讲座题目：Global Coherence: When Every Agent Is Right and the Team Is Still Wrong (全局一致性：当每个智能体都正确，团队却依然出错)
主 讲 人：Xin Heng（Tote AI 首席研究科学家兼人工智能与数据副总裁）
讲座时间：2026年11月13日（周五）上午10:00
讲座地点：钱伟长楼201会议室
欢迎感兴趣的师生前来聆听、交流！
光电信息与物理科学学院
2026年11月10日
讲座内容简介：
With the rapid advancement of Large Language Models (LLMs) and AI agents, multi-agent systems have become a promising paradigm for complex task solving. However, even when individual agents make locally valid decisions, their collaboration may still result in a loss of global coherence. This talk will introduce the global coherence problem in multi-agent systems, emphasizing that such failures arise from limitations in shared state and information availability rather than merely insufficient model intelligence.
The talk will first present the Observation-Aliasing Impossibility Theorem, which characterizes the fundamental limits of decision-making under incomplete observations. It will then introduce a Local-to-Global Semantic Framework, formulated as X = (H, C, G, F; D), which characterizes interaction topology, semantic transformations, state transitions, consistency conditions, and essential historical information in multi-agent systems. Furthermore, the talk will discuss the Coherence Harness architecture, where models propose actions while the harness supplies authoritative state, checks proposals against system constraints, and controls whether actions are accepted or rejected.
主讲人简介：
Xin Heng，博士，人工智能与大数据分析领域技术专家，现任 Tote AI 首席研究科学家兼人工智能与数据副总裁。毕业于加州理工学院，获电气工程博士学位，并获加州大学伯克利分校哈斯商学院金融工程硕士学位。研究方向包括人工智能、大数据分析、多智能体系统及企业级人工智能产品与基础设施建设。曾任 Caffeine TV 人工智能与数据领域执行顾问、Punchh 副总裁兼数据负责人、Uber 数据科学部门经理等职位。
