Diversifying Multiple Generative Agents by Aligning with Human Populations

Large Language Models (LLMs) are increasingly used in settings where systems must reflect diverse human preferences and behaviors, ranging from assistants for everyday tasks to tools for simulating human behavior in scientific research. However, prior work shows that LLMs often produce homogeneous outputs that fail to capture pluralistic human perspectives and behaviors. Rather than trying to capture this diversity using a single generative agent, we propose a framework for constructing a set of generative agents that collectively serve as faithful behavioral surrogates for a human population. Each agent is an LLM grounded in real human demonstrations (task-response pairs), whose behavior is steered to match that of a specific subpopulation. The challenge is therefore to select a representative set of agents from the exponentially large space of possible agents. We formalize this as minimizing the representation error (the average behavioral distance between each human and their nearest agent) and show that this objective is submodular, enabling methods with approximation guarantees. Extensive experiments in educational and crowdsourcing domains demonstrate that our methods construct agent sets that more faithfully represent human populations than existing methods, and that these agents reproduce subpopulation-level behavioral patterns on unseen tasks.