面向交通与运输研究的大语言模型：方法论、前沿进展与未来机遇

abcdefghijkl 1.IntroductionTraffic and transportation have been pivotal in shaping human civilization throughout history.From the rise and fall of empires driven by maritime trade routes to the development of intricateroad networks facilitating urban expansion, the movement of people and goods has always been acornerstone of societal advancement since the 20th century Before Christ (Gianpaolo et al., 2013).Efficient transportation systems have enabled economic growth, cultural exchange, and technologicalprogress, while also presenting challenges related to congestion, safety, and environmental impact.In the 20th century, the advent of computer technologies revolutionized traffic and transportationresearch. The introduction of optimization algorithms and prediction models has allowed for moresystematic and efficient planning of transportation networks. These advancements have enabled bet-ter traffic management, route optimization, and forecasting of transportation demands, significantlyimproving the functionality of transportation systems. However, despite these technological strides,several persistent issues remain unresolved. Modern transportation systems generate vast amountsof heterogeneous data, encompassing numerical metrics, videos, images, and unstructured textualinformation from diverse sources such as traffic reports, social media, and sensor logs. Traditionaloptimization and prediction algorithms, while powerful, often struggle to integrate and interpret thismultifaceted data effectively.Recent developments in artificial intelligence, particularly Large Language Models (LLMs), havethe potential to address these challenges. LLMs, such as generative pre-trained transformer(GPT)-4,bidirectional encoder representations from transformers (BERT), and their derivatives are advancedartificial intelligence (AI) systems trained on extensive datasets to understand, generate, and manipu-late human language with high proficiency. These models leverage Transformer architectures (Vaswaniet al., 2017), enabling them to capture complex linguistic patterns and contextual relationships. Be-yond natural language processing (NLP), LLMs exhibit capabilities in reasoning, data integration,and multimodal understanding, making them well-suited for applications in traffic and transportationresearch.LLMs can undertake a variety of tasks critical to enhancing transportation systems. They canautomate the extraction and summarization of information from unstructured data sources, improvethe accuracy of traffic forecasts by integrating textual and numerical data, assist in scenario generationfor planning and emergency response, and facilitate better decision-making through sophisticateddata analysis and interpretation. These capabilities not only enhance the efficiency and safety oftransportation systems but also contribute to sustainability by optimizing resource allocation andreducing emissions.The purpose of this paper is to provide a comprehensive review of recent methodologies and appli-cations of LLMs in traffic and transportation research. Our goal is to present and highlight the state ofthe art and the potential of LLMs within the traffic and transportation research community, thereby outlining promising directions for future research. The specific research questions to be addressedinclude:•In which areas of traffic and transportation research are LLMs more promising for adoption?•Which LLM methods are more appropriate to tackle specific traffic and transportation problems?•What are the challenges and future opportunities for LLMs in traffic and transportation research?Our paper is organized as follows. In Section 2, we introduce the background and the core method-ologies in LLMs. In Section 3, applications are classified into two broad categories, namelytrafficandtransportation. In Section 4, we present statistics of current research trends and future directions.Finally, we conclude this paper in Section 5. The abbreviations used in the paper are presented inTable 1.2.Background of LLMsBetween six and eleven months, a child typically starts to learn its language from the surroundingenvironment (Health, 2025). A newborn is exposed to an overwhelming amount of linguistic input –parents talking, sibling chattering, TV sounds, and even books they see. Initially, these exposures arenoise. But gradually, through consistent exposure and the pattern recognition repertoire of infants’brains, the child begins to make sense of the sea of information (Jurafsky and Martin, 2025).This remarkable process mirrors how LLMs learn, beginning with their data foundation. Just aschildren absorb massive amounts of language input during their formative years, LLMs begin theirdevelopment with enormous text datasets. The parallel extends to the processing mechanism: justas human sensory organs (eyes and ears) perform initial pre-processing of linguistic input beforeneural transmission, LLMs employ sophisticated pre-processing techniques to tr