Harnessing Disagreement: Detecting Correlated Agreement Blindness in Multi-Agent Triage
The paper identifies 'correlated agreement blindness' as a structural flaw where multi-agent systems fail to escalate dangerous cases when their improving base models converge on incorrect answers. The proposed ARAT system uses a directed-star architecture with Random Forest and k-NN agents, plus a meta-model, to mitigate this by using disagreement and conservative overrides to improve safety performance.