Safety Performance on WildJailbreak
90.1Selective Refusal Score (Δs)GSAE
Evaluation Results
| Method | Links | |
|---|---|---|
| GSAEBase Model=Llama 3 8B, Method Category=Main Method, Random Seeds=52025.12 | 90.1 | |
| Input Gate OnlyBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 82.4 | |
| Gradient CuffBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 78.7 | |
| No gatingBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 78.1 | |
| SafeSwitchBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 65.4 | |
| Random graphsBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 60.1 | |
| GSAE-1DBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 55.4 | |
| SAE steeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 48.2 | |
| CAABase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 42 | |
| Prompt guardrailsBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 18.3 | |
| No SteeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | -3 |