Behaviorally Informed Institutional Design with Game Theory and LLMs

The event will start 30 minutes before initially announced.

How can we design social institutions that account for how people actually make decisions? Behavioral research has revealed substantial heterogeneity in preferences and decision-making that traditional economic models often abstract away. Understanding these tendencies may open new possibilities for improving collective outcomes. In this talk, I present two projects addressing complementary aspects of this broader goal. First, I investigate how real-time interaction facilitates cooperation in a one-shot social dilemma, where purely self-interested behavior predicts only mutual defection. Experiments show that real-time interaction yields higher mutual cooperation, primarily through stronger reciprocity by followers. A game-theoretic model suggests that flexible decision timing helps people infer others' cooperative preferences from their actions and coordinate their responses. This illustrates how institutional tweaks can leverage existing cooperative tendencies. Second, I introduce Cross-Centaur, an LLM-based model of individual decision-making across diverse domains. We construct a dataset in which the same participants complete tasks spanning learning, decisions under risk and delay, reasoning, and social decision-making, and use it to fine-tune an LLM. The model uses behavior in one task to better predict the same individual's decisions in another, a step toward more flexible, generalizable models of decision-making. Together, these projects explore how understanding human decision-making may inform the design of better social institutions.