The abstract is read at the publisher; the summary is JClub's.
A moderated multi-agent LLM diagnostic team improved differential recall on curated CPC and MedCaseReasoning vignettes (top-3 +5.5 and +3.0 points) but reversed on 364 real emergency notes (top-3 −6.4 points).
- Why it matters: Benchmarking for diagnostic support is often based on curated reports, leaving uncertainty about whether multi-agent help generalizes to uncurated first-contact emergency text.
- What they did: The team compared a single direct call, five specialist personas, and the same roles as isolated blinded agents integrated by a moderator, with five repeat runs per case, judged against clinician-validated labels.
- The result: The label-level reversal on real emergency-department encounters occurred without a clinician-judged safety difference (68% ties), implying input curation drives direction while safety impact was not detected.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.