Text-only Jailbreak Attack Defense on GCG (test)
8.24ASRNullSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| NullSteerBackbone=MiniGPT-42026.03 | 8.24 | |
| ASTRABackbone=MiniGPT-42026.03 | 9.09 | |
| NullSteerBackbone=LLaVA-v1.52026.03 | 13.5 | |
| ASTRABackbone=LLaVA-v1.52026.03 | 14.55 | |
| NullSteerBackbone=Qwen2-VL2026.03 | 28.58 | |
| ASTRABackbone=Qwen2-VL2026.03 | 30.91 | |
| w/o defenseBackbone=MiniGPT-42026.03 | 58.18 | |
| w/o defenseBackbone=Qwen2-VL2026.03 | 81.82 | |
| w/o defenseBackbone=LLaVA-v1.52026.03 | 92.73 |