Jailbreak Robustness on StrongReject Static PAIR threshold 0.75 (test)
0ASRBCT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BCTModel=Qwen3-4B-Instruct-2507, StrongReject threshold=0.752026.05 | 0 | 100 | |
| OPCTModel=Qwen3-4B-Instruct-2507, StrongReject threshold=0.752026.05 | 0 | 100 | |
| OPCTModel=Qwen3-8B, StrongReject threshold=0.752026.05 | 0 | 100 | |
| BCTModel=gpt-oss-20b, StrongReject threshold=0.752026.05 | 0 | 100 | |
| OPCTModel=gpt-oss-20b, StrongReject threshold=0.752026.05 | 1.1 | 98.9 | |
| BCTModel=Qwen3-8B, StrongReject threshold=0.752026.05 | 4.1 | 95.9 | |
| BaselineModel=Qwen3-4B-Instruct-2507, StrongReject threshold=0.752026.05 | 31.6 | 68.4 | |
| BaselineModel=gpt-oss-20b, StrongReject threshold=0.752026.05 | 36.2 | 63.8 | |
| BaselineModel=Qwen3-8B, StrongReject threshold=0.752026.05 | 37.8 | 62.2 |