Red-teaming Strategy Diversity Analysis on Jailbreak Strategy Pool
0.282kNN DistanceTreeTeaming
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TreeTeamingOrchestrator Model=Qwen2.5-VL-7B, Number of strategies (n)=152026.03 | 0.282 | 2.723 | |
| TreeTeamingOrchestrator Model=Claude 3.5, Number of strategies (n)=152026.03 | 0.279 | 2.714 | |
| TreeTeamingOrchestrator Model=Qwen2.5-VL-7B, Number of strategies (n)=10, Note=average of 5 random runs2026.03 | 0.279 | 2.416 | |
| TreeTeamingOrchestrator Model=GPT-4o, Number of strategies (n)=152026.03 | 0.275 | 2.705 | |
| TreeTeamingOrchestrator Model=Claude 3.5, Number of strategies (n)=10, Note=average of 5 random runs2026.03 | 0.273 | 2.354 | |
| TreeTeamingOrchestrator Model=GPT-4o, Number of strategies (n)=10, Note=average of 5 random runs2026.03 | 0.268 | 2.301 | |
| Union setNumber of strategies (n)=152026.03 | 0.263 | 2.694 | |
| Union setNumber of strategies (n)=10, Note=average of 5 random runs2026.03 | 0.259 | 2.282 | |
| TreeTeamingOrchestrator Model=Qwen2.5-VL-7B, Number of strategies (n)=5, Note=average of 5 random runs2026.03 | 0.252 | 1.642 | |
| TreeTeamingOrchestrator Model=Claude 3.5, Number of strategies (n)=5, Note=average of 5 random runs2026.03 | 0.238 | 1.617 | |
| Union setNumber of strategies (n)=5, Note=average of 5 random runs2026.03 | 0.215 | 1.604 | |
| TreeTeamingOrchestrator Model=GPT-4o, Number of strategies (n)=5, Note=average of 5 random runs2026.03 | 0.205 | 1.612 |