Adversarial Attack on AdvBench
100Attack Success Rate (ASR)Standard
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| StandardTarget Model=Llama-3.1-8B, Initialization=Standard2025.02 | 100 | 21 | 22.02 | 3.16 | — | — | |
| BaselineTarget Model=Llama-3.1-8B, Initialization=Baseline2025.02 | 100 | 3 | 3.96 | 1.58 | — | — | |
| 1-CRITarget Model=Llama-3.1-8B, Initialization=1-CRI (ours)2025.02 | 100 | 7.5 | 7.6 | 6.59 | — | — | |
| 5-CRITarget Model=Llama-3.1-8B, Initialization=5-CRI (ours)2025.02 | 100 | 2 | 2.52 | 0.69 | — | — | |
| 25-CRITarget Model=Llama-3.1-8B, Initialization=25-CRI (ours)2025.02 | 100 | 7 | 7.84 | 4.06 | — | — | |
| BaselineTarget Model=Mistral-7B, Initialization=Baseline2025.02 | 100 | 4 | 8.28 | 2.2 | — | — | |
| 1-CRITarget Model=Mistral-7B, Initialization=1-CRI (ours)2025.02 | 100 | 5 | 5.5 | 4.02 | — | — | |
| 5-CRITarget Model=Mistral-7B, Initialization=5-CRI (ours)2025.02 | 100 | 2 | 2.3 | 0.41 | — | — | |
| 25-CRITarget Model=Mistral-7B, Initialization=25-CRI (ours)2025.02 | 100 | 7 | 7.02 | 2.89 | — | — | |
| StandardTarget Model=Mistral-7B, Initialization=Standard2025.02 | 92 | 14 | 27.26 | 2.71 | — | — | |
| BaselineJudge=Heuristic2026.06 | 65.18 | — | — | — | — | — | |
| SafeLMJudge=Heuristic2026.06 | 55.56 | — | — | — | — | — | |
| BaselineJudge=LLM2026.06 | 43.99 | — | — | — | — | — | |
| SafeLMJudge=LLM2026.06 | 23.55 | — | — | — | — | — | |
| SRPJudge=Heuristic2026.06 | 1.98 | — | — | — | — | — | |
| SRPJudge=LLM2026.06 | 0.22 | — | — | — | — | — | |
| Mistral-7B-Instruct-v0.3Attack=PAIR2026.06 | — | — | — | — | 15.77 | 13.85 | |
| Mistral-7B-Instruct-v0.3Attack=Prefilling2026.06 | — | — | — | — | 52.81 | 49.06 | |
| Mistral-7B-Instruct-v0.3Attack=I-GCG2026.06 | — | — | — | — | 20.96 | 15.19 | |
| Mistral-7B-Instruct-v0.3 + Trajectory AlignmentAttack=PAIR2026.06 | — | — | — | — | 0.97 | 0.19 | |
| Mistral-7B-Instruct-v0.3 + Trajectory AlignmentAttack=Prefilling2026.06 | — | — | — | — | 77.69 | 2.12 | |
| Mistral-7B-Instruct-v0.3 + Trajectory AlignmentAttack=I-GCG2026.06 | — | — | — | — | 0.38 | 0.38 |