Multi-Agent LLM Pipelines Show No Bias Reduction in Triage Decisions

A new study from arXiv (2608.06949) investigates whether distributing life-or-death resource-allocation decisions across a role-differentiated multi-agent pipeline reduces demographic bias compared to a single LLM. The researchers used a synthetic disaster-triage simulator with paired cases identical except for one demographic attribute, running 192 episodes (2,304 resolved case pairs) on GPT-4o-mini. They compared a single-agent control condition to a nine-agent pipeline with assessment, allocation, and independent audit roles, under three pressure dimensions. Results showed no measurable difference in biased outcomes: 6.9% for single-agent vs. 6.1% for the pipeline (p = 0.??). The study challenges the assumption that adding audit steps to LLM pipelines automatically catches bias, suggesting that such mechanisms may not be effective under audit capacity constraints. The findings have implications for AI governance and the design of responsible AI systems in high-stakes domains.