L-MARS: Legal Multi-Agent System with Agentic Search and Citation-Faithfulness Audit

arXiv cs.CL · 6 hours ago

Researchers developed L-MARS, a multi-agent legal question-answering system that audits whether cited sources actually support the claims made in answers, finding that a judge-driven evidence verification loop improves citation faithfulness from an F1 score of 0.13 to 0.25 on a 100-question Bar Exam audit. The system uses a six-class taxonomy to classify claims and cross-provider judging where different AI model families verify each other's work. The work demonstrates that standard retrieval-augmented generation achieves citation F1 scores near 0.75 but still leaves significant room for improvement in ensuring answers are grounded in verifiable sources.