Zhuoyun Du 杜卓耘

My research began at THUNLP, where I worked with Prof. Zhiyuan Liu and Prof. Chen Qian on how agents organize and coordinate their work. Studying multi-agent collaboration led me to a broader question: how can agents learn to improve the ways they work together?

At Zhejiang University, I explored agent self-improvement through doctor–patient coevolution. At Alibaba, I investigated communication in latent space, asking how agents could exchange information beyond natural-language messages. These experiences shaped my interest in agents that develop their own representations, communication protocols, and mechanisms for self-improvement.

More recently, in the ByteDance Talent Program, I worked on multi-turn dialogue systems, user simulation, autonomous evaluation, and LLM post-training. This experience connected my research interests to the practical challenge of generating useful interaction data and evaluating whether agents are improving. Together, these directions motivate my long-term goal of building machine-native AI agents that learn and improve through interaction.