PASM: Population Adaptive Symbolic Mixture-of-Experts Model for Cross-location Hurricane Evacuation Decision Prediction

Researchers Xiao Qian and Shangjia Dong have introduced PASM (Population-Adaptive Symbolic Mixture-of-Experts), a novel AI architecture designed to solve a critical problem in disaster response: predicting whether households will evacuate during hurricanes. Traditional machine learning models trained on data from one region, like Florida, often fail catastrophically when applied to another, like Georgia, due to systematic differences in how similar households make decisions across locations. PASM tackles this by combining two powerful techniques: it uses large language models (LLMs) to guide symbolic regression, which discovers human-readable, closed-form mathematical rules describing evacuation logic, and it employs a mixture-of-experts architecture that automatically identifies data-driven subpopulations and routes each household to a specialized 'expert' model at inference time.

In rigorous testing on real-world data from Hurricanes Harvey and Irma, PASM demonstrated superior performance. When transferring knowledge from Florida and Texas to predict behavior in Georgia using only 100 calibration samples, PASM achieved a Matthews correlation coefficient (MCC) of 0.607. This significantly outperformed leading benchmarks, including XGBoost (0.404), the tabular foundation model TabPFN (0.333), OpenAI's GPT-5-mini (0.434), and meta-learning approaches like MAML (≤0.346). Crucially, the model's routing mechanism assigns distinct, interpretable formula archetypes to different subpopulations, creating clear behavioral profiles for planners.

The research also included a comprehensive fairness audit across four demographic axes—income, race, age, and disability status—finding no statistically significant disparities in the model's performance after correction for multiple comparisons. By closing more than half of the typical cross-location generalization gap while maintaining full transparency, PASM represents a major step toward deployable, trustworthy AI for emergency management, moving beyond black-box predictions to provide actionable, understandable insights for saving lives.