Quality-Diversity Red-Teaming: Automated Generation of High-Quality and Diverse Attackers for Large Language Models
The Quality-Diversity Red-Teaming (QDRT) framework automatically generates high-quality and diverse adversarial prompts to test Large Language Models. It achieves this by training multiple specialized attackers using behavior-conditioned reinforcement learning. This allows the system to find attacks with varied styles and risk categories, significantly improving attack diversity and effectiveness against various LLMs.