Researchers propose multi-agent system to audit legal AI citations

A preprint on arXiv introduces L-MARS, an open multi-agent system for legal question answering that adds a judge-driven evidence loop to verify each citation. The system retrieves relevant sources, drafts an answer, then runs a cross-provider verification pass where a separate model family checks every atomic claim against its cited source. On a stratified 100-question Bar Exam audit, the multi-turn judge loop raised strict citation F1 from 0.13 in a naive retrieval-augmented baseline to 0.25 and cut the no-citation rate from 34% to 13%. A post-draft step called Faith-Search further reduced unreachable citations below 1% but did not improve F1 beyond the judge loop. The authors also tested a 50-question LegalSearchQA case study, finding retrieve-then-draft pipelines saturated near 0.75 citation F1 while a single-agent web-search baseline collapsed to 0.22 under external audit. The work is posted as arXiv:2509.00761v4, updated in July 2026.