Dissertation Defense: Yutong Xie

The School of Information is pleased to announce the oral defense of Yutong Xie.

Title: Towards AI Behavioral Science: Measurement, Modeling, and Objectives

Date: Tuesday, September 1, 2026

Time: 1:00-3:00pm

Location: Dow Room (LCSIB 4320)

Qiaozhu Mei, serving as committee chair, will preside over the oral defense. 

All are welcome to attend!

Abstract:

Modern artificial intelligence (AI) systems, particularly large language models (LLMs), increasingly operate in settings with consequential social, economic, and institutional stakes. In these environments, they exhibit emergent behavioral tendencies: patterns of cooperation, fairness, and risk-taking that arise from training rather than explicit design. Characterizing these tendencies is essential for understanding how AI systems behave and how they will ultimately affect the people and societies with which they interact. At the same time, the human-likeness of AI behavior makes these systems novel instruments for studying human behavior. At this intersection, AI Behavioral Science emerges as a rapidly developing interdisciplinary field. This dissertation advances the field along three connected directions: measuring the behavior of AI systems, modeling human behavior with AI, and understanding the objectives that shape AI behavior.

The first direction adapts instruments from behavioral economics and psychology to characterize the behavioral tendencies of LLMs and compare them against large human populations. The findings indicate that while AI behavior generally falls within the range of human behaviors, its distribution differs systematically. AI choices are more concentrated, and leading models are often more cooperative and altruistic than the average human. 

The second direction turns AI into an instrument for modeling human behavior. Using LLMs to interpret human behavioral motivations as natural-language "behavioral codes," this work elicits and reproduces human behavioral patterns. These codes can be further composed into mixtures that reflect the distributional diversity of real populations, rather than merely capturing their average behavior. 

The final direction investigates the origins of AI behavior, tracing them to the objectives used to train and evaluate models. This research constructs an "objective space" in which diverse objectives, ranging from payoffs in economic games to standard benchmark performances and human-preference ratings, are jointly embedded into a shared space based on how similarly they rank models.

Together, these three connected fronts demonstrate practical approaches that advance AI Behavioral Science as a rigorous interdisciplinary domain. This dissertation lays the foundation for a broad research agenda where AI serves simultaneously as an object of behavioral study, an instrument for behavioral modeling, and a system whose behavior can be shaped through the objectives used to develop it.