Topic: Leveraging Collective Advice-Taking Behavior to Infer Accuracy and Improve the Wisdom of Crowds

Speaker: Assistant Professor Yunhao (Jerry) Zhang, UCL School of Management

Moderator: Associate Professor Yunlu Yin

Time: 13:30-15:00, Tuesday, September 15, 2026

Venue: Guoshun Campus, Room 624, Siyuan Faculty Building

Abstract: Wisdom-of-crowds estimates can be compromised when some agents’ predictions are systematically biased. A natural remedy is to selectively aggregate the predictions of more accurate agents, but identifying these agents is difficult when ground truth and past performance are unavailable. I show that collective advice-taking behavior can reveal which members of a crowd are better informed and thereby improve aggregate accuracy. The core insight is that, after observing numerical advice (e.g., the group mean), better-informed agents - having already incorporated higher-quality information into their initial estimates - perceive less corrective value or novel information from the advice and therefore place less weight on it than less-informed agents. Building on this insight, I develop Cluster-Weight-on-Advice (CWOA), a two-shot aggregation method that identifies clusters of similar initial estimates and selectively aggregates the cluster that updates least in response to advice. A behavioral model characterizes when lower cluster-level weight on advice (WOA) signals greater accuracy and shows that this ordering can persist despite imperfect beliefs and psychological biases in advice-taking. Two preregistered experiments manipulating information quality provide direct evidence for the proposed mechanism: better-informed, more accurate participants place less weight on advice, even when reporting no greater confidence. Across two additional preregistered studies and five archival datasets spanning medical judgments, consumer estimates, political factual beliefs, and general knowledge, CWOA consistently improves aggregate accuracy and outperforms conventional and state-of-the-art meta-prediction-based benchmarks. In one study, where advice comes from an AI model that is more accurate than the human group mean, CWOA nevertheless produces estimates more accurate than the AI itself. These findings show that how people revise their beliefs can reveal latent information quality even when accuracy cannot be directly observed, providing a new behavioral foundation for improving the wisdom of crowds.

Bio: Yunhao (Jerry) Zhang is an Assistant Professor in the Marketing group at UCL School of Management. He received his PhD in Management from the MIT Sloan School of Management and was a postdoctoral research fellow at the Psychology of Technology Institute, jointly sponsored by the Haas School of Business at the University of California, Berkeley, and the Marshall School of Business at the University of Southern California. His research examines information aggregation, human-AI interaction, collective intelligence, belief updating, and judgment and decision-making. His work has been published in journals including Management Science, Cognition, Nature Human Behaviour, and Proceedings of the National Academy of Sciences.