TransportAgents: a multi-agents LLM framework for traffic accident severity prediction

TL;DR

TransportAgent tackles crash-severity prediction from heterogeneous, narrative-rich data by decomposing reasoning into category-specific LLM agents whose outputs are fused by a supervised MLP. This hybrid architecture blends domain-focused interpretability with data-driven fusion, achieving calibrated and robust predictions across CPSRMS and NEISS datasets and across backbone models including GPT-3.5, GPT-4o, and LLaMA-3.3. The core contributions include a data-processing and category-organization pipeline, dedicated category-aware severity agents, and a learned fusion mechanism, all validated through extensive experiments, distributional analyses, and robustness checks. The results demonstrate improved accuracy, balanced predictions across severity levels, and strong generalization, offering a generalizable blueprint for safe decision support in complex, safety-critical domains.

Abstract

Accurate prediction of traffic crash severity is critical for improving emergency response and public safety planning. Although recent large language models (LLMs) exhibit strong reasoning capabilities, their single-agent architectures often struggle with heterogeneous, domain-specific crash data and tend to generate biased or unstable predictions. To address these limitations, this paper proposes TransportAgents, a hybrid multi-agent framework that integrates category-specific LLM reasoning with a multilayer perceptron (MLP) integration module. Each specialized agent focuses on a particular subset of traffic information, such as demographics, environmental context, or incident details, to produce intermediate severity assessments that are subsequently fused into a unified prediction. Extensive experiments on two complementary U.S. datasets, the Consumer Product Safety Risk Management System (CPSRMS) and the National Electronic Injury Surveillance System (NEISS), demonstrate that TransportAgents consistently outperforms both traditional machine learning and advanced LLM-based baselines. Across three representative backbones, including closed-source models such as GPT-3.5 and GPT-4o, as well as open-source models such as LLaMA-3.3, the framework exhibits strong robustness, scalability, and cross-dataset generalizability. A supplementary distributional analysis further shows that TransportAgents produces more balanced and well-calibrated severity predictions than standard single-agent LLM approaches, highlighting its interpretability and reliability for safety-critical decision support applications.

Figures (9)

• Figure 1: Architecture of the proposed Transport Agent framework. The Data Preprocessing Team selects and organizes relevant features, the Severity Assessment Team produces category-specific evaluations, and the MLP module integrates these into a final severity level score.
• Figure 2: Illustrative roles of Transport Agent components. (Left) Feature Selection Agent determines whether a variable is relevant to severity (e.g., Primary Body Part$\rightarrow$ relevance = yes). Severity Prediction Task Linker checks whether a conceptual category is pertinent to a given variable (e.g., number of transportation modes$\rightarrow$transportation details). (Right) Conceptual Category Organizer maps selected variables to categories (e.g., Age$\rightarrow$victim demographics). Category Severity Evaluator reasons over information within one category to output a category-specific severity score.
• Figure 3: Correlation of variables within category and between severity level in the CPSRMS dataset.
• Figure 4: Correlation of variables within category and between severity level in the NEISS dataset.
• Figure 5: Comparison of true and predicted severity level distributions on the CPSRMS dataset using the GPT-3.5-turbo backbone. Results are shown for Vanilla LLM (top-left), CoT prompting (top-right), MLP (bottom-left), and the proposed TransportAgent framework (bottom-right). Baseline methods exhibit clear distributional bias, particularly under-representation of higher severity levels, whereas TransportAgent produces predictions more closely aligned with the true severity distribution across all levels.