Adaptive Jailbreak Attack Success Rate on PAIR behaviors held-out (test)
100ASRQwen3-8B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3-8BDefense=Baseline, Sampling Strategy=best-of-8x52026.05 | 100 | |
| Qwen3-4B-Instruct-2507Defense=Baseline, Sampling Strategy=best-of-8x52026.05 | 95 | |
| gpt-oss-20bDefense=Baseline, Sampling Strategy=best-of-8x52026.05 | 94 | |
| Qwen3-4B-Instruct-2507Defense=BCT, Sampling Strategy=best-of-8x52026.05 | 16.9 | |
| Qwen3-8BDefense=BCT, Sampling Strategy=best-of-8x52026.05 | 12 | |
| Qwen3-8BDefense=OPCT, Sampling Strategy=best-of-8x52026.05 | 1.2 | |
| Qwen3-4B-Instruct-2507Defense=OPCT, Sampling Strategy=best-of-8x52026.05 | 0 | |
| gpt-oss-20bDefense=BCT, Sampling Strategy=best-of-8x52026.05 | 0 | |
| gpt-oss-20bDefense=OPCT, Sampling Strategy=best-of-8x52026.05 | 0 |