AI 安全与对齐 TI-DPO（Token-Importance 引导对齐）提升训练效率；多智能体金融欺诈风险（MultiAgentFraudBench）揭示协作攻击新威胁；对齐研究从单模型扩展至多模型协同场景。 This content currently has no extended markdown.