Jailbreak Attack Robustness on Jailbreak Attack Evaluation Set Llama-3 8B
100GCG Robustness ScoreGSAE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GSAEBackbone=Llama-3 8B, Type=Graph-Regularized Sparse Autoencoder2025.12 | 100 | 95.1 | 90.1 | 92.4 | |
| SAE steeringBackbone=Llama-3 8B2025.12 | 72.5 | 68.2 | 65 | 61.4 | |
| SafeSwitchBackbone=Llama-3 8B2025.12 | 68.3 | 84 | 40.1 | 39.5 | |
| Safety-Tuned BaselineBackbone=Llama-3 8B2025.12 | 65.4 | 55.3 | 60.1 | 50.3 | |
| CAABackbone=Llama-3 8B, Type=Contrastive Vector2025.12 | 58.1 | 55 | 49.3 | 46.5 | |
| Prompting guardrailsBackbone=Llama-3 8B2025.12 | 41.2 | 36.1 | 32.4 | 28 |