PaperlayerView exact revision on arXiv

Source-linked AI summary

Exploring the Roles of Large Language Models in Reshaping Transportation Systems: A Survey, Framework, and Roadmap

Tong Nie, Jian Sun, Wei Ma

arXiv:2503.21411v2cs.AI

TL;DR

Modern transportation systems face congestion, dynamic environments, and fragmented heterogeneous information, while existing LLM transportation studies lack an integrated framework. This survey introduces LLM4TR and a four-role taxonomy, reviews applications and trends, and provides deployment guidance; it also identifies persistent information-fusion and domain-knowledge limitations.

  • Problem

    Existing LLM transportation studies focus on isolated applications and lack a principled framework explaining how LLM functionalities interconnect across transportation systems.

  • Method

    The survey introduces LLM4TR, a role-centric taxonomy covering sensing, learning, modeling, and managing through four LLM roles, and reviews applications, trends, resources, and deployment guidance.

  • Results

    The survey organizes transportation research around four LLM roles and identifies decision facilitation and component generation as prominent directions while highlighting underexplored areas.

  • Takeaways & Limitations

    LLM4TR offers a methodological basis for relating LLM capabilities to transportation-system integration and for guiding future research and real-world development.

  • Takeaways & Limitations

    Current information-processing studies mostly use simple integrations, leaving cross-domain fusion and external transportation knowledge insufficiently exploited.

Abstract

from arXiv · show

Modern transportation systems face pressing challenges due to increasing demand, dynamic environments, and heterogeneous information integration. The rapid evolution of Large Language Models (LLMs) offers transformative potential to address these challenges. Extensive knowledge and high-level capabilities derived from pretraining evolve the default role of LLMs as text generators to become versatile, knowledge-driven task solvers for intelligent transportation systems. This survey first presents LLM4TR, a novel conceptual framework that systematically categorizes the roles of LLMs in transportation into four synergetic dimensions: information processors, knowledge encoders, component generators, and decision facilitators. Through a unified taxonomy, we systematically elucidate how LLMs bridge fragmented data pipelines, enhance predictive analytics, simulate human-like reasoning, and enable closed-loop interactions across sensing, learning, modeling, and managing tasks in transportation systems. For each role, our review spans diverse applications, from traffic prediction and autonomous driving to safety analytics and urban mobility optimization, highlighting how emergent capabilities of LLMs such as in-context learning and step-by-step reasoning can enhance the operation and management of transportation systems. We further curate practical guidance, including available resources and computational guidelines, to support real-world deployment. By identifying challenges in existing LLM-based solutions, this survey charts a roadmap for advancing LLM-driven transportation research, positioning LLMs as central actors in the next generation of cyber-physical-social mobility ecosystems. Online resources can be found in the project page: https://github.com/tongnie/awesome-llm4tr.

1. Introduction

Modern transportation systems must address cyber-physical-social complexity, heterogeneous data, and dynamic mobility demands, while existing ITS approaches remain fragmented and limited. This survey responds with LLM4TR, a unified role-centric framework and taxonomy spanning transportation tasks, research trends, practical deployment, and future challenges.

  • 1.1. Motivation: Transportation systems face congestion, resilience, sustainability, adaptability, and heterogeneous-data challenges across increasingly complex mobility ecosystems.Emerging autonomous, aerial, shared, and AI-enabled mobility systems expose limitations in static models and fragmented data pipelines.
  • 1.2. Scope and focus: LLM4TR organizes transportation tasks as sensing, learning, modeling, and managing, and LLM functionalities as information processor, knowledge encoder, component generator, and decision facilitator.The framework provides the survey’s organizing taxonomy for reviewing LLM integration across the transportation system lifecycle.
  • 1.1. Motivation: LLMs offer language understanding, in-context learning, multimodal reasoning, and human-like decision making for transportation applications.Reported uses include integrating semantic context into traffic prediction, understanding multimodal scenarios, and controlling traffic through natural-language interfaces.
  • 1.1. Motivation: Existing LLM transportation studies largely examine isolated applications and lack a principled framework explaining how their functionalities interconnect.The survey identifies a need for a unified conceptual framework to guide integration across transportation systems.
  • 1.2. Scope and focus: The survey differentiates itself from application-oriented reviews through a methodological taxonomy, broad literature coverage, practical guidance, and a research-trend roadmap.It also curates datasets, benchmarks, tools, literature, libraries, hardware guidance, and an online project repository.

