel=Llama-3.2-3B, Steering Configuration=Best steering2026.01 | 84.62 | |
| Selective SteeringModel=Qwen2.5-3B, Steering Configuration=Best steering2026.01 | 84.62 | |
| ActAddModel=Llama-3.2-3B, Steering Configuration=Best steering2026.01 | 82.69 | |
| SASModel=Llama-3.2-3B, Steering Configuration=Best steering2026.01 | 82.69 | |
| Selective SteeringModel=gemma-2-2b, Steering Configuration=Best steering2026.01 | 82.69 | |
| Selective SteeringModel=Llama-3.2-1B, Steering Configuration=Best steering2026.01 | 79.81 | |
| AASModel=Llama-3.1-8B, Steering Configuration=Best steering2026.01 | 77.88 | |
| Selective SteeringModel=Llama-3.1-8B, Steering Configuration=Best steering2026.01 | 77.88 | |
| AASModel=Llama-3.2-1B, Steering Configuration=Best steering2026.01 | 76.92 | |
| ActAddModel=Llama-3.1-8B, Steering Configuration=Best steering2026.01 | 74.04 | |
| SASModel=Llama-3.1-8B, Steering Configuration=Best steering2026.01 | 74.04 | |
| Selective SteeringModel=Qwen2.5-1.5B, Steering Configuration=Best steering2026.01 | 74.04 | |
| AASModel=gemma-2-2b, Steering Configuration=Best steering2026.01 | 74.04 | |
| ActAddModel=Llama-3.2-1B, Steering Configuration=Best steering2026.01 | 70.19 | |
| SASModel=Llama-3.2-1B, Steering Configuration=Best steering2026.01 | 70.19 | |
| AASModel=Qwen2.5-3B, Steering Configuration=Best steering2026.01 | 70.19 | |
| Selective SteeringModel=gemma-2-9b, Steering Configuration=Best steering2026.01 | 68.27 | |
| AASModel=gemma-2-9b, Steering Configuration=Best steering2026.01 | 67.31 | |
| DirAblModel=Qwen2.5-7B, Steering Configuration=Best steering2026.01 | 55.77 | |
| DirAblModel=Llama-3.2-1B, Steering Configuration=Best steering2026.01 | 54.81 | |
| DirAblModel=Llama-3.2-3B, Steering Configuration=Best steering2026.01 | 53.85 | |
| DirAblModel=Qwen2.5-3B, Steering Configuration=Best steering2026.01 | 52.88 | |
| ActAddModel=Qwen2.5-3B, Steering Configuration=Best steering2026.01 | 50.96 | |
| SASModel=Qwen2.5-3B, Steering Configuration=Best steering2026.01 | 50.96 | |
| AASModel=Qwen2.5-1.5B, Steering Configuration=Best steering2026.01 | 39.42 | |
| DirAblModel=Llama-3.1-8B, Steering Configuration=Best steering2026.01 | 32.69 | |
| DirAblModel=Qwen2.5-1.5B, Steering Configuration=Best steering2026.01 | 25 | |
| DirAblModel=gemma-2-2b, Steering Configuration=Best steering2026.01 | 25 | |
| ActAddModel=Qwen2.5-1.5B, Steering Configuration=Best steering2026.01 | 13.46 | |
| SASModel=Qwen2.5-1.5B, Steering Configuration=Best steering2026.01 | 13.46 | |
| DirAblModel=gemma-2-9b, Steering Configuration=Best steering2026.01 | 11.54 | |
| ActAddModel=gemma-2-2b, Steering Configuration=Best steering2026.01 | 0 | |
| SASModel=gemma-2-2b, Steering Configuration=Best steering2026.01 | 0 | |
| ActAddModel=gemma-2-9b, Steering Configuration=Best steering2026.01 | 0 | |
| SASModel=gemma-2-9b, Steering Configuration=Best steering2026.01 | 0 |