Time-consuming but high ASR method
Maintain semantic consistency and short length to avoid Perplexity Filter
- Empty seed pool
- Mutation through LLM helper such as GPT3.5
- Role-play, Contextualization, Expand system prompts
- Jailbreaking tempaltes
- Execution and Check
- RoBERTa based fast check model
- ChatGPT based judge model
- If success → add to seed pool
