This paper presents STORM, a novel framework that bridges large language models with spatiotemporal forecasting and sequential decision-making. The method aligns multimodal inputs (time series, graph structure, text) into a unified semantic space using contrastive learning, enabling a frozen LLM backbone to jointly optimize forecasting, explanation, and decision policies. Applied to urban traffic management, STORM addresses the gap between predicting congestion and recommending control actions while generating natural language explanations.
Key findings
STORM integrates spatiotemporal graph neural networks with frozen LLM backbones via cross-modal contrastive alignment, enabling zero-shot transfer capabilities.
The framework jointly optimizes forecasting accuracy, decision policies, and explainability through a multi-objective training approach.
Novel architecture combines Graph Attention Networks for spatial dependencies with reprogrammed time series patches for temporal modeling.
Targeted application to urban traffic management demonstrates potential for 11% MAE reduction and 13% decision accuracy improvement over Time-LLM baselines.
Limitations & open questions
Empirical validation section is not included in the provided document, leaving experimental methodology and full results unseen.
The retrieval-augmented component mentioned in the title lacks detailed explanation in the visible methodology sections.