1. Bibliographic Information

• Title: Large Language Models and Their Applications in Roadway Safety and Mobility Enhancement: A Comprehensive Review
• Authors: Muhammad Monjurul Karim, Yan Shi, Shucheng Zhang, Bingzhang Wang, Mehrdad Nasri, Yinhai Wang*
  • The authors are affiliated with various research institutions, with the corresponding author, Yinhai Wang, being a prominent figure in transportation engineering research, particularly associated with the University of Washington's STAR Lab (Smart Transportation Applications and Research Laboratory). This background lends significant domain expertise to the review.
• Journal/Conference: This paper is a preprint available on arXiv.
• Publication Year: The initial version was submitted in 2024 (based on the original source link format).
• Abstract: The abstract summarizes the critical need for advanced analytical frameworks in transportation to address safety and mobility challenges. It posits that Large Language Models (LLMs) represent a paradigm shift. The paper reviews how LLMs are customized—through architectural, training, prompting, and multimodal strategies—to work with transportation's unique spatio-temporal data. It systematically covers applications in mobility (e.g., traffic prediction, signal control) and safety (e.g., crash analysis, driver behavior). The review also examines enabling technologies like Vehicle-to-Everything (V2X) and edge computing. Key challenges such as hallucinations, data governance, and safety assurance are discussed, along with promising future research directions like multimodal fusion and enhanced reasoning. The paper aims to provide a structured roadmap for researchers and practitioners.
• Original Source Link: The paper is an arXiv preprint, available at https://arxiv.org/abs/2506.06301v1. This means it has not yet undergone formal peer review for a journal or conference but is shared publicly to disseminate research findings quickly. The PDF can be accessed at https://arxiv.org/pdf/2506.06301v1.pdf.

2. Executive Summary

• Background & Motivation (Why):

  • Core Problem: Traditional engineering and statistical methods are struggling to keep pace with the increasing complexity, dynamism, and data heterogeneity of modern transportation systems. Critical challenges in roadway safety (accidents) and mobility (congestion) persist, demanding more advanced analytical tools.
  • Importance & Gaps: Road accidents cause preventable deaths, and congestion leads to significant economic and environmental costs. While AI has made inroads, the recent emergence of LLMs offers unprecedented capabilities in language understanding, reasoning, and knowledge integration that have not been systematically reviewed in the specific, interconnected context of roadway safety and mobility. Existing reviews are often too broad (covering all of transportation) or too narrow (focusing only on autonomous driving or forecasting), and they frequently neglect to detail the crucial adaptation methods required to make LLMs effective in this domain.
  • Innovation: This paper provides a focused, comprehensive review specifically on LLM applications for roadway safety and mobility enhancement. Its key innovation is the in-depth analysis of how LLMs are adapted to bridge the "modality gap" between their language-native design and the spatio-temporal, numerical data of transportation. It also provides a structured synthesis of applications, enabling technologies, challenges, and future opportunities in these two distinct but related fields.

• Main Contributions / Findings (What):

  • Systematic Customization Framework: It summarizes and provides a general framework for how LLMs are customized for transportation tasks, an often-neglected aspect in other surveys.
  • Categorization of Applications: It systematically categorizes and critically analyzes a wide array of LLM application methodologies across both mobility (traffic prediction, signal control, simulation) and safety (crash analysis, driver behavior, rule formalization).
  • Identification of Trends and Challenges: It identifies emerging trends, common architectural patterns, and synthesizes domain-specific challenges, including inherent LLM limitations (e.g., hallucinations), data governance issues, and safety-critical deployment risks.
  • Roadmap for Future Research: It uncovers unexplored research opportunities and outlines promising future directions, such as advanced multimodal fusion, enhanced spatio-temporal reasoning, and human-AI collaboration, serving as a roadmap for the field.

3. Prerequisite Knowledge & Related Work

This section explains the foundational concepts necessary to understand the paper's content and situates the work within the existing literature.

