LLM Jailbreaking on 100-query jailbreak set
46.4Jailbreak Success RateReward-Guided RRT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Reward-Guided RRTModel=DeepSeek-V32025.07 | 46.4 | 2.7 | 4.2 | 17.7 | |
| Reward-Guided RRTModel=Gemini-2.5-Flash2025.07 | 36 | 8.4 | 2.8 | 23.4 | |
| Baseline RRTModel=DeepSeek-V32025.07 | 34.8 | 6.8 | 3.8 | 7.2 | |
| Reward-Guided RRTModel=Llama-3.1-70B2025.07 | 33.8 | 13.4 | 3.4 | 27.2 | |
| Reward-Guided RRTModel=Qwen-Plus2025.07 | 31 | 7.6 | 2.4 | 18 | |
| Baseline RRTModel=Qwen-Plus2025.07 | 29.4 | 5.6 | 9.2 | 7.4 | |
| Baseline RRTModel=Llama-3.1-70B2025.07 | 27.2 | 14.4 | 7.4 | 19.4 | |
| Baseline RRTModel=Gemini-2.5-Flash2025.07 | 26.2 | 6.4 | 5 | 14.2 |