Sense, Think, Act, Reflect: Distilling Fast and Interpretable Decision Functions from LLM-Driven Crowds

Large language models (LLMs) have been shown to be capable of generating human-like agent behavior in diverse scenarios, making them useful building blocks for agent-based simulations. However, the substantial inference cost restricts the crowd sizes that can be tackled, and the opaque nature of LLM-based decision making raises reliability concerns. To address these issues, we propose the approach of Decision Function Distillation (DFD), which extracts strategies underlying the decision making of LLM agents in a rule-based and interpretable form. The final decision function is determined in an iterative process during which intermediate insights gathered from historical agent trajectories are refined and finally translated into commented code. In two variants of the approach, the intermediate insights are either generated directly as text based on few-shot examples or explicitly formalized into code snippets. Unlike black-box symbolic regression (SR), the gradual and transparent refinement process allows modelers to understand the strategies captured in the final commented decision function. We demonstrate DFD on agent-based crowd evacuation scenarios, showing that DFD outperforms both classical and a state-of-the-art LLM-based SR.