Designing Behavioral Profiles for Language Model Simulation of Out-of-Distribution Human Behavior

Large language models (LLMs) are increasingly used to simulate human behavior, but their usefulness depends on whether the simulation aligns with real human behavior, especially as the social context changes. LLM simulation is usually conditioned on descriptions of the people it represents, yet systematic ways to construct such descriptions remain underexplored. We develop a framework that generates behavioral profiles from earlier experiments, designed to improve the out-of-distribution generalization of LLM simulations. We first draw candidate behavioral features from the behavioral literature (e.g., cooperation propensity in public-goods games) and use an LLM to assign a value on a 0-10 scale when an earlier behavioral record provides direct evidence for a feature. A separate, fixed LLM then simulates earlier experiments conditioned on profiles containing different features, allowing us to determine which profile contents most consistently improve agreement with human outcome distributions. We keep the resulting profile contents unchanged and evaluate them on new participants under new experimental designs. In public-goods games, multi-party bargaining, and generalized 11-20 money-request games, profiles built from the leading features produced simulations that matched human outcome distributions more closely than no profile or general-purpose persona sources. Compact profiles containing one or two features performed at least as well as profiles containing every feature that could be assigned from a record. Profiles built from leading features in one type of game transferred unevenly to another, occasionally making simulations worse than simulating without any profile.