2. Background and Overview

Transportation research is organized around sensing, learning, modeling, and managing tasks, while LLMs provide foundational capabilities relevant to heterogeneous, dynamic transportation systems. This section introduces these tasks and the background capabilities that motivate applying LLMs to transportation.

  • Sensing: Sensing acquires traffic, environmental, and traveler data through infrastructure-based, edge, mobile, and social-media sources.These sources improve data granularity but introduce noise, sparsity, and heterogeneity.
  • Learning: Learning extracts patterns and predictive insights from large, heterogeneous transportation datasets, connecting sensing with decision-making.Deep learning includes supervised, unsupervised, and reinforcement learning, with graph-based methods representing spatio-temporal relationships.
  • Modeling: Modeling formulates and simulates transportation systems to represent traffic dynamics, travel behavior, and infrastructure interactions.Dynamic traffic assignment models simulate network-wide trajectories but struggle with scalability in large, multimodal networks.
  • Managing: Managing optimizes and controls transportation operations by using sensing, learning, and modeling insights with optimization and control methods.Examples include adaptive signal control, fleet dispatching, congestion pricing, and deep reinforcement learning.
  • LLM background: LLMs are Transformer-based models trained on massive text data whose scale supports general problem-solving capabilities.Emergent abilities include in-context learning, instruction following, and step-by-step reasoning, while in-context learning adapts to tasks without retraining.

2.3. Key techniques in LLMs

Key LLM techniques span pretraining, decoding, Transformer architecture, post-training, and prompt engineering. Together, these techniques shape model representations, output generation, task adaptation, computational efficiency, and practical use.

  • Pretraining: Pretraining learns language representations from large-scale corpora using autoregressive, autoencoding, or combined objectives.Autoregressive models predict subsequent tokens, whereas autoencoding models reconstruct masked or corrupted spans.
  • Decoding: Decoding generates text through search-based or sampling-based strategies, with temperature controlling the probability distribution used for token selection.Lower temperature increases the likelihood of high-probability words and decreases the likelihood of low-probability words.
  • Architecture: Transformer architecture uses self-attention, positional encoding, and normalization to model contextual relationships and stabilize large-scale training.Multi-head attention captures different semantic relationships, while standard self-attention has quadratic complexity for long sequences.
  • Post-training: Post-training techniques adapt pretrained models through instruction tuning, alignment tuning, and model adaptation for broader, safer, or more efficient deployment.The reviewed approaches include instruction tuning, RLHF-based alignment, and parameter-efficient fine-tuning.
  • Practical utilization: Prompt engineering crafts task descriptions, inputs, context, and styles to elicit desired outputs from LLMs.Decomposing tasks, using chain-of-thought prompting, providing few-shot demonstrations, and role-playing are identified as practical strategies.

3. The LLM4TR Framework

LLM4TR reframes transportation management around four synergistic LLM roles that connect sensing, learning, modeling, and managing. The framework replaces isolated stages with a knowledge-and-data-driven closed loop supported by reflective feedback.

  • Four roles: LLM4TR organizes transportation applications into information processors, knowledge encoders, component generators, and decision facilitators.These roles respectively process information, encode knowledge, generate system components, and support or make transportation decisions.
  • Four roles: Information processors integrate multimodal transportation information through natural-language interfaces, while knowledge encoders convert data into context-aware or structured domain knowledge.The framework links heterogeneous inputs with representations usable by downstream transportation tasks.
  • Four roles: Component generators use LLMs to automate modeling and system design by producing modular components from human-interpretable instructions.This role applies generative capabilities to transportation modeling and component construction.
  • Four roles: Decision facilitators provide principled guidance or human-like decisions after activating task-execution capabilities in transportation management.The role addresses the managing stage of the transportation cycle.
  • Closed-loop synergy: The four roles form a closed-loop cycle in which processed information informs knowledge encoding, adaptive models, management decisions, and subsequent sensing.The framework describes reflective feedback that aligns operations with evolving urban dynamics.

1. LLMs as information processors •

As information processors, LLMs contextualize and fuse heterogeneous transportation data so complex traffic information can support management and semantic understanding.

