Researchers have developed L-MARS, a novel multi-agent system designed for legal question answering that emphasizes citation faithfulness. This system incorporates agentic search and a judge-driven audit process to verify that cited sources actually support the legal rules presented in the answers. Evaluations on a bar examination dataset showed that while retrieval alone had minimal impact on accuracy, the multi-turn auditing loop significantly improved citation F1 scores and reduced instances of uncited claims. AI
IMPACT This research highlights the critical need for verifiable citations in legal AI, potentially influencing future development of trustworthy AI systems in specialized domains.
RANK_REASON This is a research paper detailing a new system and its evaluation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
