Mitigating Hallucination Cascades in Multi-Agent Systems via Consensus-Aware Knowledge Verification

Large Language Model (LLM)-based multi-agent systems have demonstrated strong capabilities in complex task automation. However, these systems suffer from a critical vulnerability: hallucination cascade, where erroneous information generated by one agent propagates through the system and is amplified by downstream agents. We propose Dynamic Consensus Auditing (DCA), a novel framework that introduces a lightweight audit agent to monitor inter-agent message flows. DCA combines Retrieval-Augmented Generation (RAG) with uncertainty quantification based on semantic entropy to detect potentially hallucinated content. When high-uncertainty messages are detected, DCA triggers a peer review mechanism where multiple agents vote on the validity of the information. Experiments on HotpotQA, TriviaQA, and MuSiQue benchmarks demonstrate that DCA reduces error propagation rates by 42.5% (relative reduction) on average compared to baseline multi-agent systems, with a token overhead of 15.3% (averaged across datasets). Our results suggest that, in the context of multi-hop question answering with three-agent pipelines, consensus-aware verification provides a practical balance between reliability and computational efficiency in multi-agent collaboration.