The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams
Having agents share complete answers can make them less useful as a team.
The paper says different model families often start with different solution structures, but those differences collapse after agents read one another’s full outputs. In 11 verifier-scored optimization tasks under matched budgets, independent proposals avoided that convergence better than full-solution interaction. Critique helped only when the broken rule was easy for the LLM to spot and repair. ArXiv · AI/CL/LG's note
The paper says different model families often start with different solution structures, but those differences collapse after agents read one another’s full outputs. In 11 verifier-scored optimization tasks under matched budgets, independent proposals avoided that convergence better than full-solution interaction. Critique helped only when the broken rule was easy for the LLM to spot and repair. ArXiv · AI/CL/LG's note
score 5