• Foundational Concepts:

  • Large Language Models (LLMs): These are massive neural networks, typically based on the Transformer architecture, which was introduced in the paper "Attention Is All You Need" [26]. The key mechanism is self-attention, which allows the model to weigh the importance of different words (or tokens) in an input sequence to better understand context and long-range dependencies. LLMs like GPT [7], Llama [13], and Gemini [8] are pre-trained on vast amounts of text and code, giving them general knowledge and reasoning abilities.
  • Fine-Tuning & Domain Adaptation: After pre-training, an LLM can be adapted for a specific domain or task. Fine-tuning updates the model's parameters using a smaller, domain-specific dataset (e.g., transportation safety manuals). Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) [32] make this process more efficient by only training a small number of new parameters, reducing computational cost and the risk of "catastrophic forgetting" (where the model forgets its general knowledge).
  • Prompt Engineering: This is the art of designing the input text (the prompt) given to an LLM to guide it toward the desired output without changing the model itself. A key technique is Chain-of-Thought (CoT) prompting, where the prompt instructs the model to "think step-by-step," breaking down a complex problem into intermediate reasoning steps. This often improves accuracy on logical tasks and makes the model's reasoning process more transparent.
  • Retrieval-Augmented Generation (RAG): To combat hallucinations and provide up-to-date, domain-specific information, RAG systems connect an LLM to an external knowledge base (e.g., a database of traffic laws). When a query is received, the system first retrieves relevant information from the knowledge base, augments the original prompt with this information, and then has the LLM generate a response grounded in the provided facts.
  • LLM Agents & Tool Use: This paradigm treats the LLM as a central "agent" or orchestrator that can use external "tools." The LLM plans how to solve a problem by calling on specialized modules, such as a traffic simulator, a calculator, or another AI model. This leverages the LLM's reasoning while relying on the precision of dedicated tools.
  • Multimodal Large Language Models (MLLMs): These models can process and understand information from multiple modalities, not just text. Vision-Language Models (VLMs) are a common type, combining a vision encoder (to process images/video) with an LLM. This allows them to answer questions about visual scenes, making them highly relevant for transportation, which is rich in data from cameras and sensors.

• Previous Works: The paper critiques several existing review articles on LLMs in transportation.

  • Broad-Scope Surveys: Works by Yan et al. [19], Nie et al. [20], and Wandelt et al. [21] cover the entire transportation sector. The paper argues that while comprehensive in breadth, they lack the necessary depth on the specific, nuanced challenges of roadway safety and mobility.
  • Narrow-Scope Surveys: Reviews by Cui et al. [23] and Yang et al. [24] focus deeply on Autonomous Driving (AD), but this scope is too narrow as it omits broader safety topics (like accident analysis from reports) and network-level mobility issues (like traffic signal control). Similarly, a review by Zhang et al. [25] concentrates only on time-series forecasting, overlooking other critical mobility and safety applications.

• Differentiation: This review carves out a unique niche by:

  1. Focusing specifically on the interconnected domains of roadway safety and mobility.
  2. Placing a strong emphasis on how LLMs are adapted to handle transportation's non-textual data.
  3. Providing a general framework for this customization.
  4. Thoroughly covering the role of cross-cutting enabling technologies like edge computing and V2X.

4. Methodology (Core Technology & Implementation)

As a review paper, its methodology is the systematic framework used to survey and structure the field. A core part of this is the analysis of specialized architectures designed to bridge the "modality gap" between LLMs and transportation data.

Specialized Architectures for Transportation

The paper highlights four key strategies for adapting LLMs to the transportation domain.

1. Integration of Explicit Spatio-Temporal Modules: Standard LLMs lack the "inductive biases" to understand spatial relationships (how locations relate to each other in a road network) and temporal dynamics (how traffic evolves over time). This approach adds dedicated modules to the LLM architecture to explicitly handle this.

Fig. 1. Conceptual framework for augmenting Large Language Models (LLMs) with explicit Spatio-Temporal (S-T) modules for transportation applications.

