Jailbreaking on HarmBench V2 (val)
72Attack Success RateAttention Redistribution Attack (ARA)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Attention Redistribution Attack (ARA)Model=Mistral-7B, k=5, steps=500, prompts=2002026.04 | 72 | 29.5 | 76.4 | 86.6 | |
| Attention Redistribution Attack (ARA)Model=LLaMA-3-8B, k=5, steps=500, prompts=2002026.04 | 60 | 23.8 | 59.6 | 81.4 | |
| Attention Redistribution Attack (ARA)Model=Gemma-2-9B, k=5, steps=500, prompts=2002026.04 | 2 | 0.1 | 45.8 | 80.1 |