Dissertation Committee Chair/Advisor:
Dr. Robert C. Wilson – (Georgia Institute of Technology)

Dissertation Committee Members:
Dr. Thomas L. Griffiths – (Princeton University)
Dr. Anna Ivanova – (Georgia Institute of Technology)
Dr. Sashank Varma – (Georgia Institute of Technology)
Dr. Travis E. Baker – (Rutgers University, Newark)

Title: Language as Process Evidence for Modeling Human Cognition

Abstract:
Different cognitive processes can produce the same choice. This dissertation uses large language models to test what verbal reports add to behavioral measurements by linking what people say to what they do. In two risky-choice experiments, think-aloud reports improved predictions of the choice made on the same trial. Prediction accuracy remained above the task-only baseline after explicit choice statements were masked. In sorting, programs generated from participants’ written strategies predicted their actions more accurately on average than programs generated from other participants’ reports. In stochastic learning and risky choice, verbal reports were mapped to strategy labels and numerical decision variables. Supplying think-aloud reports during automated model discovery improved model fit and changed the computational structures of the resulting models. Reinforcement learning on human choice proportions elicited candidate explanations of risky choices. Training on synthetic choices generated by an expected-value rule shifted the explanations toward expected-value comparisons. Together, these studies provide methods for using verbal reports to formulate and test computational accounts of how people make decisions and carry out strategies.