As shown in Image 1, the process is:

• Input: Raw spatio-temporal data (e.g., speed, location, time) is fed into a Spatio-temporal processing module.
• Processing: This module uses separate encoders to create embeddings (vector representations) for spatial, temporal, and feature information. These are then fused into a Unified S-T Representation.
• Tokenization: This unified representation is converted into Enriched S-T "Tokens" that the LLM can process. Unlike word tokens, these tokens are imbued with spatial and temporal context.
• LLM Processing & Output: The LLM processes this context-aware sequence to produce transportation-specific analysis, like a traffic forecast. Examples include STGLLM-E [96], which uses a special spatial attention mechanism, and UrbanGPT [98], which uses a Temporal Convolutional Network (TCN).

2. Novel Input Representation and Tokenization Strategies: This strategy focuses on transforming raw numerical and spatio-temporal data into a format that a standard LLM can understand.

Fig. 2. Overview of input representation strategies for transforming spatiotemporal transportation data into LLM-compatible formats.

Image 2 illustrates several approaches:

• Textualization: Converting numerical data into natural language sentences. For example, a data point (sensor_id=5, speed=30mph, time=8:00, weather=rainy) could be converted to the text: "At 8:00 AM under rainy conditions, sensor 5 recorded a speed of 30 mph." This is used by xTP-LLM [35].
• Patch Tokenization: Borrowed from Vision Transformers (ViTs), this method breaks a time-series into smaller segments or "patches." Each patch is then converted into a single token, similar to how an image is broken into patches. This is used by STGLLM-E [96].
• Semantic Tokenization: Converting numerical values or patterns into descriptive text tokens like 'short up' or 'steady down', as done in TIME-LLM [61].
• Intermediate Structured Representations: The LLM acts as a planner, generating structured code (e.g., JSON) that instructs other specialized models on how to process the raw numerical data. LCTGen [99] uses this approach.

3. Partially Frozen Model Adaptation: Fully fine-tuning a massive LLM is computationally expensive and can lead to catastrophic forgetting. This approach strikes a balance by freezing most of the model's pre-trained parameters and only fine-tuning a small, selective portion.

• For example, ST-LLM [97] freezes the lower layers of a GPT-2 model (which capture general patterns) and only fine-tunes the upper attention layers to specialize them for capturing complex spatio-temporal dependencies in traffic data.
• PEFT techniques like LoRA [32] are a popular form of this. LoRA injects small, trainable "adapter" matrices into the model while keeping the large original weights frozen. The change in a weight matrix $W_0$ is approximated by a low-rank product AB: $W_{adapted} = W_0 + AB$ Here, only the small matrices $A$ and $B$ are trained, drastically reducing the number of trainable parameters.

4. LLM-as-Orchestrator Integration Frameworks: This architectural pattern uses the LLM as an intelligent coordinator that manages a suite of external, specialized tools. The LLM's role is to understand a user's request, plan the steps, call the right tools, and synthesize the results. The paper formalizes this workflow in Algorithm 1:

1. Decompose: The LLM breaks the user's request ($R_{user}$) into a set of subtasks ($S$).
2. Loop through subtasks:
   • Select Tool: For each subtask $s_i$, the LLM selects the appropriate tool $t_i$ from a tool library $\mathcal{T}$.
   • Format Parameters: The LLM prepares the necessary inputs $p_i$ for the selected tool.
   • Execute Tool: The external tool is executed, producing an intermediate result $r_i$.
3. Synthesize: The LLM gathers all intermediate results ($R_{intermediate}$) and generates a final, comprehensive response ($A_{final}$). Examples include TrafficGPT [49], which orchestrates traffic simulation and analysis models, and Open-TI [70], which calls tools for map processing and demand generation.

5. Experimental Setup

The paper is a review, so it surveys the experimental setups of the works it cites.

• Datasets: A wide variety of datasets are mentioned across different applications:

  • Autonomous Driving & Scene Understanding: nuScenes [204], KITTI [203], DAIR-V2X [92], BDD100k.
  • Trajectory Prediction: highD, ETH-UCY [138], SDD [139].
  • Crash Analysis & Safety: CrashEvent [174], government reports like MMUCC [211].
  • Mobility & Forecasting: Real-world traffic sensor data from various cities, public transit data (GTFS), and travel survey data (Swissmetro).

