Machine Learning and Friends Lunch: Shilong Liu, A Taxonomy of Self-evolving Agents
Content
Speaker
Shilong Liu (Princeton University)
Abstract
Self-evolving agents are becoming a popular research direction, but the term is used for systems that improve in very different ways. Some agents iteratively optimize code, algorithms, scientific findings, or robot policies. Some update their own prompts, memory, tools, and skills without changing model weights. Others learn through self-training, self-play, or weak feedback from the environment.
This talk presents a taxonomy organized around three parts of an agent system: the model, the agent harness, and the artifacts it produces. These parts lead to three main research directions: artifact iterative optimization, agent harness self-improvement, and model learning without gold answers. I will explain how these directions differ, where their boundaries are beginning to blur, and how they connect to related ideas such as continual learning, test-time training, and recursive self-improvement
Speaker Bio
Shilong Liu is a Peretsman Scully Postdoctoral Research Fellow at the Princeton AI Lab, working with Prof. Mengdi Wang. He will join Columbia University’s Department of Electrical Engineering as a tenure-track Assistant Professor in Fall 2027. He received his Ph.D. in Computer Science and Technology from Tsinghua University, advised by Profs. Lei Zhang, Hang Su, and Jun Zhu, and his B.Eng. from Tsinghua University in 2020. He has also worked and interned at ByteDance Seed, NVIDIA Research, Microsoft Research, and IDEA Research.