Jailbreak Defense on Unified main malicious pool 2416 examples (held-out)
1.57ASRNeuroArmor-ReAnchor
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| NeuroArmor-ReAnchorModel=Llama-3-8B2026.06 | 1.57 | 21 | 0.87 | 96.61 | |
| NeuroArmor-ReAnchorBackbone=Llama-3-8B-Instruct2026.06 | 1.57 | — | — | — | |
| NeuroArmor-ReAnchorModel=Gemma-2-9B2026.06 | 1.9 | 32 | 1.32 | 95.69 | |
| Direct Safety SteeringModel=Llama-3-8B2026.06 | 23.92 | 443 | 18.34 | 28.55 | |
| Direct Safety SteeringBackbone=Llama-3-8B-Instruct2026.06 | 23.92 | — | — | — | |
| Runtime Intervention BaselineModel=Llama-3-8B2026.06 | 24.92 | 530 | 21.94 | 14.52 | |
| Runtime Intervention BaselineBackbone=Llama-3-8B-Instruct2026.06 | 24.92 | — | — | — | |
| Llama-Guard-3Model=Llama-3-8B2026.06 | 26.61 | 410 | 16.97 | 33.87 | |
| Llama-Guard-3Backbone=Llama-3-8B-Instruct2026.06 | 26.61 | — | — | — | |
| VanillaModel=Llama-3-8B2026.06 | 41.56 | 620 | 25.66 | 0 | |
| NeuroArmor-DetectModel=Llama-3-8B2026.06 | 41.56 | 620 | 25.66 | 0 | |
| VanillaBackbone=Llama-3-8B-Instruct2026.06 | 41.56 | — | — | — | |
| VanillaModel=Gemma-2-9B2026.06 | 53.52 | 743 | 30.75 | 0 | |
| NeuroArmor-DetectModel=Gemma-2-9B2026.06 | 53.52 | 743 | 30.75 | 0 |