Abstract

Modern transportation systems face pressing challenges due to increasingdemand, dynamic environments, and heterogeneous information integration. Therapid evolution of Large Language Models (LLMs) offers transformative potentialto address these challenges. Extensive knowledge and high-level capabilitiesderived from pretraining evolve the default role of LLMs as text generators tobecome versatile, knowledge-driven task solvers for intelligent transportationsystems. This survey first presents LLM4TR, a novel conceptual framework thatsystematically categorizes the roles of LLMs in transportation into foursynergetic dimensions: information processors, knowledge encoders, componentgenerators, and decision facilitators. Through a unified taxonomy, wesystematically elucidate how LLMs bridge fragmented data pipelines, enhancepredictive analytics, simulate human-like reasoning, and enable closed-loopinteractions across sensing, learning, modeling, and managing tasks intransportation systems. For each role, our review spans diverse applications,from traffic prediction and autonomous driving to safety analytics and urbanmobility optimization, highlighting how emergent capabilities of LLMs such asin-context learning and step-by-step reasoning can enhance the operation andmanagement of transportation systems. We further curate practical guidance,including available resources and computational guidelines, to supportreal-world deployment. By identifying challenges in existing LLM-basedsolutions, this survey charts a roadmap for advancing LLM-driven transportationresearch, positioning LLMs as central actors in the next generation ofcyber-physical-social mobility ecosystems. Online resources can be found in theproject page: https://github.com/tongnie/awesome-llm4tr.