Understanding the Interplay of Human Modeling and Social Intelligence via LLM Simulations

This research project aims to advance the field of social intelligence in Large Language Models (LLMs) by investigating the interplay between human modeling and social behavior. I challenge the current unsystematic approach to social intelligence benchmarking in LLMs and propose a novel framework incorporating personalized mental models and contextual understanding to determine the components with the biggest influence on the discourse. Through a combination of longitudinal psychometric studies, gamified data collection and evaluation, advanced representation learning, and explainable AI techniques, I will investigate the role of fine-grained personal and social context in shaping human and LLM interactions. Through extensive simulations via social LLM agents, combined with data analysis techniques, this research will help to shed light on the relation between psychological traits and processes, their demonstrations in social discourse, and the representation and controllability of these in LLMs, and ultimately contribute to the development of more socially adept and effective LLMs for potential applications in more vulnerable domains such as assisted living or health counseling.