Harmful Refusal on StrongREJECT Adversarial
37.1Attack Success Rate (ASR)SELF-REDTEAM
Evaluation Results
| Method | Links | |
|---|---|---|
| SELF-REDTEAMBackbone=Qwen2.5-7B-IT2026.05 | 37.1 | |
| Qwen2.5-7B-ITBackbone=Qwen2.5-7B-IT2026.05 | 47 | |
| Qwen2.5-3B-ITBackbone=Qwen2.5-3B-IT2026.05 | 49.7 | |
| SELF-REDTEAMBackbone=Qwen2.5-3B-IT2026.05 | 55.9 | |
| ABSBackbone=Qwen2.5-7B-IT2026.05 | 64.4 | |
| Qwen2.5-14B-ITBackbone=Qwen2.5-14B-IT2026.05 | 64.7 | |
| ABSBackbone=Qwen2.5-3B-IT2026.05 | 66.9 | |
| SELF-REDTEAMBackbone=Qwen2.5-14B-IT2026.05 | 70.1 | |
| ABSBackbone=Qwen2.5-14B-IT2026.05 | 74.8 |