• Evaluation Metrics: Various metrics are used depending on the task. The most common ones mentioned are:

  • F1-Score: Used for classification tasks (e.g., crash severity, driver distraction). It measures the harmonic mean of precision and recall, providing a balanced measure of a model's accuracy, especially when classes are imbalanced.
    • Conceptual Definition: The F1-score balances the trade-off between Precision (the model's accuracy in its positive predictions) and Recall (the model's ability to find all actual positive instances). A high F1-score indicates both high precision and high recall.
    • Mathematical Formula: $$F1 = 2 \cdot \frac{\mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}}$$
    • Symbol Explanation:
      • $\mathrm{Precision} = \frac{TP}{TP + FP}$, where TP is True Positives and FP is False Positives.
      • $\mathrm{Recall} = \frac{TP}{TP + FN}$, where FN is False Negatives.
  • ROUGE (Recall-Oriented Understudy for Gisting Evaluation): Used to evaluate the quality of automatically generated text summaries or narratives (e.g., in crash report analysis).
    • Conceptual Definition: ROUGE compares a machine-generated summary to one or more human-written reference summaries. It works by counting the number of overlapping units, such as n-grams, word sequences, and word pairs. For instance, ROUGE-1 measures the overlap of unigrams (single words).
    • Mathematical Formula (for ROUGE-N): $$\text{ROUGE-N} = \frac{\sum_{S \in \{\text{RefSummaries}\}} \sum_{\text{gram}_n \in S} \text{Count}_{\text{match}}(\text{gram}_n)}{\sum_{S \in \{\text{RefSummaries}\}} \sum_{\text{gram}_n \in S} \text{Count}(\text{gram}_n)}$$
    • Symbol Explanation:
      • $n$: The length of the n-gram (e.g., 1 for unigrams).
      • $\{\text{RefSummaries}\}$: The set of human-created reference summaries.
      • $\text{Count}_{\text{match}}(\text{gram}_n)$: The number of times an n-gram in the machine-generated summary also appears in a reference summary.
      • $\text{Count}(\text{gram}_n)$: The total number of n-grams in the reference summaries.
  • Other Metrics: The paper also implicitly references metrics for forecasting (e.g., Mean Absolute Error, Root Mean Squared Error) and reinforcement learning (e.g., average delay, queue length, collision rate).

• Baselines: The reviewed papers compare LLM-based approaches against a range of traditional and deep learning models, including:

  • Time-Series Models: ARIMA, SVR [105].
  • Early Deep Learning: LSTM [106], GRU [107].
  • Spatio-Temporal Models: Graph Convolutional Networks (GCNs) [109], Temporal Convolutional Networks (TCNs) [110].
  • Reinforcement Learning (RL): Standard RL agents for tasks like traffic signal control.
  • Traditional Machine Learning: Random Forest, XGBoost.

6. Results & Analysis

The paper's core results are presented as a synthesis of findings from the reviewed literature, summarized in two main tables. The following are manual transcriptions of these tables as presented in the source text.

LLM Contributions to Mobility Enhancement

The paper summarizes how LLMs contribute to various mobility applications. The key findings are compiled in Table I.

Transcription of Table I: Qualitative Summary of LLM Contributions to Mobility Enhancement

LLM Contributions to Roadway Safety Enhancement

Similarly, the paper summarizes LLM applications focused on safety in Table II.

Transcription of Table II: Qualitative Summary of LLM Contributions to Roadway Safety Enhancement

Analysis Summary: The results synthesized across the reviewed literature consistently show that LLMs, when properly adapted, can outperform or significantly augment traditional methods. Key themes include:

• Enhanced Performance: LLMs often achieve higher accuracy in prediction and classification tasks, especially in few-shot or zero-shot settings where data is scarce.
• Interpretability: A major advantage is the ability to generate natural language explanations for their decisions (e.g., via CoT), addressing the "black-box" problem of many deep learning models.
• Data Fusion: LLMs and MLLMs excel at integrating diverse and unstructured data sources (text, images, sensor data, contextual information) into a unified analytical framework.
• Automation: They can automate laborious tasks, such as analyzing crash narratives, generating simulation scenarios from text descriptions, and creating natural language interfaces for complex tools.

