LLM-Assisted Traffic Management Plan Generation: An Empirical Study of WisDOT's WisTMP System

The Pain Points of Work Zone Traffic Management

Construction work zones are among the most critical and hazardous segments of transportation infrastructure. To maintain both road safety and traffic flow during construction, engineers must develop detailed Transportation Management Plans (TMPs). Yet TMP preparation has long been a labor-intensive process that relies heavily on practitioner experience — time-consuming, costly, and inconsistent in quality.

A recent study published on arXiv, based on the Wisconsin Department of Transportation's (WisDOT) WisTMP system, proposes a compelling approach: using large language models (LLMs) to automate TMP content generation, freeing planners from tedious documentation work.

A TMP is the core document required for obtaining construction permits and coordinating multi-party work activities. It typically covers temporary traffic control device layouts, detour route planning, construction phasing, pedestrian and cyclist protection measures, and emergency response plans. State Departments of Transportation (DOTs) enforce strict format and content standards for TMPs, requiring planners to be well-versed in technical references such as FHWA's Manual on Uniform Traffic Control Devices (MUTCD). In practice, a moderately complex TMP can take days to weeks to prepare manually, and quality varies widely depending on the preparer's familiarity with the local road network, construction type, and past projects. WisDOT's WisTMP system is an online assisted-preparation platform designed to lower the barrier through standardized workflows — and this research takes it a step further by integrating LLMs to increase automation.

Local Deployment and Data Security as Design Priorities

The research team built an LLM-assisted framework with two notable design considerations.

The first is model diversity. The framework fine-tuned multiple open-source LLMs across different parameter scales (from 7B/8B to 14B) to compare how model size affects task performance.

The second is local deployment. Because traffic management documents often involve government projects and sensitive engineering data, the framework runs models locally rather than calling cloud-based APIs, ensuring data security. This tradeoff carries real-world significance for government and public-sector adoption of AI tools — regulatory compliance often matters more than raw model performance when determining whether a technology can actually be deployed.

Turning Historical Documents into Training Data

Adapting a general-purpose LLM to the specialized task of TMP preparation requires domain-specific data. The research team built a dedicated domain dataset from historical WisTMP documents.

The approach involved converting raw PDF files into structured question-answer pairs stored in JSON format for supervised fine-tuning. This pipeline — transforming unstructured professional documents into Q&A training corpora — also serves as a reusable paradigm for LLM deployment in other vertical domains.

Supervised Fine-Tuning (SFT) is currently one of the most mainstream methods for adapting general pretrained LLMs to specialized domains. The core logic is that general-purpose models already possess language understanding and generation capabilities, but lack domain-specific terminology, preferred document formats, and business logic. By continuing training on domain-specific Q&A pairs or instruction-response data, the model learns "what outputs conform to domain standards." Converting PDFs into structured Q&A pairs is itself quite challenging: it requires identifying document structure, extracting key fields, and designing appropriate question forms that will correctly trigger the model during inference. The quality of this data engineering step directly determines the ceiling of fine-tuning performance, and is typically one of the most labor-intensive stages in vertical-domain LLM projects.

Experimental Findings: Fine-Tuning Works, but Scale Returns Are Limited

The experimental results yielded several clear conclusions.

Fine-Tuning Significantly Improves Generation Quality

On standard text generation metrics, domain-fine-tuned models significantly outperformed their non-fine-tuned counterparts. This confirms that in professional document generation scenarios, domain adaptation is the critical step to making LLMs useful — rather than simply relying on a base model's general capabilities.

Bigger Is Not Always Better

One particularly illuminating finding is that scaling from 7B/8B to 14B yielded only marginal performance gains. This contrasts with the intuitive assumption that larger models are always better, suggesting that blindly adding parameters is not cost-effective for certain structured document generation tasks. For government and engineering organizations that require local deployment under constrained compute budgets, mid-sized models may offer better value for money.

A model's parameter count (in billions, or B) is often used as a rough proxy for capability, but this rule breaks down on specific tasks. Researchers partially attribute this to "task-model fit": for structured document tasks with fixed formats and limited domain vocabulary, small models that are thoroughly fine-tuned can often approach or even surpass large models that haven't been fine-tuned. Larger models' advantages are more apparent in open-ended tasks requiring broad knowledge or complex reasoning chains. Furthermore, in local deployment scenarios, a 14B parameter model demands significantly more VRAM and inference hardware than a 7B/8B model (typically requiring 24GB+ VRAM), and if performance gains are marginal, this additional hardware cost is difficult to justify under government budget constraints. The study's findings therefore have direct decision-making value for resource-constrained institutions choosing model sizes.

Models Still Have Clear Weaknesses

Through detailed section-wise and strategy-level analysis, the research team identified several notable weaknesses in current LLMs:

• Over-generation of strategies: Models tend to enumerate too many management strategies without exercising selectivity.
• Difficulty providing project-specific justifications: Generated content leans toward generic templates and fails to produce convincing rationales for specific projects.
• Inaccurate cost estimation: Sections involving precise figures and cost calculations show notably poor reliability.

These weaknesses cluster precisely around areas requiring professional judgment and quantitative reasoning, reinforcing the gap that still exists between LLMs' ability to "generate fluent text" and their ability to "make accurate, professional decisions."

Implications for Industry Deployment

The value of this research lies not only in demonstrating that LLMs can improve TMP preparation efficiency, but also in its honest delineation of technical boundaries. LLMs are well-suited to handle repetitive work such as drafting initial versions and filling in templates, but high-stakes tasks like project justification and cost accounting still require engineering review and oversight.

In other words, the more realistic positioning at this stage is "AI-assisted" rather than "AI-replaced" — let the model do what it does well in content generation, and leave professional judgment to humans. The research team plans to release source code and demo videos (project page: zihaosheng.github.io/TMP-LLM/), which will serve as a practical resource for subsequent research at the intersection of transportation engineering and AI.

Summary

Starting from the relatively niche vertical domain of work zone traffic management, this work successfully navigates the full technical pipeline from "historical documents → structured dataset → open-source model fine-tuning → local deployment," and uses rigorous experimentation to define the current capabilities and limitations of LLMs. It serves both as a useful example of AI empowering infrastructure management and as a reminder that in vertical-domain LLM deployment, data security, model scale tradeoffs, and human-AI task division often deserve more careful consideration than the model itself.