Using LLM to generate function-based models of crowds.

What is it about?

This work explores how large language models can be used to create realistic decision-making models for crowd simulation, such as building evacuation. While LLM-based agents can behave in human-like ways, they are expensive to run at large scale, and the black-box nature of their decision-making raises reliability concerns. To address these issues, we propose Decision Function Distillation, a method that turns the behavior of LLM agents into simpler, rule-based decision functions. Instead of relying on an LLM every time an agent makes a decision, our approach studies past agent behavior, extracts the main decision strategies in natural language, and gradually converts them into clear, commented code. This makes the resulting agents faster, cheaper to run, and easier for researchers to inspect and trust. We test this method in crowd evacuation scenarios. The results show that our approach can reproduce useful decision-making behavior more effectively than traditional symbolic regression methods and a recent LLM-based symbolic regression method. Overall, the study shows that LLM agents can be used not only as components of simulation models, but also as tools for building simpler and more interpretable decision rules for large-scale agent-based simulations.

Featured Image

Photo by Bernard Hermant on Unsplash

Photo by Bernard Hermant on Unsplash

Why is it important?

As agent-based simulation becomes increasingly popular, generating accurate models and providing explanations for agents’ decisions have become significant concerns. The method proposed in this article leverages the capabilities of LLMs in commonsense reasoning and code generation to construct explicit decision-making functions. It provides human experts with a way to better understand and scrutinize the behavior of simulated agents.