  • Function: LLMs process and fuse text, sensor data, task descriptions, and user feedback through contextual encoding, reasoning, and multimodal integration.This produces unified representations of complex traffic patterns and reduces downstream processing complexity.
  • Examples: Applications include analyzing sensory traffic data and accident reports and converting user language queries into task-specific commands.These examples cover both transportation observation and natural-language interaction.

2. LLMs as knowledge encoders •

As knowledge encoders, LLMs transform unstructured transportation knowledge into explicit rules, latent representations, or other computable forms for downstream use.

  • Function: LLMs extract and formalize transportation domain knowledge through explicit rule extraction and latent semantic embedding.This role bridges unstructured knowledge in transportation data and computable or comprehensible representations.
  • Examples: Examples include traffic-rule knowledge bases, traffic-scenario knowledge graphs, and representation vectors for subsequent computation.These outputs support traffic management, scenario formalization, and downstream modeling.

3. LLMs as component generators •

As component generators, LLMs create functional algorithms, synthetic environments, and evaluation frameworks that automate transportation-system design, testing, and refinement.

  • Function: LLMs generate functional algorithms, synthetic environments, and evaluation frameworks through instruction-followed content generation.The role uses generative capabilities to automate component design, testing, and refinement.
  • Examples: Applications include designing reinforcement-learning reward functions, synthesizing virtual driving environments, and providing model-component refinement feedback.These examples span algorithm design, environment construction, and iterative improvement.

4. LLMs as decision facilitators •

Decision facilitators use LLMs to predict traffic dynamics, optimize decisions, and simulate human-like reasoning across transportation management tasks. The taxonomy positions them at the action-oriented end of a progression from data processing to system implementation.

  • Decision facilitators predict traffic dynamics, optimize decisions, and simulate human-like reasoning for micro-level agent behaviors and macro-level system states.
  • Their applications include autonomous-driving control and planning, safety-critical action guidance, and traffic-state forecasting.
  • The taxonomy distinguishes processors, encoders, generators, and facilitators by emphasizing data transformation, knowledge formalization, content synthesis, and action prediction, respectively.
  • Processors consume raw data, encoders build domain knowledge, generators create downstream artifacts, and facilitators produce recommendations or predictions supporting operational decisions.
  • Decision facilitators attract the most existing research interest, especially in traffic prediction, autonomous driving, and traffic signal control.

4. LLMs as Information Processors

As information processors, LLMs encode context, reason over transportation data, and fuse modalities to support scenario understanding and analysis. Applications span simulation, situational awareness, safety analytics, and retrieval-augmented driving systems, but real-world robustness remains limited.

  • LLMs process heterogeneous transportation data through contextual encoding, analytical reasoning, and multimodal integration enabled by in-context learning and multimodal extensions.
  • Context encoders translate natural-language instructions into domain-specific representations and simulator configurations for transportation problem solving.
  • ChatSUMO reduces traffic-scenario creation time from 15 minutes to 30 seconds while achieving 96% simulation accuracy.
  • Hybrid architectures combine textual, numerical, and visual inputs to support customer-flow prediction and privacy-preserving pedestrian-activity analysis.
  • LLMs extract actionable patterns from unstructured crash narratives, with one approach reaching 1.0 recall and 0.93 precision for alcohol-involvement identification.
  • Current systems often depend on benchmark datasets or constrained simulations, assume clear user inputs, and align no more than two modalities.
  • Future work calls for unified fusion across geographic, traffic, social-media, and environmental sources and across spatio-temporal, visual, and textual modalities.

5. LLMs as Knowledge Encoders

LLMs act as knowledge encoders by using pretrained world knowledge to extract, structure, and apply transportation-specific insights. The survey distinguishes explicit knowledge extraction from latent knowledge representation.

  • LLMs serve as repositories of implicit and explicit world knowledge derived from pretraining on regulations, geographical semantics, and behavioral patterns.
  • The survey categorizes knowledge-encoding studies into explicit knowledge extraction for structured reasoning and latent knowledge representation through embedding spaces.

B Knowledge Encoder Knowledge Encoder

Knowledge encoders formalize transportation information as explicit structured knowledge or reusable latent embeddings. The surveyed approaches support classification, forecasting, navigation, and operational planning, while knowledge freshness remains a central limitation.

