Adversarial Attack on Harmful Query Stress Test
5.4ASRGCG
Evaluation Results
| Method | Links | |
|---|---|---|
| GCGTTS Strategy=Best-of-N, Model=Gemma3-27B, N=82025.10 | 5.4 | |
| AutoDAN-TurboTTS Strategy=MCTS, Model=Gemma3-27B2025.10 | 15.6 | |
| AutoDAN-TurboTTS Strategy=Best-of-N, Model=Gemma3-27B, N=82025.10 | 17.1 | |
| GCGTTS Strategy=Best-of-N, Model=Llama3.1-8B, N=82025.10 | 17.6 | |
| GCGTTS Strategy=MCTS, Model=Llama3.1-8B2025.10 | 25.4 | |
| GCGTTS Strategy=Best-of-N, Model=Qwen3-8B, N=82025.10 | 33.5 | |
| GCGTTS Strategy=MCTS, Model=Gemma3-27B2025.10 | 33.6 | |
| AutoDANTTS Strategy=Best-of-N, Model=Llama3.1-8B, N=82025.10 | 36.8 | |
| AutoDAN-TurboTTS Strategy=Best-of-N, Model=Llama3.1-8B, N=82025.10 | 39.7 | |
| GCGTTS Strategy=MCTS, Model=Qwen3-8B2025.10 | 40 | |
| AutoDAN-TurboTTS Strategy=Best-of-N, Model=Qwen3-8B, N=82025.10 | 41.4 | |
| AutoDAN-TurboTTS Strategy=MCTS, Model=Llama3.1-8B2025.10 | 44.6 | |
| REFDIVTTS Strategy=Best-of-N, Model=Llama3.1-8B, N=82025.10 | 46.5 | |
| AutoDAN-TurboTTS Strategy=MCTS, Model=Qwen3-8B2025.10 | 48.5 | |
| AutoDAN-TurboTTS Strategy=Best-of-N, Model=Mistral-7B, N=82025.10 | 73.3 | |
| AutoDANTTS Strategy=Best-of-N, Model=Gemma3-27B, N=82025.10 | 74.9 | |
| AutoDAN-TurboTTS Strategy=MCTS, Model=Mistral-7B2025.10 | 76.4 | |
| AutoDANTTS Strategy=MCTS, Model=Llama3.1-8B2025.10 | 83.1 | |
| GCGTTS Strategy=Best-of-N, Model=Mistral-7B, N=82025.10 | 87.7 | |
| AutoDANTTS Strategy=MCTS, Model=Gemma3-27B2025.10 | 90.4 | |
| REFDIVTTS Strategy=Best-of-N, Model=Gemma3-27B, N=82025.10 | 92.6 | |
| REFDIVTTS Strategy=MCTS, Model=Llama3.1-8B2025.10 | 96.7 | |
| AutoDANTTS Strategy=Best-of-N, Model=Mistral-7B, N=82025.10 | 97.3 | |
| REFDIVTTS Strategy=Best-of-N, Model=Mistral-7B, N=82025.10 | 97.6 | |
| REFDIVTTS Strategy=MCTS, Model=Gemma3-27B2025.10 | 98.9 | |
| REFDIVTTS Strategy=Best-of-N, Model=Qwen3-8B, N=82025.10 | 99.5 | |
| AutoDANTTS Strategy=Best-of-N, Model=Qwen3-8B, N=82025.10 | 99.6 | |
| GCGTTS Strategy=MCTS, Model=Mistral-7B2025.10 | 99.6 | |
| AutoDANTTS Strategy=MCTS, Model=Qwen3-8B2025.10 | 100 | |
| REFDIVTTS Strategy=MCTS, Model=Qwen3-8B2025.10 | 100 | |
| AutoDANTTS Strategy=MCTS, Model=Mistral-7B2025.10 | 100 | |
| REFDIVTTS Strategy=MCTS, Model=Mistral-7B2025.10 | 100 |