LLM-Based Agent Simulations for Modeling Individual Differences in Social Interactions (LASSI)

A key challenge in psychological science is the lack of formal models that capture individual differences in social interactions. At the same time, recent advances in artificial intelligence offer unprecedented opportunities to simulate and assess interpersonal behavior at scale. In this work, we bring these developments together by introducing LASSI, an agent-based modeling framework based on Contemporary Integrative Interpersonal Theory (CIIT) that uses large language models (LLMs) to simulate social interactions. Within a dyadic interaction cycle, an LLM-based agent continually adjusts its interpersonal behavior (i.e., agency and communion) in response to the perceived behavior of its interaction partner (either a human or another LLM-based agent). Technically, an agent’s interpersonal behavior is governed by a finite-state Markov model within a discretized agency-communion space. The agent’s state distribution is iteratively updated as a function of its prior state distribution, the interaction partner’s behavior, and the combined influence of complementarity (the tendency of communal behavior to evoke communal behavior and of agentic behavior to evoke less agentic behavior), behavioral defaults (individual differences in general behavioral tendencies), and interpersonal reactivities (individual differences in the strength of reactions to defined partner behaviors). We show how LASSI can be used to simulate, assess, and experimentally manipulate interpersonal dynamics, derive a mathematical model that formalizes how key interpersonal phenomena jointly shape interpersonal behavior, and demonstrate the framework’s flexibility for advancing theories of interpersonal behavior and supporting practical applications within and across psychological science.