  • Explicit extraction converts unstructured or semi-structured transportation data into formal representations such as text and knowledge graphs for systematic reasoning.
  • Specialized models such as TransGPT retain transportation knowledge across single-modal and multimodal data for analysis, recommendation, and related tasks.
  • IncidentResponseGPT synthesizes regional guidelines with real-time incident data to generate tailored, actionable traffic incident plans.
  • Embedding-based encoders transform contextual knowledge into high-dimensional vectors that serve as inputs for subsequent computational processes.
  • LLM-derived embeddings improve spatiotemporal forecasting and city-wide traffic-demand estimation, with evaluations showing zero-shot transferability.
  • ALT-Pilot combines language-based landmark embeddings with LiDAR, camera, and vision-language inputs to enable open-vocabulary navigation and enhance localization.
  • Static fine-tuning and prompt-based extraction cannot keep pace with evolving regulations, infrastructure, and mobility patterns, requiring continual knowledge updates.

6. LLMs as Component Generators

LLMs act as component generators that create algorithms, simulated environments, synthetic data, and evaluations for transportation systems. These capabilities support automated design and iterative refinement, while physical fidelity and semantic completeness remain deployment constraints.

  • Algorithmic function design: LLMs translate natural-language transportation requirements into executable functions, including traffic-management and reinforcement-learning reward functions.In-context learning and pretrained language modeling support code or rule-based function design and refinement.
  • World simulation: LLM-based world models generate controllable driving images, videos, trajectories, and maps for autonomous-driving training and evaluation.These systems combine sensor data, semantic information, temporal dynamics, geometric controls, and textual prompts.
  • Data synthesis: LLMs synthesize system parameters and rare safety-critical traffic scenarios to address data scarcity in data-centric intelligent transportation systems.LLMScenario uses prompt engineering, parameter generation, and evaluation-feedback tuning for traffic scenarios.
  • System evaluation: LLMs interpret training episodes and performance records to evaluate failures and suggest refinements for autonomous-driving policies, objectives, and architectures.CRITICAL forms a feedback loop that continuously generates critical scenarios from analyses of traffic dynamics and risk metrics.
  • Summary and outlook: The component-generator role spans algorithm design, environment generation, data synthesis, and evaluation, but generated artifacts may diverge from real-world physics or omit critical semantics.The survey identifies physical fidelity and alignment with regulatory constraints as important deployment requirements.

7. LLMs as Decision Facilitators

LLMs facilitate transportation decisions through human-like reasoning, action guidance, and spatiotemporal forecasting. Their integration with transportation control and prediction methods broadens decision support, while computational, representational, and evaluation limitations remain.

  • Decision maker: LLMs support traffic-signal control by reasoning over real-time conditions, reducing waiting times, generalizing across datasets, and tolerating sensor outages.LLMLight uses GPT-3.5/4, while LightGPT is pretrained on traffic patterns and LA-Light combines LLM reasoning with reinforcement-learning outputs.
  • Decision maker: LLMs model travel behavior and generate personalized itineraries by combining semantic reasoning with structured prompts, empirical preferences, optimized points-of-interest sequences, and contextual descriptions.Persona-based few-shot learning addresses behavioral misalignment observed with zero-shot models.
  • Decision maker: In autonomous driving and mixed-autonomy traffic, LLMs act as cognitive decision makers or coordinating agents using environmental interpretation, memory, tools, and inter-vehicle communication.CoMAL uses multiple LLM agents to allocate roles and discuss strategies in real time.
  • Guidance: LLMs guide lower-level control by generating action candidates, language instructions, controller selections, or trajectory plans that connect semantic reasoning with model-predictive control.VLM-MPC assigns high-level trajectory prediction to a vision-language model and dynamic execution to MPC.
  • Spatial-temporal predictor: Spatiotemporal tokenization and traffic-specific pretraining enable LLM-based models to capture traffic-flow semantics for forecasting.TrafficBERT and the Transportation Foundation Model use self-attention or graph-based representations to model temporal and system interactions.
  • Summary and outlook: LLM decision facilitation remains constrained by numerical and geometric representation gaps, computationally intensive adaptation, and a lack of standardized transportation benchmarks.These limitations affect precise physical-space modeling, infrastructure-scale optimization, real-time decisions, and reproducible evaluation.

8. Practical Guidance

The survey provides practical guidance for deploying LLMs in transportation through datasets, literature and software resources, and hardware planning. It emphasizes language-enhanced data, parameter-efficient adaptation, and the preliminary nature of hardware estimates.

