arXiv paper shows multi-agent AI beats single LLMs in legal reasoning

A new arXiv paper by Cor Steging, Ludi van Leeuwen, and Tadeusz Zbiegień explores multi-agent deliberation (MAD) for legal reasoning using LLMs. While agentic AI is gaining traction in law, multi-agent systems remain largely unexplored. The researchers propose two novel frameworks inspired by courtroom procedures and legal argumentation, enabling multiple LLM agents to deliberate and arrive at answers collaboratively. Experiments on legal and non-legal benchmarks revealed that multi-agent frameworks achieve comparable overall accuracy to baseline single LLMs, but with a critical difference: they generate significantly distinct answers. In many cases, the multi-agent approach successfully solved problems that the monolithic baseline could not, and vice versa, indicating complementary strengths.

A qualitative evaluation highlighted scenarios where multi-agent frameworks outperform single models—particularly for questions that demand critical thinking from multiple perspectives. The findings position multi-agent systems as a promising direction for AI in the legal domain, with potential to improve access to justice by handling nuanced legal reasoning. The paper is accepted for presentation at the AIDA2J Workshop during the 21st International Conference of AI & Law in Singapore (June 2026).