7. Conclusion & Reflections

• Conclusion Summary: The paper concludes that LLMs represent a transformative technology for roadway safety and mobility. By adapting LLMs to handle transportation's unique data modalities, researchers are unlocking new capabilities in prediction, optimization, analysis, and simulation. LLMs are shown to improve efficiency in the mobility domain and provide new tools for proactive analysis in the safety domain. While the potential is immense, the authors stress that significant challenges related to LLM limitations, data governance, and safety assurance must be overcome for responsible and effective real-world deployment. The review provides a structured overview of this rapidly evolving field, acting as a guide for future innovation.

• Limitations & Future Work: The paper presents a thorough discussion of challenges and future directions in Section VII.

  • Inherent LLM Limitations:
    • Hallucination: Generating factually incorrect information.
    • Physical Grounding: Difficulty connecting symbolic knowledge to dynamic, real-world sensor data.
    • Numerical Reasoning: Weakness in precise mathematical calculations and optimization.
    • Scalability & Latency: High computational cost and slow inference times are barriers to real-time applications.
    • Consistency: Non-deterministic outputs pose reliability risks for safety-critical systems.
  • Data, Privacy, and Bias:
    • Data Scarcity: Need for large, high-quality, diverse transportation datasets.
    • Concept Drift: Models may become outdated as traffic patterns evolve.
    • Algorithmic Bias: Risk of perpetuating societal biases present in data.
    • Privacy: Handling sensitive data like trajectories and video feeds requires robust privacy-preserving techniques (e.g., Federated Learning, Differential Privacy).
  • Deployment Challenges:
    • Sim-to-Real Gap: Models trained in simulation may not perform well in the real world.
    • Interoperability: Integrating LLMs with heterogeneous legacy transportation systems is difficult.
    • Robustness: Ensuring systems are resilient to sensor failures, communication disruptions, and adversarial attacks.
  • Future Research Avenues: The authors highlight several promising areas:
    1. Advanced Multimodal Fusion: Integrating a wider range of sensor data.
    2. Native Spatio-Temporal Reasoning: Building LLM architectures that inherently understand space and time.
    3. Causal Inference: Moving beyond correlation to understand cause-and-effect relationships.
    4. Human-AI Collaboration: Designing effective shared control and decision-making systems.
    5. Continuous Learning: Developing models that can adapt to changing environments over time.
    6. Verified Explainability & Safety: Creating formal methods to verify the safety and trustworthiness of LLM-based systems.
    7. Efficient, Domain-Specific Foundation Models: Building smaller, specialized models for transportation.
    8. Hybrid AI Architectures: Combining LLMs with symbolic reasoning and optimization solvers.
    9. Edge Deployment: Optimizing models to run efficiently on vehicles and roadside infrastructure.

• Personal Insights & Critique:

  • This is an exceptionally well-structured and comprehensive review paper. Its primary strength lies in its focused scope—bridging mobility and safety—and its detailed analysis of how to adapt LLMs, which is a practical and critical contribution for researchers entering the field.
  • The categorization of applications and the summary tables (Table I and Table II) are highly effective at providing a clear, at-a-glance overview of the state of the art.
  • The paper does an excellent job of balancing optimism about the potential of LLMs with a sober assessment of their very real limitations and the ethical hurdles to deployment. The emphasis on challenges like hallucination, bias, and the sim-to-real gap is crucial for promoting responsible innovation.
  • One potential improvement could be a more quantitative meta-analysis, comparing the reported performance gains of LLM-based methods across different studies for a specific task (e.g., traffic forecasting). However, given the heterogeneity of datasets and evaluation protocols, this would be extremely challenging and is a reasonable omission for a qualitative review.
  • Overall, the paper serves as an outstanding roadmap. It not only documents what has been done but also clearly illuminates the path forward, making it an invaluable resource for graduate students, researchers, and industry practitioners looking to apply LLMs to solve real-world transportation problems.