Jailbreak Attack on CommonsenseQA
1Attack Success Rate (ASR)H-COT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| H-COTVictim LRM=Qwen2.5-Max2026.04 | 1 | 2 | |
| H-COTVictim LRM=OpenAI o4-mini2026.04 | 1 | 2 | |
| Cognitive OverloadVictim LRM=OpenAI o4-mini2026.04 | 10 | 2.6 | |
| Cognitive OverloadVictim LRM=Qwen2.5-Max2026.04 | 11 | 2.36 | |
| H-COTVictim LRM=DeepSeek R12026.04 | 60 | 2.63 | |
| Cognitive OverloadVictim LRM=DeepSeek R12026.04 | 72 | 2.58 | |
| PRJAVictim LRM=OpenAI o4-mini2026.04 | 86 | 3.3 | |
| PRJAVictim LRM=DeepSeek R12026.04 | 91 | 3.63 | |
| PRJAVictim LRM=Qwen2.5-Max2026.04 | 95 | 3.42 |