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

Xiao Qian; Shangjia Dong

cs.LGMar 31, 2026

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

Authors: Xiao Qian, Shangjia Dong

Organizations: Dept. of Civil, Environmental, and Construction Engineering, University of Delaware, Newark, DE, USA

Abstract

Accurate prediction of evacuation behavior is critical for disaster preparedness, yet models trained in one region often fail elsewhere. Using a multi-state hurricane evacuation survey, we show this failure goes beyond feature distribution shift: households with similar characteristics follow systematically different decision patterns across states. As a result, single global models overfit dominant responses, misrepresent vulnerable subpopulations, and generalize poorly across locations. We propose Population-Adaptive Symbolic Mixture-of-Experts (PASM), which pairs large language model guided symbolic regression with a mixture-of-experts architecture. PASM discovers human-readable closed-form decision rules, specializes them to data-driven subpopulations, and routes each input to the appropriate expert at inference time. On Hurricanes Harvey and Irma data, transferring from Florida and Texas to Georgia with 100 calibration samples, PASM achieves a Matthews correlation coefficient of 0.607, compared to XGBoost (0.404), TabPFN (0.333), GPT-5-mini (0.434), and meta-learning baselines MAML and Prototypical Networks (MCC $\leq$ 0.346). The routing mechanism assigns distinct formula archetypes to subpopulations, so the resulting behavioral profiles are directly interpretable. A fairness audit across four demographic axes finds no statistically significant disparities after Bonferroni correction. PASM closes more than half the cross-location generalization gap while keeping decision rules transparent enough for real-world emergency planning.

Figures & tables

Figure 1: Evacuation probability difference in UMAP (Uniform Manifold Approximation and Projection) space: $P(\text{Evac}|\text{Georgia})-P(\text{Evac}|\text{Florida})$ (left) and $P(\text{Evac}|\text{Georgia})-P(\text{Evac}|\text{Texas})$ (right).

Figure 2: Overview of the PASM framework.

Model	MCC	ROC-AUC	Accuracy
XGBoost	0.404	0.680	0.692
TabPFN	0.333	0.751	0.654
GPT-5-mini(medium)	0.434	0.692	0.692
PASM	0.607	0.840	0.769

Table 1: Cross-location adaptation performance in Georgia.

Method	MCC	ROC-AUC	Accuracy
Matching Networks	0.280 $\pm$ 0.117	0.732 $\pm$ 0.096	0.589 $\pm$ 0.045
MAML	0.314 $\pm$ 0.158	0.753 $\pm$ 0.061	0.631 $\pm$ 0.071
Prototypical Networks	0.313 $\pm$ 0.162	0.763 $\pm$ 0.092	0.650 $\pm$ 0.078
HierClust+LR	0.346 $\pm$ 0.164	0.746 $\pm$ 0.100	0.635 $\pm$ 0.071
PASM (Ours)	0.607	0.840	0.769

Table 2: Comparison with meta-learning and heterogeneity-aware baselines.

Method	MCC	ROC-AUC	Accuracy
Top-1 Average	0.234	0.766	0.615
Top-5 Average	0.548	0.769	0.731
No SR (LR + Router)	0.434	0.752	0.731
PASM	0.607	0.840	0.769

Table 3: Ablation on aggregation strategy.

Coefficient Strategy	MCC	ROC-AUC	Accuracy
Fixed Coefficients	0.389	0.763	0.692
Learnable (Static)	0.365	0.828	0.654
PASM (Full)	0.607	0.840	0.769

Table 4: Ablation on coefficient adaptation strategy.

Figure 3: Cross-state transferability: models trained on Florida and evaluated on Florida (intra-state) vs. Georgia/Texas (cross-state).

	Florida			Georgia			Texas
Model	MCC	AUC	Accuracy	MCC	AUC	Accuracy	MCC	AUC	Accuracy
GPT-5-mini	.736	.867	.867	.571	.791	.787	.736	.867	.867
TabPFN	.757 $\pm$ .046	.962 $\pm$ .009	.877 $\pm$ .023	.313 $\pm$ .067	.773 $\pm$ .018	.678 $\pm$ .026	.717 $\pm$ .055	.953 $\pm$ .012	.854 $\pm$ .030
XGBoost	.729 $\pm$ .040	.944 $\pm$ .017	.864 $\pm$ .020	.325 $\pm$ .076	.719 $\pm$ .029	.686 $\pm$ .034	.714 $\pm$ .040	.927 $\pm$ .016	.855 $\pm$ .021

Table 5: Cross-state transfer performance when training on Florida (100 shots) and testing on each state.

Model	Train $\to$ Test	MCC	AUC	Accuracy
GPT-5-mini	FL $\to$ FL	.736	.867	.867
	FL $\to$ GA	.571	.791	.787
	GA $\to$ GA	.295	.691	.795
TabPFN	FL $\to$ FL	.757 $\pm$ .046	.962 $\pm$ .009	.877 $\pm$ .023
	FL $\to$ GA	.313 $\pm$ .067	.773 $\pm$ .018	.678 $\pm$ .026
	GA $\to$ GA	.274 $\pm$ .101	.763 $\pm$ .066	.729 $\pm$ .037

Table 6: Within-state heterogeneity analysis (100 training shots).

