Abstract
Highway incident management is a safety-critical domain where decision-making speed and accuracy directly impact public safety. This paper evaluates two distinct approaches to automating the decision loop: a generative Large Language Model (LLM) system and a hybrid neuro-symbolic system combining LLMs with classical optimization. Using real-world data from Italy's primary highway operator, we demonstrate that while LLMs are superior interfaces for parsing unstructured incident reports, grounding their outputs in a Markov Decision Process (MDP) is essential for maintaining the consistency and reliability required in emergency environments.
Introduction
In the "radio room" of a highway operations center, operators are tasked with managing chaotic, high-stakes situations. When an incident occurs, they must synthesize fragmented information—weather conditions, vehicle types, and crash severity—while consulting extensive procedural manuals to deploy resources like ambulances, police, or lane closures.
This manual process suffers from two primary bottlenecks: cognitive overload during peak traffic periods and subjective inconsistency between different operators. To address these challenges, this research explores "automating the loop"—creating a system capable of transforming natural language incident descriptions into standardized, optimal action sequences.
Methodology
The study compares two architectural philosophies for decision support: a pure generative approach and a hybrid pipeline.
1. The Neuro-Symbolic Hybrid (LLM + Optimization)
This architecture treats the LLM as a translator rather than a decision-maker. The pipeline consists of three stages:
- Translation: An LLM parses unstructured incident reports into a structured data format.
- Optimization (The Brain): The core logic is handled by a Stochastic Markov Decision Process (MDP). The system models the highway environment as a graph where nodes represent states and edges represent actions. To find the most efficient response, the system employs the Improved Prioritized Sweeping (IPS) algorithm—an extension of Dijkstra’s algorithm designed for stochastic environments—to calculate the path with the lowest expected resolution time.
- Output Generation: A secondary LLM translates the mathematical optimal actions back into human-readable instructions for the operator.
2. The Generative Baseline (Full LLM)
This model utilizes GPT-4o mini as the primary agent. To bridge the gap between general knowledge and domain-specific requirements, the authors implemented several advanced prompting techniques:
- Retrieval-Augmented Generation (RAG): The model accesses a 600-page highway procedure manual and a database of historical reports via semantic chunking.
- Chain of Thought (CoT): The model is prompted to generate step-by-step reasoning before arriving at a final recommendation to improve logical flow.
Experimental Setup and Data
The models were evaluated using a proprietary dataset from Autostrade per l’Italia, the primary Italian highway operator. The dataset includes thousands of historical incident reports, detailing the specific characteristics of each event, the actions taken by human operators, and the eventual resolution times. The data was split into an 80% training/reference set (used to build the MDP graph and RAG database) and a 20% test set of unseen incidents.
Key Findings
The Consistency Gap
The most significant finding of the study is the disparity in reliability between the two models. The LLM + Optimization hybrid achieved nearly 100% consistency across all test cases. Because the MDP relies on mathematical optimization over a fixed graph, it provides stable recommendations regardless of how an incident is phrased.
In contrast, the Full LLM approach exhibited high sensitivity to "minimal perturbations." Slight changes in the wording of an incident description caused the model's consistency to drop to between 30% and 50% for certain incident types. This highlights the inherent risk of hallucinations and variability in purely generative models for safety-critical tasks.
LLMs as Interfaces, Not Deciders
The research suggests a paradigm shift in how AI is applied to infrastructure management. LLMs are highly effective as the "ears and mouth" of a system—parsing messy human language and explaining complex results. however, they lack the mathematical guarantees required to serve as the "brain." By delegating the actual decision logic to a symbolic optimizer (the MDP), the system gains the benefits of natural language processing without sacrificing safety.
Limitations and Future Directions
While the hybrid model is superior in reliability, it faces challenges regarding scalability. As the variety of highway scenarios increases, the MDP graph can become computationally expensive to navigate. Future work will explore clustering techniques to simplify state representation.
Additionally, the authors aim to improve the "explainability" of the system. While the hybrid model is consistent, providing operators with the specific historical precedents or manual sections that justify a recommendation will be crucial for building trust in automated decision-support tools.
Summary updated