Large Language Models in Public Transportation

• Large Language Models (LLMs) are transformer-based AI tools that process unstructured transit data, enabling zero-shot reasoning and robust demand forecasting.
• They integrate techniques such as retrieval-augmented generation and fine-tuning to optimize routing, schedule management, and adaptive control in transit systems.
• LLMs improve customer interactions and policy support by automating trip planning, extracting feedback, and providing clear, interpretable reasoning.

LLMs are transforming the landscape of public transportation research, operations, planning, and customer-facing applications. By leveraging pre-trained, high-capacity transformer architectures—such as GPT-4, RoBERTa, and Llama 3—these models have demonstrated the ability to process unstructured data, encode contextual and semantic knowledge, perform robust reasoning, and deliver interpretable outputs for a range of public transit use-cases. This article presents a comprehensive overview, rooted in recent research, of the methodological paradigms, application domains, evaluation strategies, performance profiles, and ongoing challenges in deploying LLMs across public transport systems.

1. Methodological Foundations and Model Architectures

LLMs applied in public transportation are typically based on large-scale transformer architectures—decoder-only (e.g., GPT-3.5, GPT-4), encoder-only (e.g., BERT, RoBERTa), and encoder–decoder (Seq2Seq) variants. Off-the-shelf models are employed for few-shot and zero-shot reasoning, often augmented with prompt engineering or fine-tuning on domain data. Important paradigms include:

• Zero-shot and few-shot learning: LLMs can infer relationships or make predictions without explicit model parameter training on task-specific datasets. Few-shot chain-of-thought (CoT) prompting, which encourages the model to provide explicit, step-by-step reasoning, yields improved forecasting and interpretability (Mo et al., 2023).
• Retrieval-Augmented Generation (RAG): Contextual augmentation mechanisms retrieve relevant documents (e.g., schedules, alerts, station gazetteers) based on embedding similarity and prepend them to the prompt, enabling accurate, up-to-date information extraction (Wang et al., 2024).
• Fine-tuning and domain adaptation: Models like MetRoBERTa utilize fine-tuning on transit-CRM feedback to acquire transit-specific representations for topic classification (Leong et al., 2023).
• Plug-in and multi-agent architectures: LLMs can orchestrate specialized sub-agents for system tasks such as multimodal narrative generation, API tool usage, or real-time knowledge retrieval (Ma et al., 17 Nov 2025, Wang et al., 2024).

2. Prediction, Planning, and Demand Forecasting

LLMs have been employed for demand forecasting, travel behavior prediction, and operational decision support under both routine and disrupted conditions:

• Event-based human mobility prediction: The LLM-MPE framework processes raw, unstructured event descriptions via LLMs, standardizes them, and combines these with historical inflow/outflow decompositions to predict event-induced demand surges, offering point forecasts with explicit rationales (Liang et al., 2023). Compared to gradient boosting baselines, LLM-MPE achieves a 13.9% RMSE reduction and a 17.8% MAE reduction on event days.
• Integration with spatiotemporal models: Rather than direct end-to-end forecasting, textual contextual data (weather, events) is encoded via LLMs into high-dimensional embeddings, which are then injected as auxiliary nodes in spatiotemporal graph neural networks (GNNs), leading to substantial MAE and RMSE improvements for both city-level and node-specific predictions (Huang, 2024).
• Passenger travel choice under disruptions: DelayPTC-LLM models individual passenger responses to metro delays using GPT-4, incorporating passenger heterogeneity and delay features through natural-language prompts. On highly imbalanced, sparse datasets, DelayPTC-LLM attains the highest accuracy (0.83) relative to classical baselines, with balanced recall and precision, and provides reasoning chains that aid operations (Chen et al., 2024).
• Travel behavior modeling: Zero-shot LLMs can infer individual mode choice from scenario and traveler attributes, producing results close to multinomial logit, random forest, and neural network methods, with the additional benefit of interpretable, text-based rationales (Mo et al., 2023).

3. Information Management, Customer Interaction, and Policy Support

LLMs are augmenting traditional databases and APIs to deliver adaptive, context-aware, and conversational information systems:

• Natural language interfaces: Through prompt chaining, tool integration, and retrieval-augmented techniques, LLMs automate trip planning, schedule queries, and policy navigation. These systems mediate between unstructured user queries, structured GTFS feeds, and internal policy repositories, providing real-time recommendations and policy guidance (Wang et al., 2024, Jonnala et al., 7 Jan 2025).
• Social media and CRM feedback extraction: Transit-topic–aware models such as MetRoBERTa classify open-ended feedback into actionable topic labels with 90% accuracy, outperforming keyword-based baselines. Pipelines extend to include sentiment analysis, demographic inference, and spatio-temporal visualization, enabling large-scale, real-time feedback aggregation (Leong et al., 2023).
• Real-time social media mining: Retrieval-augmented LLMs (Llama 3 + RAG) process transit-related tweets, simultaneously extracting station names, sentiment, sarcasm, and problem summaries with higher accuracy and interpretive nuance than traditional NLP methods, surfacing emergent issues and enhancing responsiveness (Wang et al., 2024, Ruan et al., 2024).
• Multi-agent and multimodal analytics: In domains such as fuel-efficiency analytics, specialized LLM agents coordinate to transform multimodal sensor data, model outputs, and visualizations into concise, stakeholder-oriented narratives with high factual accuracy (>97%), governed by a “judge agent” for deterministic quality assurance (Ma et al., 17 Nov 2025).

