Neuro-Symbolic Generative Adapters for Urban Foundation Models

Urban foundation models (UFMs) have emerged as powerful tools for understanding and optimizing urban systems, but adapting these large models to diverse urban tasks remains challenging. This paper introduces Neuro-Symbolic Generative Adapters (NSyGA), a novel parameter-efficient fine-tuning (PEFT) framework. NSyGA integrates symbolic reasoning with neural generative policies to dynamically produce task-specific adapter parameters. NSyGA parses urban task descriptions into symbolic representations, learns adaptation strategies through reinforcement learning, and generates compact adapter parameters that modify only the necessary components of the foundation model. Through extensive experiments on real-world urban datasets, we demonstrate that NSyGA outperforms state-of-the-art parameter-efficient methods in accuracy (9.3% improvement), parameter efficiency (29% reduction), and interpretability. Our comprehensive analysis of computational complexity, scalability, and fairness highlights NSyGA's practical applicability to resource-constrained urban computing scenarios. The neuro-symbolic approach enables better generalization to compositionally novel tasks and provides insights into the relationship between task characteristics and adaptation strategies.