Safety Performance on JBB
53Refusal Score (CR)CAST
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CASTModel=QWEN 3 1.7B2025.10 | 53 | — | |
| AdaSteerModel=QWEN 3 1.7B2025.10 | 53 | — | |
| AlphaSteerModel=QWEN 3 1.7B2025.10 | 52 | — | |
| SurgicalModel=QWEN 3 1.7B2025.10 | 51 | — | |
| QWEN 3 1.7BModel=QWEN 3 1.7B2025.10 | 49 | — | |
| ELSModel=QWEN 3 1.7B2025.10 | 43 | — | |
| system promptModel=QWEN 3 1.7B2025.10 | 27 | — | |
| CASTModel=QWEN 3 14B2025.10 | 17 | — | |
| SurgicalModel=QWEN 3 14B2025.10 | 16 | — | |
| AdaSteerModel=QWEN 3 14B2025.10 | 16 | — | |
| CASTModel=QWEN 3 8B2025.10 | 14 | — | |
| QWEN 3 14BModel=QWEN 3 14B2025.10 | 14 | — | |
| AlphaSteerModel=QWEN 3 14B2025.10 | 14 | — | |
| AdaSteerModel=LLAMA3.1-8B-INST2025.10 | 13 | — | |
| SurgicalModel=QWEN 3 8B2025.10 | 13 | — | |
| AdaSteerModel=QWEN 3 8B2025.10 | 13 | — | |
| CASTModel=LLAMA3.1-8B-INST2025.10 | 12 | — | |
| QWEN 3 8BModel=QWEN 3 8B2025.10 | 12 | — | |
| AlphaSteerModel=QWEN 3 8B2025.10 | 12 | — | |
| SurgicalModel=LLAMA3.1-8B-INST2025.10 | 11 | — | |
| AlphaSteerModel=LLAMA3.1-8B-INST2025.10 | 11 | — | |
| ELSModel=QWEN 3 8B2025.10 | 11 | — | |
| LLAMA3.1-8B-INSTModel=LLAMA3.1-8B-INST2025.10 | 10 | — | |
| ELSModel=LLAMA3.1-8B-INST2025.10 | 10 | — | |
| ELSModel=QWEN 3 14B2025.10 | 10 | — | |
| CASTModel=LLAMA2-7B-CHAT2025.10 | 7 | — | |
| AlphaSteerModel=LLAMA2-7B-CHAT2025.10 | 6 | — | |
| system promptModel=QWEN 3 8B2025.10 | 6 | — | |
| SurgicalModel=LLAMA2-7B-CHAT2025.10 | 5 | — | |
| AdaSteerModel=LLAMA2-7B-CHAT2025.10 | 5 | — | |
| system promptModel=LLAMA3.1-8B-INST2025.10 | 3 | — | |
| LLAMA2-7B-CHATModel=LLAMA2-7B-CHAT2025.10 | 3 | — | |
| ELSModel=LLAMA2-7B-CHAT2025.10 | 3 | — | |
| system promptModel=QWEN 3 14B2025.10 | 3 | — | |
| system promptModel=LLAMA2-7B-CHAT2025.10 | 0 | — | |
| CAABase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | 30.1 | |
| Gradient CuffBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | 68 | |
| GSAEBase Model=Llama 3 8B, Method Category=Main Method, Random Seeds=52025.12 | — | 76.2 | |
| GSAE-1DBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | — | 40.1 | |
| Input Gate OnlyBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | — | 70.3 | |
| No gatingBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | — | 64.1 | |
| No SteeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | -9.1 | |
| Prompt guardrailsBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | 10.2 | |
| Random graphsBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | — | 44.2 | |
| SAE steeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | 36 | |
| SafeSwitchBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | — | 51.4 |