The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams
Does multi-agent LLM interaction help or hurt? Some work reports gains from debate (Du et al., 2024), critique loops (Chen et al., 2025), and mixture-of-agents synthesis (Wang et al., 2025), while other work finds that interaction adds cost without improving quality under equal budgets (Tran & Kiela, 2026; Xu et al., 2026; Jarrett et al., 2025), or that independent sampling already captures…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
