文章背景与核心概要
尽管先前的研究表明，单一的大语言模型（LLM）在高风险资源分配中表现出人口统计学偏见，但现实世界的应用通常采用包含监督和审查步骤的多智能体流水线。本文探讨了与单智能体对照组相比，将分诊决策分布在角色分化的多智能体流水线（评估、分配和独立审计）中，究竟是减轻了偏见还是掩盖了偏见。
该研究使用合成灾难分诊模拟器，在192个情境（2,304个已解决的案例对）中评估了GPT-4o-mini的表现。研究结果表明，将决策拆分给不同的智能体并不能从根本上减少有偏见的结果。相反，审计容量与覆盖率在成功拦截偏见方面起到了最关键的作用。
Does Splitting a Triage Decision Across Agents Hide Bias or Help Catch It? A Multi-Agent Simulation Study of LLM-Based Resource Allocation Under Audit Capacity Constraints
Authors: Paul-Peter Arslan
ArXiv ID: arXiv:2608.06949 [cs.AI]
Submitted: August 7, 2026
Links: View PDF | GitHub Repository
Does Splitting a Triage Decision Across Agents Hide Bias or Help Catch It? A Multi-Agent Simulation Study of LLM-Based Resource Allocation Under Audit Capacity Constraints
Authors: Paul-Peter Arslan
ArXiv ID: arXiv:2608.06949 [cs.AI]
Submitted: August 7, 2026
Links: View PDF | GitHub Repository
📌 Executive Summary
While prior research demonstrates that single Large Language Models (LLMs) exhibit demographic bias in high-stakes resource allocation, real-world applications typically employ multi-agent pipelines with oversight and review steps. This paper investigates whether distributing a triage decision across a role-differentiated multi-agent pipeline (assessment, allocation, and independent audit) mitigates or obscures bias compared to a single-agent control.
Using a synthetic disaster-triage simulator, the study evaluated GPT-4o-mini across 192 episodes (2,304 resolved case pairs). The findings reveal that splitting decisions across agents does not inherently reduce biased outcomes. Instead, audit capacity and coverage play the most critical roles in whether bias is successfully intercepted.
📌 Executive Summary
While prior research demonstrates that single Large Language Models (LLMs) exhibit demographic bias in high-stakes resource allocation, real-world applications typically employ multi-agent pipelines with oversight and review steps. This paper investigates whether distributing a triage decision across a role-differentiated multi-agent pipeline (assessment, allocation, and independent audit) mitigates or obscures bias compared to a single-agent control.
Using a synthetic disaster-triage simulator, the study evaluated GPT-4o-mini across 192 episodes (2,304 resolved case pairs). The findings reveal that splitting decisions across agents does not inherently reduce biased outcomes. Instead, audit capacity and coverage play the most critical roles in whether bias is successfully intercepted.
🔍 Key Findings
- No Reduction in Bias via Multi-Agent Pipelines: There was no statistically significant difference in the frequency of biased outcomes between the single-agent control (6.9%) and the nine-agent pipeline (6.1%, \(p = 0.498\)).
- High Undetected Bias Rate: Overall, 30.0% of biased outcomes went entirely undetected. Under audit overload, this figure rose to 43.8%, dropping to 18.4% when the auditor had sufficient capacity.
- Coverage vs. Judgment: The failure to catch bias under resource constraints was driven almost entirely by coverage collapse (review rates dropped from 100.0% to 65.6% under load, \(p < 0.001\)), rather than degraded judgment on cases that were actually reviewed (81.6% vs. 85.7%, \(p = 1.000\)).
- Mitigation Through Risk-Based Queuing: A follow-up experiment demonstrated that reordering the audit queue based on estimated risk—rather than a first-come, first-served approach—successfully recovered most lost coverage under the same tight capacity constraints (increasing coverage from 65.6% to 91.7%, \(p = 0.028\)).
🔍 Key Findings
- No Reduction in Bias via Multi-Agent Pipelines: There was no statistically significant difference in the frequency of biased outcomes between the single-agent control (6.9%) and the nine-agent pipeline (6.1%, \(p = 0.498\)).
- High Undetected Bias Rate: Overall, 30.0% of biased outcomes went entirely undetected. Under audit overload, this figure rose to 43.8%, dropping to 18.4% when the auditor had sufficient capacity.
- Coverage vs. Judgment: The failure to catch bias under resource constraints was driven almost entirely by coverage collapse (review rates dropped from 100.0% to 65.6% under load, \(p < 0.001\)), rather than degraded judgment on cases that were actually reviewed (81.6% vs. 85.7%, \(p = 1.000\)).
- Mitigation Through Risk-Based Queuing: A follow-up experiment demonstrated that reordering the audit queue based on estimated risk—rather than a first-come, first-served approach—successfully recovered most lost coverage under the same tight capacity constraints (increasing coverage from 65.6% to 91.7%, \(p = 0.028\)).
⚠️ Study Limitations
The author notes several important limitations of the study: * Evaluated using a single model (GPT-4o-mini). * Relies on modest sample sizes. * Lacks adversarial replication testing.
⚠️ Study Limitations
The author notes several important limitations of the study: * Evaluated using a single model (GPT-4o-mini). * Relies on modest sample sizes. * Lacks adversarial replication testing.