As large language models are increasingly deployed for legal question answering, evaluations typically focus on multiple-choice accuracy. However, this measure overlooks a common failure: whether the citation source attached to an answer exists and supports the rule attributed by the system. We present L-MARS, an open multi-agent legal QA system with agentic search and judge-driven evidence checks, auditing it claim by claim against its cited source.
Each atomic claim is labeled with a six-class taxonomy and scored with strict-ALCE under cross-provider judging. In a stratified audit of 100 Bar Exam questions, retrieval barely improves accuracy, yet the multi-turn judge loop lifts strict citation F1 from 0.13 (naive RAG) to 0.25 and reduces the no-citation rate from 34% to 13%.
We further introduce Faith-Search, a post-draft step that re-verifies and repairs unreachable citations, dropping the unreachable rate below 1% but not improving F1 over the multi-turn loop; thus, we report it as a targeted reachability intervention rather than a faithfulness breakthrough.
A 50-question LegalSearchQA case study confirms the trend: retrieve-then-draft pipelines saturate near 0.75 citation F1, while a single-agent web-search baseline collapses to 0.22 under external audit.
Blogger's Review: The design of L-MARS effectively enhances citation accuracy in legal QA systems, particularly through the integration of multi-turn auditing and agentic search. Despite Faith-Search performing well in reachability, it does not significantly boost overall F1, highlighting the complexity and challenges of citation. The multi-agent architecture offers a fresh perspective for automation in legal question answering.