Clinical multi-agent systems launder demographic triage disparity
Original title: Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems
On 120 MIMIC-IV cases, all 20 model-topology cells gave Black patients more acute triage; pooling perspectives cut the gap 67–94%.
A full factorial of five model families and four aggregation topologies on 120 real MIMIC-IV emergency cases under counterfactual demographic swaps found that no configuration neutralizes disparity. All 20 cells commit the direction of agents' independent answers, assigning Black patients more acute triage. Fairness instructions and demographic masking leave the gap untouched, and per-agent self-debiasing widens it. A diverse-perspective ensemble that pools all five perspectives every round reduces the disparity by 67 to 94% without retraining; the topology that never pools them cannot be reliably repaired.
Unconfirmed: Results are from the authors' counterfactual swaps on 120 MIMIC-IV cases, not a deployed clinical system.
- Every model has a topology that amplifies its baseline gap, but which one varies by model.
- The authors say bias is already encoded in ordinary clinical numbers.
- Interventions close the gap only when counterfactual versions are visible and pooled.
- Single-model, single-output, single-version audits can certify parts but not the committed decision.
From the original article
1 min read at arxiv.orgClinical large language models are increasingly fielded as multi-agent teams, on the assumption that deliberation averages out individual error. It does not average out bias.
By Xiaoyang Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.