Large Language Models in Transportation Systems Management and Operations: From Text Reasoning to Multi-modal Decision Support
A comprehensive survey examines how large language models and multimodal LLMs are being applied to transportation systems management and operations across three domains: operations, fleet services, and decision support. The research identifies LLMs as promising decision-support tools while highlighting key challenges in real-time inference, data integration, and explainability that must be addressed for operational deployment.
This survey documents a significant convergence between artificial intelligence capabilities and critical infrastructure management. Transportation agencies face unprecedented data complexity—sensor streams, incident reports, traveler feedback, and visual observations arrive constantly and require rapid interpretation for safe, efficient operations. LLMs and multimodal variants offer a novel abstraction layer to process these heterogeneous inputs simultaneously, moving beyond siloed analysis systems.
The research reflects broader AI industry maturation. As organizations move beyond early-stage AI pilots, they increasingly seek integration mechanisms that consolidate structured databases, unstructured text, and visual data into unified decision frameworks. Transportation represents an ideal proving ground: the stakes are high, data sources are genuinely heterogeneous, and human operators currently perform much of this integration manually.
The practical implications are substantial but nuanced. For technology vendors, this survey validates market opportunity in decision-support applications—the most promising near-term use case. For transportation agencies and fleet operators, the identified challenges reveal that production deployment requires solving real-time inference latency, building interpretable AI systems (critical for liability), and establishing governance frameworks. The emphasis on "operationally oriented applications" versus prototypes suggests a maturation threshold: systems must move beyond demonstrations to handle production constraints.
Looking forward, success hinges on localized adaptation—transportation needs differ significantly by geography and infrastructure type. Edge deployment becomes essential where cloud latency proves unacceptable. Benchmarking standards remain absent, creating friction for comparative evaluation. Cross-agency collaboration could accelerate standards development, but institutional silos persist in transportation. The survey implicitly argues that LLM-based systems will augment rather than replace human operators in this domain.
- →Multimodal LLMs show particular promise for integrating text, visual, and sensor data in transportation decision-making
- →Real-time inference latency and explainability remain critical unsolved challenges for operational deployment
- →LLMs function most effectively as decision-support layers augmenting human operators rather than autonomous systems
- →Localized adaptation and edge deployment are essential requirements for production transportation applications
- →Benchmarking standards and cross-agency collaboration frameworks remain missing from the LLM-for-transportation ecosystem