L-MARS: multi-agent search with a citation audit

The paper L-MARS: Legal Multi-Agent System with Agentic Search and Citation-Faithfulness Audit distributes a legal question among agents for planning, search, source filtering, verification and answer synthesis. For evaluation, LegalSearchQA was created: 50 questions across five areas of law whose answers depend on events after the models’ training cutoff. The system reached 96% accuracy versus 58% without search, while an ordinary chain of thought gave 30%: extended internal reasoning without reliable search made the result worse. The dataset is small, and the answers require updating as the law changes. Two conclusions matter for Lexometrica: compare agentic search not only with a plain query but also with chain of thought and single-pass RAG, and version the currency date of each case and of the legislation corpus.