Managing Hallucinations with LLM-based Multi-Agent Workflows

Large Language Model (LLM)-based AI agents can support complex knowledge tasks by planning, searching, and synthesizing information. However, their tendency to generate inaccurate or fabricated content remains a major barrier to reliable use. Our study examines whether the design of a multi-agent workflow can reduce GenAI hallucinations. Our results show that certain architectures of multi-agent workflow substantially improved output accuracy, reduced both factual and faithful hallucinations, and increased overall task success. These findings suggest that the way AI agents are organized and supervised may matter than the raw capability of underlying models.

Jan 5, 2027 12:00 PM

The 60th Hawaii International Conference on System Sciences (HICSS) 2027