Automating the loop in traffic incident management on highway
Matteo Cercola, Nicola Gatti, Pedro Huertas Leyva, Benedetto Carambia, Simone Formentin · 2025
Read Automating the loop in traffic incident management on highway by Matteo Cercola, Nicola Gatti, Pedro Huertas Leyva, Benedetto Carambia, Simone Formentin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-12085.txt.
Introduction
Matteo Cercola and co-authors ask whether large language models can help run highway incident response. Radio room operators now carry that load, making fast, high-stakes calls under pressure. The team builds decision support that joins language models with older, steadier tools.
Their 2025 paper tests two designs on past event data from the Italian operator Autostrade per l'Italia. One design pairs the language model with an optimizer, so free text meets hard constraints. The other lets the language model decide on its own. Both showed promise, but the hybrid proved more reliable, a trait that matters most where errors cost lives. The work points to a guarded path for language models in control rooms.
Catalog record
| CREATOR | Matteo Cercola, Nicola Gatti, Pedro Huertas Leyva, Benedetto Carambia, Simone Formentin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-12085
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Automating the loop in traffic incident management on highway Matteo Cercola, Nicola Gatti, Pedro Huertas Leyva, Benedetto Carambia, Simone Formentin Abstract Effective traffic incident management is essential for ensuring safety, minimizing congestion, and reducing response times in emergency situations. Traditional highway incident management relies heavily on radio room operators, who must make rapid, informed decisions in high-stakes environments. This paper proposes an innovative solution to support and enhance these decisions by integrating Large Language Models (LLMs) into a decision-support system for traffic incident management. We introduce two approaches: (1) an LLM + Optimization hybrid that leverages both the flexibility of natural language interaction and the robustness of optimization techniques, and (2) a Full LLM approach that autonomously generates decisions using only LLM capabilities. We tested our solutions using historical event data from Autostrade per l’Italia. Experimental results indicate that while both approaches show promise, the LLM + Optimization solution demonstrates superior reliability, making it particularly suited to critical applications where consistency and accuracy are paramount. This research highlights the potential for LLMs to transform highway incident management by enabling accessible, data-driven decision-making support. keywords Artificial Intelligence, Large Language Model, Decision Support System, Highway Management, Optimization. 1 Introduction Large Language Models (LLMs) are trained on vast datasets of text, learning to predict the next word in a sequence, which enables them to generate coherent and contextually relevant text. Their power lies in their ability to understand and generate human-like language, handle a wide range of tasks, and adapt to various contexts. LLMs excel at tasks that require natural language processing, including answering questions, summarizing information, and even performing basic reasoning by leveraging patterns learned during training. Recent studies ( Wei et al. 2022 ; Rao et al. 2023 ; Sha et al. 2023 ; Bian et al. 2024 ; Chung et al. 2024 ; Fu et al. 2024 ) have demonstrated that LLMs exhibit impressive reasoning capabilities. While their training is primarily based on language data, these models have shown the ability to make inferences, connect facts, and simulate logical reasoning to some extent. Recent studies demonstrate the potential of large language models to tackle complex language-based tasks such as planning ( Liu et al. 2023 ; Ahn et al. 2022 ; Singh et al. 2023 ), generating code ( Gur et al. 2024 ; Finnie-Ansley et al. 2022 ; Li et al. 2022 ), and optimization ( Pryzant et al. 2023 ; Liu et al. 2024 ; Nie et al. 2024 ; Guo et al. 2024 ; Yao et al. 2024 ). Building on this progress, our research investigates the application of LLMs for decision-making in critical scenarios. Specifically, we propose two alternative approaches that integrate natural language interaction into decision-support systems: • LLM + Optimization: This hybrid approach combines the flexibility of LLMs with the robustness of classical optimization techniques, ensuring solution reliability while enabling a natural language interface. • Full LLM: In this approach, decisions are generated exclusively by LLMs, offering a fully language-based solution without relying on external algorithms. We have tested our proposed solutions on the problem of managing highway events, where decisions must be both efficient and reliable. By employing GenerativeAI these models can enable users to engage with the systems through simple, conversational exchanges, thereby facilitating the interaction. Some works on this topic are Grigorev et al. 2024 and Goecks and Waytowich 2023 , where they use LLMs to generate response plans based on accident descriptions by learning from a knowledge base of incident response guidelines. A key advantage of our approach is the integration of real historical data, which includes detailed event descriptions and management records provided by Autostrade per l’Italia. This data-driven approach enables our systems to learn directly from historical patterns and responses, fostering decision-making grounded in actual operational experience. We evaluated both the Full LLM and LLM + Optimization solutions in the context of traffic incident management, and both showed promising results. However, given the critical nature of this application, we also assessed the reliability of each approach. Our findings indicate that the LLM + Optimization solution, by leveraging optimization techniques, offers more reliable outcomes, making it better suited for high-stakes environments where consistency and accuracy are crucial. 2 Problem state and case description Effective management of highway events, such as traffic accidents, breakdowns, or risky weather conditions, is crucial to maintain traveler safety, minimize congestion, and ensure smooth traffic flow. The ability to quickly and accurately respond to these incidents can significantly speed up the process of assisting injured people and reducing the risk of secondary accidents. Currently, this critical responsibility falls on the shoulders of radio room operators, who are tasked with managing highway events in real-time. Radio room operators are trained with lengthy manual procedures that are often not fully applicable in real-world scenarios. The nature of their work demands high levels of attention and quick decision-making, especially during long shifts that are mentally and physically taxing. These operators must assess a wide range of variables, from traffic patterns to weather conditions, all while coordinating responses across multiple stakeholders, including emergency services, road maintenance teams, and police forces. Furthermore, the decisions they make are influenced by subjective factors, such as their individual experience and interpretation of unfolding events, which can lead to inconsistencies in how similar incidents are handled. Innovative solutions are urgently needed to support operators in making rapid, consistent, and well-informed decisions, particularly in high-stakes environments where errors can lead to severe consequences. Given the critical and overwhelming job of the operators, such solutions should facilitate decision-making through natural and intuitive interaction, enhancing reliability without overburdening the user. 3 Methodologies The following are the methodologies we developed to tackle the problem: • LLM + Optimization Block : Combines the flexibility of large language models with optimization algorithms, ensuring that decisions are both contextually sound and operationally robust. • Full LLM : Relies exclusively on LLMs to make decisions, for the purpose of testing them in decision problems. Both methodologies take a natural language event description as input (e.g., type of event, number of vehicles involved, location, time, etc.) and produce a natural language output detailing the sequence of decisions to take (e.g., calling the police, closing a lane, etc.). 3.1 LLM + Optimization Block Figure 1: LLM + Optimization Block pipeline. The proposed LLM + decision block solution combines the strengths of LLMs and optimization algorithms to establish a robust framework for managing highway events, building upon the approach presented in Liu et al. 2023 . LLMs excel at processing unstructured inputs and facilitating natural language interactions, while optimization algorithms provide mathematically grounded decision-making guarantees. By merging these two approaches, we aim to develop a system that is both user-friendly and theoretically reliable. The solution, visible in Figure 1 , includes two translation blocks (LLMs) at the beginning and end of the pipeline and a decision block in between. This structure enables radio room operators to engage with the optimization algorithm through natural, conversational interactions, thus reducing the technical complexity of the interface. In this system, the two LLM components act as translators: the first converts natural language inputs into a structured format that can be processed by the optimization algorithm and the second retranslates