Model	Perception	MCC	AUC	Acc
GPT-5-mini	Without	.571	.791	.787
GPT-5-mini	Oracle	.629	.822	.813
TabPFN	Without	.313 $\pm$ .067	.773 $\pm$ .018	.678 $\pm$ .026
TabPFN	Oracle	.335 $\pm$ .055	.797 $\pm$ .021	.694 $\pm$ .021
XGBoost	Without	.325 $\pm$ .076	.719 $\pm$ .029	.686 $\pm$ .034
XGBoost	Oracle	.349 $\pm$ .038	.765 $\pm$ .024	.699 $\pm$ .018

Table 7: PADM-inspired two-stage prediction: Florida

\to

Georgia transfer (100 shots).

Cluster	Expert Formula	Demographic Profile	Behavioral Pattern
C0	$\text{TimesAsked}+\text{EvacPctZip}$	Youngest (age 33.4), unmarried, low education	Direct response to external pressure
C1	$\frac{\text{EvacPctZip}}{c_{0}}+\ldots-\log(\text{Age})$	High age (69.5), non-standard marital status	Geographic signal vs. age resistance
C2	$\frac{\text{EvacPctZip}}{c_{0}}+\ldots-\log(\text{Age})$	Near-mean demographics, married females	Calibration baseline group
C3	$\frac{\text{EvacPctZip}}{c_{0}}-\log(\text{Age})$	Older (59.0), widowed females	Community behavior vs. age
C4	$\frac{\text{EvacPctZip}}{c_{0}}-\log(\text{Age})$	Young (34.5), divorced, low education	Geographic rate vs. age
C5	$\frac{\text{EvacPctZip}}{c_{0}}-\log(\text{Age})$	Young (33.1), high education, males with children	Minimal age resistance, high compliance (96.9%)

Table 8: Cluster profiles showing routed expert formulas, demographic characteristics, and behavioral interpretations. Three formula archetypes emerge: (A) linear additive (C0 only), (B) geographic rate vs. log-age (C3,4,5,6,9), and (C) multi-factor with cosine and root transforms (C1,2,7,8).

Metric	Value	Threshold
Adjusted Rand Index (ARI)	0.876	$>$ 0.8 (strong)
Normalized Mutual Information (NMI)	0.916	$>$ 0.8 (high)
Co-clustering Jaccard	0.822	$>$ 0.75 (reliable)
Cluster count (mode)	8	range [6,11]
Noise fraction	0.020	—

Table 9: Clustering stability metrics for Experiment A (20 runs with varying UMAP random state, fixed data split).

Method	MCC	LLM Calls	LLM Tokens	Wall Time (s)
PASM (Full)	0.607	2,068	4,508,756	24,699
No Routing (Top-1 SR)	0.234	150	313,483	1,718
No SR (LR + Router)	0.434	0	0	45

Table 10: Computational cost breakdown. GPU memory: Ollama LLM server 5.720 GiB peak, router training 1,576 MiB. Throughput: 131 tok/s (A6000 Ada), 79 tok/s (RTX 4070 Ti).

Calibration Shots	MCC
20	0.426
30	0.488
50	0.548
80	0.566
100	0.607

Table 11: MoE router calibration MCC across target-domain sample sizes. Source-domain experts and Georgia test set are held fixed.

Explore similar work

LLM-powered reasoning in agent-based modeling
Jul 7, 2026Sifat Afroj Moon, Dakotah Maguire, Adam Spannaus +5Agent-Based Modeling Large Language Model-Based Social Simulation
Oak Ridge National Laboratory · Computational Science and Engineering Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830, USA · Geospatial Science and Human Security Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830, USA +1
ARMove: Learning to Predict Human Mobility through Agentic Reasoning
Apr 19, 2026Chuyue Wang, Jie Feng, Yuxi Wu +2Transfer Learning Agentic Reasoning
Tencent · Zhongguancun Academy · Alibaba Group +1
Hierarchical Generative Agents for Simulating Sequential Human Behavior
Jun 12, 2026Maria G. Mendoza, Lucas Waldburger, Jin Lee +1Cognitive Modeling Human Behavior Simulation
University of California, Berkeley

Jul 7, 2026J/K move · Enter open · S save

LLM-powered reasoning in agent-based modeling

Sifat Afroj Moon, Dakotah Maguire, Adam Spannaus, Joe Tuccillo+4

Oak Ridge National Laboratory · Computational Science and Engineering Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830, USA · Geospatial Science and Human Security Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830, USA+1

Agent-Based Modeling Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

Agent-based modeling (ABM) has the capability to model millions of individuals and their interactions, which is useful for policy making. However, ABMs have traditionally relied on static prior, which prevents the models from adapting to real-time changes. Our research provides a novel approach to addressing this information gap. Large language models (LLMs) offer new opportunities to predict human decision-making. Here, we introduce a scalable Hybrid Agent-based and Language-driven Epidemic (HALE) modeling framework that leverages LLMs to predict human decision-making in an ABM simulation. As a proof-of-concept, we use HALE to simulate COVID-19 and its effects in Salt Lake County, UT.

Abstract

Explore similar work

LLM-powered reasoning in agent-based modeling

ARMove: Learning to Predict Human Mobility through Agentic Reasoning

Hierarchical Generative Agents for Simulating Sequential Human Behavior