Adversarial Attack on MMLU 16 subjects filtered subset
81.24ASR@30SECA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SECATarget Model=Llama-3-8B2025.10 | 81.24 | 0.33 | 0 | |
| SECATarget Model=Llama-3-3B2025.10 | 80.29 | 0.6 | 0 | |
| RawTarget Model=Llama-3-8B2025.10 | 63.52 | 1.08 | 0 | |
| RawTarget Model=Llama-3-3B2025.10 | 48.2 | 1.08 | 0 | |
| SECATarget Model=Qwen-2.5-7B2025.10 | 36.86 | 1.06 | 0 | |
| RawTarget Model=Qwen-2.5-7B2025.10 | 10.19 | 1.08 | 0 | |
| GCGTarget Model=Llama-3-8B2025.10 | 9.86 | 307.68 | 0.98 | |
| GCGTarget Model=Llama-3-3B2025.10 | 6.26 | 1,255.04 | 0.97 | |
| GCGTarget Model=Qwen-2.5-7B2025.10 | 0.57 | 1,036.62 | 0.96 |