Safety Red-teaming on UnsafeBench
76.2ASRRedEdit
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RedEditTarget Detector=Qwen3.6-35B, k (proposals)=32026.06 | 76.2 | 93 | 1.26 | 164 | 125 | |
| Single-BestTarget Detector=Qwen3.6-35B, k (sampling budget)=162026.06 | 59.5 | 92.4 | 1 | 158 | 94 | |
| RandomTarget Detector=Qwen3.6-35B, k (sampling budget)=162026.06 | 32.5 | 91.4 | 1.94 | 160 | 52 | |
| ReAct AgentTarget Detector=Qwen3.6-35B2026.06 | 30.1 | 91.7 | 2.62 | 156 | 47 |