CS Colloquium: Hua Wei on Toward Mindful Agents: Knowing When to Act, Adapt, and Defer

Please join us in ECCR 265 or on Zoom!

Abstract: AI agents are moving from answering questions to taking actions: controlling robots and infrastructure, reasoning through multi-step problems, and working alongside people. Yet agents are trained in one setting and deployed in another, and they rarely know when they are out of their depth. This talk argues that reliable agents need self-knowledge: the ability to recognize when they might be wrong and to respond by acting, adapting, or deferring. I will first examine the gap between where agents are trained and where they are deployed, in both physical control and language agents, and how to narrow it. I will then show how agents can locate what they do not know, in their inputs and in their own reasoning. Finally, I will discuss when agents should hand decisions back to people, from safety oversight to education. I will close with open problems toward agents whose self-knowledge is built into how they are trained, aligned, and deployed.

Bio: Hua Wei is an Assistant Professor in the School of Computing and Augmented Intelligence at Arizona State University. He builds mindful AI agents that act in the real world, quantify and adapt to their own uncertainty, and support people who make consequential decisions, drawing on reinforcement learning, uncertainty quantification, and multi-agent systems. He received the NSF CAREER Award 2025, the IJCAI 2026 Early Career Spotlight, the ECML-PKDD 2020 Best Applied Data Science Paper Award, and Amazon, Cisco, and Toyota faculty research awards. He received his Ph.D. from Penn State.