L-MARS is a system that reduces confusion and uncertainty in legal question answering through multi-agent reasoning and retrieval. Unlike single-pass Augmented Search Generation (RAG), L-MARS decomposes questions into subproblems, performs targeted searches across disparate sources (Serper web, local RAG, CourtListener case law), and utilizes judge agents to validate sufficiency, jurisdiction, and temporal validity before synthesizing answers. This iterative inference-retrieval-verification loop ensures consistency, filters out noisy evidence, and grounds answers in authoritative law. We evaluated L-MARS on LegalSearchQA, a new benchmark comprised of 200 state-of-the-art multiple-choice legal questions from 2025. The results demonstrate that L-MARS significantly improves factual accuracy, reduces uncertainty, and achieves higher preference scores for both human experts and LLM-based judges. This study demonstrates that multi-agent reasoning via agent search provides a scalable and reproducible blueprint for deploying LLM in high-stakes areas requiring accurate legal search and deliberation.