  • Resources: The practical-guidance review covers transportation datasets, literature collections, software libraries, and hardware requirements.These resources are intended to facilitate deployment and practical problem solving in transportation domains.
  • Datasets: Language-enhanced datasets couple raw traffic data with descriptions or labels to ground LLMs in transportation domains.The survey summarizes emerging datasets for LLM development in intelligent transportation and autonomous driving.
  • Literature and software: The resource review organizes influential LLM surveys and open-source libraries supporting application development, deployment, evaluation, and experimentation.Tables 5 and 6 respectively summarize surveys and widely adopted libraries.
  • Adaptation: Parameter-efficient methods such as LoRA and QLoRA make transportation-specific LLM customization more accessible on mid-range hardware.Pretraining a foundational transportation LLM may remain infeasible because of resource consumption.
  • Hardware planning: The hardware analysis summarizes approximate fine-tuning requirements and performance statistics for LLaMA models across sizes and methods.The estimates are compiled from peer-reviewed studies, community benchmarks, and industry experiments.
  • Hardware planning: Hardware statistics are rough planning guides whose values vary with training setup, precision, and quantization, so current benchmarks and vendor documentation remain necessary.The survey recommends reviewing up-to-date community resources and vendor documentation for detailed planning.

9. Discussion

The discussion identifies deployment, grounding, interpretability, and shared infrastructure as priorities for moving LLM4TR toward real-world transportation systems. It also highlights safety, computational, privacy, and ethical constraints that require mitigation.

  • Future opportunities of LLM4TR: Five future directions are proposed to address current gaps and shape LLM-driven transportation research.The discussion emphasizes opportunities for real-world deployment alongside methodological advances.
  • Future opportunities of LLM4TR: Experimental LLM transportation prototypes still face standardization, computational scalability, and compatibility challenges before real-world deployment.Open-TI combines conversational interfaces with SUMO and CityFlow to automate multistep workflows, but adaptation to mainstream industry software remains difficult.
  • Resources and computational guidance: Table 7 organizes rough fine-tuning and inference hardware estimates across LLaMA model sizes, with inference rates generally measured at batch size 1.The estimates aggregate community benchmarks and articles and may vary by configuration.
  • Future opportunities of LLM4TR: Hybrid knowledge grounding should combine live spatiotemporal data, transportation theory, and policy information to reduce rule-violating solutions.The proposed hierarchy includes sensor and V2X grounding, explicit transportation principles, and political grounding.
  • Future opportunities of LLM4TR: Future systems require formal verification, simulation-based interpretability, high-quality multimodal corpora, standardized evaluation pipelines, and open ecosystems.These measures are presented as part of the broader effort to integrate LLMs into cyber-physical-social transportation systems.
  • Challenges and concerns: LLM deployment in transportation must address bias, hallucinations, probabilistic opacity, and the need for rigorous performance guarantees.The paper identifies RAG, curated transportation fine-tuning, bias-aware objectives, and post-hoc validation as mitigation approaches.
  • Challenges and concerns: Computational demand and privacy risks constrain LLM adaptation, especially for practitioners with limited hardware and systems handling sensitive mobility data.PEFT methods, lightweight models, and cloud inference can reduce resource requirements, while location, travel, and driver-behavior data raise privacy concerns.

10. Conclusion

The conclusion presents LLM4TR as a unified framework for positioning LLMs across sensing, learning, modeling, and managing in transportation systems. It identifies a roadmap centered on safer, more adaptable deployment despite persistent challenges in safety, bias, and scalability.

  • Conclusion: LLM4TR categorizes LLMs as information processors, knowledge encoders, component generators, and decision facilitators.These roles respectively harmonize multimodal data, distill domain expertise, automate system design, and support reasoning for real-time control.
  • Conclusion: The framework describes LLMs as enablers of cyber-physical-social intelligence rather than merely natural-language-processing tools.The conclusion frames their integration as a shift toward a unified, language-driven transportation paradigm.
  • Conclusion: Future research should embed LLMs in physics-aware simulations, validate systems through industry-academia collaboration, and address ethical concerns in equitable access.The conclusion also notes unresolved safety, bias, and deployment-scaling challenges.
Loading 2503.21411v2…