Efficient large language model safety evaluation via multi-agent debate
Efficient LLM Safety Evaluation through Multi-Agent Debate
AI summary
The researchers propose the Multi-Agent Judge framework, in which three agents—a critic, a defender, and a judge—debate according to a unified safety scoring standard, improving the reliability of large language model safety evaluation at low cost. The accompanying HAJailBench is a manually annotated jailbreak benchmark containing 11,100 annotated interactions, covering multiple attack methods and target models. On this benchmark, the framework outperforms matched small-model prompting baselines and existing multi-agent judge methods, and is more economical than GPT-4o according to the pricing snapshot evaluated.