Ryan Liu will present his General Exam " From Overthinking to Simulation: Cognitive Science Approaches to Understanding and Applying LLMs " on Friday, October 3, 2025 at 3:30 PM in Friend 108 and via zoom. Zoom link: https://princeton.zoom.us/j/5754794053 Committee Members: Tom Griffiths (advisor), Lydia Liu , Manoel Ribeiro Abstract: The rapid development of large language models (LLMs) brings challenges as well as opportunities. Techniques like reasoning and RLHF have made models difficult to understand, but the resulting models are capable of novel applications like simulating human behavior. My research aims to deepen our understanding and broaden the applications of these models, using foundational concepts in cognitive and social sciences to inform LLM research. To understand models, I ground ideas such as “overthinking” and “helpfulness/honesty” into theoretically sound cognitive frameworks (deliberation and Gricean maxims, respectively). This approach lets me evaluate LLMs in rigorous experiments with intuitive results, such as relating chain-of-thought to human overthinking, studying how LLMs ascribe rationality to humans, and showing how LLMs navigate conflicts between honesty and helpfulness. To explore the capacity of LLMs to simulate people, I have worked on applications like simulating audiences to improve communication and emulating human research subjects. The former is inspired by episodic future thinking, whereas the latter tackles a classic challenge in social science. Broadly, I believe LLMs can benefit greatly from the social and cognitive sciences, and hope to continue my work bridging these fields. Reading List: https://docs.google.com/document/d/1LvO5lFA6qX2enACuY00975duqfRcLWxiHD7SE6Lw... Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.