1
RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates
多智能体辩论机制让AI自相检验安全漏洞，红队自动生成对抗性攻击，大幅提升大模型回复安全性。
arXiv:2506.11083v3 Announce Type: replace Abstract: We introduce RedDebate, a novel multi-agent debate framework that provides the foundation for Larg…