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

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.

💡 Research Summary

The paper introduces TransportAgents, a hybrid multi‑agent framework that leverages large language models (LLMs) together with a supervised multilayer perceptron (MLP) to predict traffic‑accident severity. The authors argue that single‑agent LLMs, despite their impressive reasoning abilities, struggle with heterogeneous crash data that combines structured variables (e.g., driver age, road type) and unstructured narrative descriptions. Moreover, single‑agent pipelines often exhibit bias, instability, and limited interpretability when applied to safety‑critical domains.

Framework Overview
TransportAgents consists of three sequential components:

1. Data Pre‑processing Team – Raw records are first stripped of the target label and split into structured and textual parts. A Feature Selection Agent performs semantic annotation and filters out irrelevant fields, reducing hallucination risk. The remaining attributes are then passed to a Conceptual Category Organizer, which groups them into high‑level concepts such as Victim Demographics, Transportation & Environmental Context, and Incident Details.

2. Severity Assessment Team – For each conceptual group a dedicated Evaluation Agent (implemented with an LLM) receives only the information belonging to its domain. Prompt engineering, including few‑shot examples and Chain‑of‑Thought (CoT) reasoning, guides the agent to produce a category‑specific severity score (s_k). This parallelized design ensures that each agent can focus on a narrow slice of the data, mitigating the tendency of a monolithic LLM to overlook or over‑emphasize certain cues.

3. Integration Manager Module – The vector of intermediate scores (\mathbf{s} =