Multi-turn Jailbreak Defense on MTJ-Pro
82.6Purpose Reverse Success RatePAT
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| PATTarget LLM=GPT-42026.01 | 82.6 | 25 | 42.9 | 15.8 | 38.9 | 10 | 31.6 | 30.7 | |
| RPOTarget LLM=LLaMa2026.01 | 82.6 | 0 | 42.9 | 10.5 | 22.2 | 0 | 36.8 | 22.2 | |
| RPOTarget LLM=Gemini2026.01 | 78.3 | 25 | 57.1 | 26.3 | 33.3 | 40 | 42.1 | 37.2 | |
| RPOTarget LLM=GPT-3.52026.01 | 60.9 | 0 | 42.9 | 26.3 | 44.4 | 60 | 15.8 | 34.3 | |
| RPOTarget LLM=GPT-42026.01 | 56.5 | 50 | 57.1 | 36.8 | 55.6 | 40 | 36.8 | 39.4 | |
| PATTarget LLM=Gemini2026.01 | 47.8 | 25 | 14.3 | 15.8 | 16.7 | 20 | 21.1 | 26.4 | |
| PATTarget LLM=GPT-3.52026.01 | 43.5 | 25 | 42.9 | 21.1 | 33.3 | 30 | 26.3 | 31.7 | |
| Self-ReminderTarget LLM=Gemini2026.01 | 43.5 | 25 | 42.9 | 26.3 | 22.2 | 40 | 31.6 | 30.7 | |
| GoalPriorityTarget LLM=LLaMa2026.01 | 39.1 | 25 | 42.9 | 10.5 | 38.9 | 30 | 31.6 | 31.4 | |
| PATTarget LLM=LLaMa2026.01 | 30.4 | 25 | 28.6 | 5.3 | 22.2 | 40 | 21.1 | 26.1 | |
| Self-ReminderTarget LLM=LLaMa2026.01 | 30.4 | 50 | 14.3 | 21.1 | 27.8 | 50 | 36.8 | 35.3 | |
| GoalPriorityTarget LLM=Gemini2026.01 | 26.1 | 50 | 14.3 | 5.3 | 38.9 | 30 | 31.6 | 30.1 | |
| Self-ReminderTarget LLM=GPT-3.52026.01 | 21.7 | 25 | 0 | 10.5 | 5.6 | 30 | 15.8 | 18.4 | |
| HoneyTrapTarget LLM=LLaMa2026.01 | 21.7 | 0 | 0 | 0 | 16.7 | 30 | 5.3 | 14 | |
| GoalPriorityTarget LLM=GPT-42026.01 | 17.4 | 0 | 0 | 5.3 | 0 | 10 | 10.5 | 11.8 | |
| Self-ReminderTarget LLM=GPT-42026.01 | 13 | 0 | 14.3 | 10.5 | 0 | 20 | 10.5 | 16.3 | |
| HoneyTrapTarget LLM=GPT-42026.01 | 13 | 0 | 14.3 | 0 | 5.6 | 0 | 5.3 | 5.7 | |
| HoneyTrapTarget LLM=Gemini2026.01 | 13 | 25 | 14.3 | 0 | 11.1 | 0 | 15.8 | 9.4 | |
| GoalPriorityTarget LLM=GPT-3.52026.01 | 8.7 | 0 | 0 | 15.8 | 11.1 | 10 | 5.3 | 11.4 | |
| HoneyTrapTarget LLM=GPT-3.52026.01 | 8.7 | 25 | 0 | 5.3 | 0 | 0 | 10.5 | 6.4 |