加州大学默塞德分校助理教授王艺炜博士学术讲座通知
报告题目：模块化大语言模型在多智能体系统中的潜力与局限性
Potential and Limitations of Modular LLMs in Multi-Agent Systems
报 告 人：王艺炜，加州大学默塞德分校计算机科学系的助理教授
时间：2025年5月15日，周四上午9:30-12:00
地点：西五楼 404
报告人简介：
王艺炜博士现为加州大学默塞德分校计算机科学系的助理教授。2023 年，他在亚马逊西雅图担任应用科学家;2024 年于加州大学洛杉矶分校自然语言处理组从事博士后研究。他于 2023 年获得新加坡国立大学博士学位，研究方向为自然语言处理与多模态大语言模型。近期的研究重点包括大语言模型的可控生成以及其在医疗健康领域的应用。他曾获得 2021年 SDSC 研究奖学金。更多信息请见个人主页。https://wangywust.github.io/.
Yiwei Wang is an Assistant Professor in the Computer Science Department of University of California at Merced. He was an Applied Scientist in Amazon (Seattle) in 2023 and a Postdoc in UCLA NLP Group in 2024. He obtained his Ph.D. degree from National University of Singapore in 2023. His research lies in natural language processing and multi-modal large language models. His recent research focuses on the controllable generation of LLMs and applications of LLMs in healthcare. He was awarded the 2021 SDSC Research Fellowship. Additional information is available at https://wangywust.github.io/.
报告摘要：
大语言模型(LLMs)在自然语言生成方面表现出令人瞩目的能力，但其在现实世界应用中的实用性仍受多项关键挑战限制，例如生成不可控、缺乏上下文忠实性以及安全边界对齐不足等问题。在本次报告中，我们探讨将块化大语言模型集成至多智能体系统中的潜力与局限性，其中模块化和任务分解被视为相较于训练时大规模扩展的测试时替代方案。我们提出了一种“分而治之”的框架，能够通过协调多个专门化智能体，实现可控生成、可扩展推理与知识编辑。我们的研究成果包括用于满足约束的文本生成新算法、多智能体协作生成图表的方法，以及基于解码约束的知识编辑模块化流程。我们还指出了诸如输出前缀越狱等新兴安全漏洞，并提出了对齐人类价值的评估框架。这些工作共同描绘了构建可信，可扩展的测试时AI 系统的路径，能够应用于医疗健康、软件开发等多个领域。
Large Language Models (LLMs) have shown impressive capabilities in natural language generation, but their utility in real-world applications remains limited by key challenges such as uncontrollability, lack of context-faithfulness, and misaligned safety boundaries. In this talk, we explore the potential and limitations of modular LLMs when integrated into multi-agent systems, where modularization and task decomposition serve as a test-time alternative to massive training-time scaling. We present a divide-and-conquer framework that enables controllable generation, scalable reasoning, and knowledge editing by orchestrating multiple specialized agents. Our research includes novel algorithms for constraint-satisfying text generation, multi-agent collaboration for chart production, and modular pipelines for knowledge editing using decoding constraints. We also highlight emerging safety vulnerabilities, such as output prefix jailbreaks, and propose human-aligned evaluation frameworks. Together, these contributions outline a pathway for building trustworthy and test-time scalable AI systems that can support healthcare, software development, and beyond.
