Collaborative Research: III: Foundations for Reliable and Generalizable Spatial-Temporal Large Language Models in Urban Dynamics

Urban dynamics prediction is increasingly powered by Large Language Model (LLM)-based spatial-temporal forecasting models. This approach includes multi-sensor observations (e.g., traffic speeds and travel demand) being tokenized into sequential representations and processed by LLMs to generate numerical forecasts. Although this paradigm has substantially improved predictive accuracy on benchmark datasets, accuracy alone is insufficient for deployment in real-world urban systems. Existing approaches primarily exploit the sequence modeling capabilities of LLMs without pretraining them for spatial-temporal dynamics, leaving their behavior in urban conditions poorly understood and difficult to trust. Consequently, current LLM-based spatial-temporal forecasting models lack: (i) theoretical foundations explaining when and why they succeed or fail; (ii) robustness under distributional and structural shifts, such as sensor failures and disruptive events; and (iii) transparent, controllable reasoning that identifies the spatial regions, temporal contexts, and external events driving predictions while enabling interactive, instruction-guided forecasting under user-specified constraints and objectives. This project addresses these fundamental challenges by advancing LLM-based spatial-temporal forecasting from generic sequence modeling toward theoretically grounded, constraint-aware, and reasoning-capable systems for urban dynamics. By unifying theory, learning, reasoning, and verification within a single framework, the project will establish the scientific foundations for trustworthy, interactive, and scalable spatial-temporal foundation models for urban intelligence. Beyond its technical contributions, the project integrates research with education and workforce development through curriculum innovation, student mentoring, outreach activities, and collaborations with industry and public-sector partners. The technical goal of this project is to develop a novel Constraint-Aware Spatial-Temporal LLM that establishes a new paradigm for theoretically grounded, constraint-aware, and reasoning-capable urban dynamics prediction. The research is organized around three tightly integrated thrusts. The first thrust develops the theoretical foundations for understanding when and why Large Language Model-based spatial-temporal forecasting models succeed or fail by establishing diagnostic theory that links spatial-temporal tokenization, cross-modal representation alignment, and distributional drift to post-fine-tuning generalization and failure regimes, yielding measurable criteria for predicting model breakdown before deployment. Building upon these theoretical insights, the second thrust develops robust and constraint-faithful learning algorithms for spatial-temporal tokenization, representation alignment, and cross-task and cross-city generalization, with guarantees on prediction feasibility and robustness under complex urban constraints. The third thrust introduces interactive reasoning and verification mechanisms that enable users to query, guide, and refine predictions through human-interpretable instructions while automatically verifying consistency with physical laws, operational requirements, and statistical constraints. Collectively, these efforts will establish the scientific foundations for trustworthy, interpretable, and controllable LLM-based systems for urban dynamics. All resulting models, algorithms, datasets, and evaluation frameworks will be released as open-source resources to accelerate research and innovation in urban intelligence and spatial-temporal foundation models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.