Shuai Wu, Xue Li, Yanna Feng, Yufang Li, Zhijun Wang

Researchers propose Council Mode, a multi-agent consensus framework that routes queries to multiple heterogeneous LLMs in parallel and synthesizes their outputs through a dedicated consensus model to reduce hallucinations and bias. The system uses intelligent triage to classify query complexity, dispatches to diverse frontier models simultaneously, and applies structured consensus synthesis to identify agreement, disagreement, and unique findings. Evaluation shows 35.9% relative reduction in hallucination rates on HaluEval and 7.8-point improvement on TruthfulQA versus the best individual model, while maintaining lower bias variance across domains.