Multi-Agent Teams Hold Experts Back
Multi-Agent Teams Hold Experts Back
AI BriefingIn self-organizing multi-agent LLM systems, teams fail to make effective use of expert members’ specialized knowledge. Across multiple benchmarks, even when explicitly told who the experts are, teams still underperform relative to the best member (the expert agent) acting alone, with performance losses of up to 41.1%. The main reason for failure is not identifying experts, but failing to make effective use of their input. Conversation analysis shows that teams tend toward “integrative compromise,” averaging expert and non-expert views. This tendency grows as team size increases and is negatively correlated with performance. At the same time, this consensus-seeking behavior improves robustness against malicious agents, revealing a fundamental trade-off between collaborative alignment and effective use of expertise.
Source: Apple Machine Learning Research · machinelearning.apple.com