TLDR: MARBLE is a novel multi-agent AI system designed to accurately predict traffic accident severity. It decomposes the complex prediction task across specialized agents, each focusing on specific feature subsets (e.g., environmental, temporal), and utilizes efficient Small Language Models (SLMs). By coordinating these agents through a robust rule-based mechanism, MARBLE achieves nearly 90% accuracy on real-world datasets, significantly outperforming traditional machine learning and advanced prompting methods, especially in scenarios with incomplete or imbalanced data. The framework offers enhanced interpretability and is highly resilient to varying class distributions, making it a promising solution for safety-critical applications.
Traffic accidents are a major concern globally, leading to significant societal costs and demanding effective strategies for prevention and emergency response. A crucial aspect of enhancing road safety is accurately predicting the severity of accidents, which involves forecasting injury outcomes from minor incidents to fatalities. This task is inherently complex due to incomplete data, strong dependencies between various factors, and a severe imbalance in data where high-severity cases are rare and difficult to detect.
Traditional methods, including classical machine learning and deep learning models, often struggle with these real-world complexities. They may fail to capture intricate interactions between diverse features, overfit on common accident types, or lack transparency, making it difficult to understand their decisions in safety-critical situations. Even recent advancements in large language models (LLMs) and agentic AI face hurdles when applied directly to structured prediction tasks like accident severity classification. Full-scale LLMs are computationally expensive, and using a single, general-purpose LLM for high-dimensional tabular data can lead to ‘prompt saturation,’ where too many features overwhelm the reasoning process, reducing accuracy and obscuring causal factors.
Introducing MARBLE: A Multi-Agent Approach
To address these challenges, researchers Kaleem Ullah Qasim and Jiashu Zhang have proposed MARBLE: a Multi-Agent Rule-Based LLM Reasoning Engine for accident severity prediction. MARBLE introduces a novel framework that leverages the power of agentic AI by distributing the prediction task across a team of specialized reasoning agents. This modular design allows each agent to focus on a specific semantic subset of features, such as spatial, environmental, or temporal factors, preventing context overload and enabling more focused reasoning.
The MARBLE system incorporates an interchangeable machine learning (ML) agent that provides a baseline prediction, alongside multiple domain-specific agents powered by computationally efficient Small Language Models (SLMs). The use of SLMs ensures that MARBLE retains the powerful inference capabilities of transformer models while being practical for real-time deployment. Predictions from this diverse ensemble of agents are then coordinated by a central mechanism, which can be either an interpretable rule-based logic or an LLM-guided reflection process. This structured coordination accounts for the rarity of certain accident classes and the confidence levels of individual agents.
How MARBLE Works
MARBLE’s architecture is an ensemble of heterogeneous agents. The primary ML agent processes the complete set of accident features, while specialized SLM agents receive only a relevant subset of features tailored to their expertise. For instance, an environmental agent would analyze weather conditions, temperature, and visibility, while a temporal agent would look at the day of the week, time of day, and month. This strategic distribution of features provides each SLM agent with a focused context, reduces computational overhead, and improves system robustness by isolating agents from irrelevant information.
Each agent produces a prediction, a calibrated confidence score, and a reasoning trace. These individual outputs are then synthesized by a coordinator. The research explored two coordination strategies: a deterministic rule-based approach and an LLM-based approach. The rule-based coordination, which uses weighted voting and specific override logic for high-confidence ML predictions, proved to be significantly more effective in practice.
Also Read:
- AI’s New Frontier: Detecting Road Crashes with Language Models
- Natural Language: The Key to Smarter, Safer Autonomous Vehicle Communication
Remarkable Performance and Robustness
MARBLE has demonstrated exceptional performance on both UK and US accident datasets. It consistently outperforms traditional machine learning classifiers and state-of-the-art prompt-based reasoning methods, including Chain-of-Thought (CoT) and Tree-of-Thought (ToT). While other methods often plateau below 48% accuracy, MARBLE achieves nearly 90% accuracy. This significant improvement redefines the practical ceiling for accident severity classification, especially under real-world conditions characterized by noise and extreme class imbalance.
An ablation study, where individual agents were systematically removed, highlighted the critical contributions of each component. The environmental agent, ML agent, and the central coordinator were found to be particularly pivotal, underscoring that MARBLE’s high predictive performance stems from the coordinated interplay between its specialized reasoning components. Furthermore, MARBLE proved to be highly robust to class imbalance, maintaining high accuracy even when rare severity classes dominated the data distribution. This resilience is a key advantage for real-world safety applications where severe accidents are infrequent but critical to predict.
In conclusion, MARBLE represents a significant step forward in accident severity prediction. By combining structured ML models with a distributed network of domain-specialized SLM agents and a robust coordination mechanism, it offers a generalizable and interpretable framework for reasoning under uncertainty in safety-critical applications. For more details, you can refer to the full research paper: MARBLE: A Multi-Agent Rule-Based LLM Reasoning Engine for Accident Severity Prediction.
