Adversarial Attack on AdvBench 80/20 (test)
82.7ASR@1ATOXIA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ATOXIATarget LLM=VICUNA-7B2025.10 | 82.7 | 92.3 | 4.53 | |
| ATOXIATarget LLM=MISTRAL-7B2025.10 | 73.1 | 99.2 | 54.42 | |
| RED-BANDIT (ϵ-GREEDY)Target LLM=VICUNA-7B2025.10 | 71.2 | 100 | 1.85 | |
| RED-BANDIT (UCB)Target LLM=MISTRAL-7B2025.10 | 70.2 | 100 | 2.31 | |
| RED-BANDIT (UCB)Target LLM=VICUNA-7B2025.10 | 63.5 | 100 | 2.37 | |
| AUTODAN-UNIVERSALTarget LLM=VICUNA-7B2025.10 | 63.2 | 84.9 | 76.33 | |
| RED-BANDIT (ϵ-GREEDY)Target LLM=MISTRAL-7B2025.10 | 61.5 | 100 | 3.49 | |
| ADVPROMPTER-WARMSTARTTarget LLM=MISTRAL-7B2025.10 | 58.7 | 95.9 | 41.16 | |
| ADVPROMPTERTarget LLM=MISTRAL-7B2025.10 | 54.3 | 96.1 | 41.6 | |
| AUTODAN-UNIVERSALTarget LLM=MISTRAL-7B2025.10 | 51.9 | 86.5 | 57.41 | |
| GCG-UNIVERSALTarget LLM=MISTRAL-7B2025.10 | 46.2 | 99 | — | |
| RED-BANDIT (ϵ-GREEDY)Target LLM=LLAMA2-7B2025.10 | 39.4 | 96.2 | 3.5 | |
| GCG-UNIVERSALTarget LLM=VICUNA-7B2025.10 | 36.7 | 82.7 | — | |
| ADVPROMPTER-WARMSTARTTarget LLM=VICUNA-7B2025.10 | 35.6 | 85.6 | 13.02 | |
| RED-BANDIT (UCB)Target LLM=LLAMA2-7B2025.10 | 33.7 | 99 | 2.4 | |
| ADVPROMPTERTarget LLM=VICUNA-7B2025.10 | 33.4 | 87.5 | 12.09 | |
| ATOXIATarget LLM=LLAMA2-7B2025.10 | 18.3 | 41.4 | 5.8 | |
| ADVPROMPTER-WARMSTARTTarget LLM=LLAMA2-7B2025.10 | 12.5 | 46.1 | 158.5 | |
| GCG-UNIVERSALTarget LLM=LLAMA2-7B2025.10 | 1 | 2.1 | — | |
| AUTODAN-UNIVERSALTarget LLM=LLAMA2-7B2025.10 | 1 | 2.1 | 373.72 | |
| ADVPROMPTERTarget LLM=LLAMA2-7B2025.10 | 1 | 7.7 | 86.8 |