Answer legal questions with checked sources and citations
Researchers built L-MARS, a team of agents that searches the web and case law, then checks its evidence; they report more accurate legal answers than leading chatbots.
Done with Custom agent
What they did
Researchers Ziqi Wang and Boqin Yuan built L-MARS, a team of agents for legal questions. One agent splits a question into smaller ones. Searches then run across the web, a local legal database and the CourtListener court-case archive. A checking agent asks whether the evidence is enough, from the right place and still current. A final agent writes the answer. They tested it on 200 new multiple-choice questions about 2025 law.
How it went
The authors report much better factual accuracy and less uncertainty than standard GPT, Claude and Gemini models. Both human reviewers and AI reviewers preferred its answers.
Worth knowing
It has two modes: a fast one-pass mode and a slower mode that keeps searching until the evidence is enough. The code is publicly available on GitHub.
Try it yourself
Answer this legal question: [your question], for [jurisdiction]. First, break it into sub-questions and show them to me. Then search official and reputable sources, including statutes and case law, for each sub-question. Before answering, check that each source applies to [jurisdiction], is current as of [date], and actually supports the point. If evidence is missing, refine your search and try again. Give a final answer with a citation for every legal claim and note anything you could not confirm.
Swap the highlighted [parts] for your own details.
Source: arXiv · Ziqi Wang, Boqin Yuan · Aug 31, 2025
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
