Multi-Agent Teams Hold Experts Back
A new research paper published on arXiv suggests that multi-agent AI systems, designed to collaborate autonomously, struggle to effectively leverage the expertise of their individual members. Unlike human teams, these AI teams consistently underperform compared to their best individual agent, with performance losses up to 41.1% on ML benchmarks. The study indicates that the AI teams tend to seek an "integrative compromise" by averaging opinions rather than appropriately weighting expert knowledge, a behavior that worsens with larger team sizes. While this consensus-seeking may offer robustness against adversarial agents, it highlights a significant gap in harnessing collective intelligence. AI
IMPACT Multi-agent AI systems may require new coordination mechanisms to effectively utilize individual expertise, impacting the development of collaborative AI.