The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams
A study on arXiv (cs.AI) published 2026-08-24 introduces the 'interaction tax' in multi-agent LLM systems. It finds that when agents read each other's complete outputs, their proposals converge within one round, erasing diversity. Testing 11 verifier-scored optimization tasks under matched budgets, full-solution interaction is a weak default; independent proposal generation avoids collapse. Critique helps only if the violated rule is easy for the LLM to find and fix.
The paper argues that contradictory findings on multi-agent LLM interaction stem from a missing distinction: not all communication is equal. Different model families find structurally different solutions, but full-solution interaction causes convergence to the first solution seen, reducing diversity. Independent proposal generation avoids this collapse. The study tested 11 verifier-scored optimization tasks under matched budgets.
Full-solution interaction leads to rapid convergence, erasing diversity that motivates multi-model use. Independent proposal generation preserves diversity. Critique is effective only when the violated rule is easily identifiable and fixable by the LLM.
For multi-agent system designers, defaulting to full-solution communication may negate benefits of model diversity. Independent generation or selective communication could be more cost-effective under equal budgets.
Avoiding the interaction tax can reduce compute costs and improve solution quality in multi-agent LLM applications, making them more viable for optimization tasks.
Expect further research on communication protocols that preserve diversity, such as partial information sharing or structured critique. Practical multi-agent frameworks may adopt independent generation as a default.