Jailbreak Attack Defense on SorryBench
56.8FFR (Reasoning)Paraphrase
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ParaphraseModel=R1-Qwen-14B2025.08 | 56.8 | 77.3 | |
| SmoothLLMModel=R1-Qwen-14B2025.08 | 54.5 | 70.5 | |
| ParaphraseModel=R1-Qwen-32B2025.08 | 52.3 | 79.5 | |
| No DefenseModel=R1-Qwen-14B2025.08 | 50 | 72.7 | |
| SmoothLLMModel=R1-Qwen-32B2025.08 | 50 | 79.5 | |
| No DefenseModel=R1-Qwen-32B2025.08 | 43.2 | 77.3 | |
| Self-ReminderModel=R1-Qwen-14B2025.08 | 25 | 38.6 | |
| SAFEPATH-ZSModel=R1-Qwen-32B2025.08 | 22.7 | 40.9 | |
| SAFEPATH-ZSModel=R1-Qwen-14B2025.08 | 20.5 | 47.7 | |
| Self-ReminderModel=R1-Qwen-32B2025.08 | 20.5 | 40.9 | |
| ThinkingIModel=R1-Qwen-14B2025.08 | 15.9 | 34.1 | |
| ReasoningGuardModel=R1-Qwen-14B2025.08 | 13.6 | 22.7 | |
| RealSafe-R1Model=R1-Qwen-14B2025.08 | 9.1 | 13.6 | |
| ReasoningGuardModel=R1-Qwen-32B2025.08 | 9.1 | 20.5 | |
| ThinkingIModel=R1-Qwen-32B2025.08 | 6.8 | 47.7 | |
| SafeKeyModel=R1-Qwen-14B2025.08 | 4.5 | 20.5 | |
| RealSafe-R1Model=R1-Qwen-32B2025.08 | 2.3 | 9.1 |