Unlocking Many Voices in LLM Social Simulation

Constructing populations for LLM social simulation requires agents that reflect differences among real people. Prompting many agents from a shared model often produces a small set of similar voices, compressing the diversity of the resulting society. To address this problem, we propose ReMap, a method that learns individual variation from real writing and composes it into synthetic populations. We first encode individual variation into model weights through lightweight adaptation. We then compose these learned adaptations across model depth to create agents assembled from multiple source individuals. Experiments across three domains and six backbones show that ReMap populations stay close to human writing while supplying far more mutually distinguishable agents than the library contains, and that they produce more diverse collective behavior than prompted populations inside external simulators. These results suggest a path toward more representative LLM populations for studying collective behavior.