Huazheng Wang is an Assistant Professor in the School of Electrical Engineering and Computer Science at Oregon State University. He was a postdoctoral research associate at Princeton University. He received his PhD in Computer Science from University of Virginia, and his B.E. from University of Science and Technology of China. His research is in reinforcement learning, information retrieval, and multi-agent systems. He has recently focused on developing provably efficient and trustworthy reinforcement learning algorithms for recommendation, ranking, and Large Language Model (LLM) agents. He served as Area Chair at ICML, NeurIPS, ICLR, Senior PC at SIGIR, and Action Editor at TMLR. He is the recipient of AAAI 2025 New Faculty Highlight and SIGIR 2019 Best Paper Award.

Large language models are increasingly being used as agentic systems, where multiple LLM agents collaborate through roles, workflows, and environments. While such systems can solve complex tasks beyond single-agent inference, they also introduce new challenges: failures become harder to diagnose, credit is difficult to assign across agents, and fixed workflows may limit further improvement. This talk asks two questions: how should we evaluate multi-agent LLM systems, and how can we train them to improve? In the first part, I will introduce automated failure attribution as a new research direction: when an agentic system fails, can we identify which agent caused the failure and when the decisive error occurred? I will present our work on Who&When and Who&When Pro, which benchmark and scale automated failure attribution for LLM agent systems across diverse tasks and modalities. In the second part, I will turn from evaluation to training. I will first show what we learn from using reinforcement learning to post-train multi-agent LLM workflows: performance can improve, but the gains and stability depend strongly on workflow topology, model scale, and policy sharing. Building on this motivation, I will focus on MetaAgent-X, an end-to-end reinforcement learning framework for automatic multi-agent systems that moves beyond fixed workflows by jointly optimizing both the system designer and executor agents. Together, these works point toward a new generation of trainable and diagnosable agentic systems that can learn from failures and improve through interaction.