Enhancing Transparency and Compliance in Automated Decision-Making: A Multi-Agent System Approach Using Language Models

The emergence of large language models has significantly advanced the feasibility of automated problem-solving using agents. However, despite promising results, these systems often function as "black boxes", raising concerns about their ability to comply with requirements due to opaque decision-making processes. To mitigate these issues, we introduce a multi-agent system powered by language models. This system segments the decisionmaking process into three agent-driven stages: proposing queries, identifying norms, and retrieving facts, while delegating final judgment to a logical reasoner. We evaluated our system in simulated driving scenarios governed by a limited set of traffic regulations. Results indicate that our approach markedly enhances compliance with decision-making accuracy and offers a more interpretable and traceable method compared to methods that rely solely on language models.