Computer Vision Seminar Series: Dr. Xin (Eric) Wang, "Think, Act, Evolve: Building Self-Improving AI Agents."

Abstract: What would it take for an AI agent to get better at its job through experience, the way people do? This talk follows one line of work toward that goal.

An agent first needs to reason well. I start with Soft Thinking, which lets models reason in continuous concept space instead of one discrete token at a time, and GRIT, which grounds multimodal reasoning in visual evidence. Reasoning alone does not get tasks done, though. In the Agent S series, we turn reasoning models into computer-use agents that plan, remember, and operate real software. Once agents act, they collect experience, and the next question is whether that experience can make them better. Through our recent work, such as Group-Evolving Agents (GEA), I examine how agents can turn experience into cumulative progress, and what determines whether this process stays stable. This work also points to a practical limit: an agent is often held back less by its model than by how it observes and acts. Finally, I present Evolving Agent Interface, which builds reusable interfaces and supporting knowledge from task experience, so the agent finds better ways to observe and act.

Speaker Bio.: Xin (Eric) Wang is an Assistant Professor in the Computer Science Department at UC Santa Barbara and Head of Research at Simular. He directs the UCSB Center for Responsible Machine Learning and co-directs the UCSB AI Group. His research spans natural language processing, computer vision, and machine learning, advancing multimodal AI agents that perceive and reason about the world, collaborate with humans, and act on real-world tasks. His long-term vision is to build agents that are capable, generalizable, efficient, and socially responsible. Previously, he was a faculty member at UC Santa Cruz and worked at Google Research, Meta FAIR, Microsoft Research, and Adobe Research.

Eric has served as a Senior Area Chair or Area Chair for ACL, NAACL, EMNLP, ICLR, NeurIPS, and ICML. His honors include the CVPR Best Student Paper Award, Google Research Faculty Award, Amazon Alexa Prize awards, and faculty research awards from JPMorganChase, Cisco, and eBay. His work has been featured by VentureBeat, MIT Technology Review, and Hugging Face Daily Papers.