4. Operations, Routing, and Adaptive Control

LLMs can reason about constraints, disruptions, and multi-modal user demands for both strategic planning and real-time operations:

• Flexible, constraint-aware routing: Approaches such as TraveLLM allow users to specify complex constraints (e.g., avoidance of crowded or disrupted stations, multimodal options) via multi-modal inputs (text and map images). The LLM planner produces detailed alternative paths, and a downstream summarization LLM enforces structured outputs; experimental evaluation shows that LLMs outperform conventional apps in handling disruptions and user preferences, with GPT-4 excelling in connectivity and disruption avoidance (Fang et al., 2024).
• Alignment in transportation policy: Multi-agent LLM simulations model democratic referenda on transit policies, capturing heterogeneity and context sensitivity of community preferences via approval or ranked-choice voting from agent LLMs (e.g., GPT-4o, Claude-3.5). LLM referenda approximate, though may systematically diverge from, classical optimization solutions and highlight model-specific behavioral biases and sentiment patterns (Yan et al., 15 Oct 2025).
• Automated network and scenario generation: LLM-integrated frameworks (e.g., NGAI) interface with network modeling/simulation plugins to drive the end-to-end design of road and transit networks from natural language or multimodal prompts, supporting rapid prototyping and scenario-based evaluation (Chen et al., 2024).

5. Evaluation Protocols, Benchmarks, and Empirical Findings

Empirical evaluations are conducted via both standard machine learning and domain-specific metrics:

• Forecasting: MAE, RMSE, MAPE, and $R^2$ for demand/inflow predictions; Jensen–Shannon divergence for statistical similarity in synthetic mobility chains (Liu et al., 2024, Liang et al., 2023, Zhang et al., 2024).
• Classification and reasoning: Topic, sentiment, and mode detection use accuracy, precision, recall, F1-score, as well as human-verified benchmarks (Leong et al., 2023, Ruan et al., 2024).
• Operational benchmarks: Exact-match accuracy on GTFS-based “understanding” and “retrieval” tasks reach up to 98% and 90% (simple queries) on off-the-shelf GPT-4 (Jonnala et al., 2024, Jonnala et al., 7 Jan 2025).
• Narrative generation: Informativeness, narrative accuracy, and readability indices are employed in multi-agent experimental setups (Ma et al., 17 Nov 2025).
• Engineering reasoning: The TransportBench benchmark assesses LLM ability in undergraduate-level transit problems, with leading models achieving up to 60%+ accuracy zero-shot, and highlighting inconsistencies and hidden reasoning flaws (Syed et al., 2024).

6. Limitations, Challenges, and Future Directions

While LLMs bring flexibility, transparency, and high-level reasoning, several open challenges remain:

• Scalability and cost: High compute and API expense, especially for GPT-4-class models, limit deployment frequency and scope. Hybrid pipelines (routine ML for simple cases, LLMs for events/disruptions) and the adoption of open-source, lighter-weight models (PEFT, quantization) are proposed (Liang et al., 2023, Yan et al., 27 Mar 2025).
• Information reliability: LLMs may hallucinate or miscalculate, especially in arithmetic-intensive or logic-heavy chains. Retrieval-augmentation, human-in-loop verification, and structured output enforcement are critical mitigations (Liang et al., 2023, Wang et al., 2024).
• Context and domain knowledge gaps: LLMs may lack up-to-date or detailed domain understanding unless explicitly fine-tuned or coupled to retrieval modules. Domain-adapted fine-tuning and knowledge-grounded prompting are active areas of development (Wang et al., 2024, Leong et al., 2023).
• Privacy, ethics, and regulatory issues: Real-time transit data and customer queries pose data privacy and equitable access challenges, motivating research on federated training, explainable outputs, and fairness auditing (Shoaib et al., 2024, Yan et al., 27 Mar 2025).
• Interoperability with operational pipelines: Standardized, modular toolkits and task-oriented prompt libraries, as well as robust APIs for legacy database and simulation integration, are needed for broader industry adoption (Chen et al., 2024).

7. Perspectives and Research Opportunities

LLMs have established themselves as versatile tools for integrating open-ended data—textual, multimodal, behavioral, and operational—across public transit workflows. Promising research directions include:

• Composition with classical optimization and simulation frameworks to create hybrid human–AI loops for planning and operational control (Yan et al., 15 Oct 2025, Yan et al., 27 Mar 2025).
• Real-time, on-device inference for dynamic control applications (bus holding, load balancing), leveraging lightweight, quantized LLMs (Yan et al., 27 Mar 2025).
• Expansion of public-domain benchmarks and multi-city, multi-modal datasets for robust, comparative evaluation (Syed et al., 2024, Zhang et al., 2024).
• Algorithmic fairness analysis, explainability, and user-centric evaluation in both routine and disaster/disruption scenarios (Ruan et al., 2024, Yan et al., 27 Mar 2025).

LLMs are thus positioned as foundational components in the emerging toolkit for intelligent, responsive, and human-centric public transportation systems. Their integration demands rigorous methodology, robust validation, and continuous adaptation to the evolving technical, operational, and societal landscape of urban mobility.