Zhou Yu is the co-founder and chief executive of Arklex AI and an associate professor of computer science at Columbia University. She builds conversational systems and autonomous agents that can reason through unfamiliar situations, recover from mistakes, and recognize when people should retain control.
Yu studied computer science and linguistics at Zhejiang University before earning her doctorate from Carnegie Mellon University’s Language Technologies Institute in 2017. She worked on interactive systems at Microsoft Research and Educational Testing Service during graduate school, then joined the University of California, Davis, as an assistant professor. There, she led the team behind Gunrock, the social chatbot that won the 2018 Amazon Alexa Prize. Team member Chun-Yen “Arbit” Chen later became her Arklex co-founder and chief technology officer.
At Columbia, Yu established a conversational AI course in 2021 and co-directs DAPLab, where her interests span dialogue, language-model reasoning, human-computer interaction, computer systems, and security. Her research follows a practical progression from sustaining open-ended conversations to planning consequential actions and evaluating whether agents remain dependable outside controlled demonstrations.
- Strategic conversational planning. Her co-authored GDP-Zero research uses open-loop Monte Carlo tree search to simulate possible user responses and compare conversational strategies. In charitable-persuasion experiments, agents learned to vary their tactics and avoid soliciting donations prematurely.
- On-policy self-improvement. With TriPosT, Yu and her collaborators train smaller language models on their own attempted solutions, feedback, and revisions. Stronger models refine that feedback without substituting unfamiliar reasoning for the smaller model’s actual mistakes; improvement remains constrained by the quality of the verifier.
- Reflective search and exploratory learning. ExACT extends tree search to browser and computer-use agents through reusable memory, contrastive reflection, and debate among evaluators. Its exploratory learning method teaches agents to investigate alternatives, assess progress, and backtrack instead of learning exclusively from successful final actions.
- Simulation-based agent evaluation. Arklex’s open-source ArkSim tests agents through realistic, multi-turn simulated conversations, while its agent-orchestration framework combines task graphs, specialized workers, integrations, and human oversight.
Yu calibrates autonomy to consequences: supervised back-office tasks require different safeguards from customer-facing decisions or agents operating across sensitive systems. Her AI Engineer Summit presentation extends that concern to environments where multiple agents and users share resources, creating additional challenges around scheduling, security, efficiency, and timely human handoffs.
