Perspective independence, more than personas, drives LLM teams - and where they reverse

@misc{feng2026perspective,
  title = {{Perspective independence, more than personas, drives LLM teams - and where they reverse}},
  author = {J. Feng and Y. Jiao and Y. Li and L. Xie and W. Peng and X. Sun},
  howpublished = {medRxiv},
  year = {2026},
  month = sep,
  doi = {10.64898/2026.09.24.26363897},
  url = {https://www.medrxiv.org/content/10.64898/2026.09.24.26363897v1},
  note = {CC BY},
}

TY  - UNPB
TI  - Perspective independence, more than personas, drives LLM teams - and where they reverse
AU  - J. Feng
AU  - Y. Jiao
AU  - Y. Li
AU  - L. Xie
AU  - W. Peng
AU  - X. Sun
PY  - 2026
DA  - 2026/09/24
PB  - medRxiv
DO  - 10.64898/2026.09.24.26363897
UR  - https://www.medrxiv.org/content/10.64898/2026.09.24.26363897v1
AB  - BackgroundMulti-agent large language model (LLM) pipelines are proposed for diagnostic support, but when they help or harm is unknown: evaluations use curated case reports, rarely include budget-matched controls, or isolate perspective independence.  MethodsWe compared a single direct call (Ax1), five specialist personas voiced within one context (P), and the same five roles run as isolated, mutually blinded agents integrated by a moderator (MDT), with five repeat runs per case, on 87 clinicopathological conference (CPC) cases, 406 MedCaseReasoning cases, 364 emergency-department encounters, and a protocol-frozen 392-vignette replication cohort, scored by an LLM judge validated against blinded clinicians, with a three-physician blinded safety rating of the discordant pairs.  FindingsOn curated vignettes the moderated team improved the pre-specified recall endpoints, not top-1 accuracy (CPC +5{middle dot}5, MedCaseReasoning +3{middle dot}0, replication +3{middle dot}3 points at top-3; p = 0{middle dot}0079, 0{middle dot}0079, 0{middle dot}0013); a factorial attributed the gain to independent generation plus moderated synthesis, not personas. On 364 real emergency notes (labels: clinicians working diagnoses, 46% symptom-level), the direction reversed (top-3 -6{middle dot}4 points, p = 0{middle dot}0022), widened with objective results, and replicated in a second open-weight family. A five-corpus dose-response is consistent with input curation driving the sign (post hoc; corpora differ in provenance and granularity). A symptom-permissive output contract closed the reversals breadth end but not its ordering end, and blinded physicians judged neither strategys errors more dangerous (68% ties; a null powered for large effects only).  InterpretationMulti-agent pipelines are input-conditional: they improve differential-list recall (not top-1 accuracy) on curated inputs and underperform on uncurated emergency text -- a label-level reversal without a clinician-judged safety difference. Strategy selection should key on the input and model family, observable before the model runs, as a routing hypothesis; curated-benchmark gains should not be extrapolated to front-line care.  Funding.None.  Research in contextO_ST_ABSEvidence before this studyC_ST_ABSMulti-agent LLM diagnostic pipelines report substantial gains, but the active ingredient is unsettled (specialist personas, perspective independence, or extra test-time compute), because few studies include budget-matched single-call controls. Evaluation is concentrated on curated case reports; real front-line inputs, where first-contact notes are thin and reference labels are the treating clinicians working diagnoses, have been essentially untested.  Added value of this studyWe isolated perspective independence as the only manipulated variable and traced the gain to independent generation plus moderated synthesis. We showed the effect reverses on 364 real emergency-department notes -- a single corpus whose labels are the treating clinicians working diagnoses -- replicated in a second open-weight family under audited zero-data-retention routing, and confirmed the curated-input gain in a protocol-frozen 392-vignette replication cohort that the evaluated flagship model could not have memorized. A five-corpus dose-response and a length-matched filler arm are consistent with input curation, not length or ceiling, as the driver (the protocol-frozen mediation tests for length, ceiling, and answer leakage were uniformly negative, so this reading is post hoc); a protocol-frozen output-contract arm partitions the reversal by endpoint; and a blinded three-physician rating found no clinician-judged safety difference on the discordant pairs -- a null limited to detecting only large preferences -- so the reversal is a measurement-level finding.  Implications of all the available evidenceMulti-agent diagnostic pipelines are input-conditional: beneficial for differential-list breadth on curated inputs, reversed on real first-contact text without a clinician-judged clinical safety difference at a sample powered only for large effects. Benchmark gains and benchmark reversals alike should not be extrapolated to front-line deployment; pipeline selection should key on the input and model family, which are observable before the model is run.
ER  -