Strategic Algorithmic Monoculture: LLMs Show Dangerously High Coordination Similarity in Multi-Agent Systems

arXiv·medium signalXBlueskyLinkedInCopy link

Ballestero, Hosseini, and Khanna experimentally distinguish primary algorithmic monoculture (baseline action similarity) from strategic monoculture (agents adjusting similarity in response to incentives). LLMs exhibit high baseline similarity and, like humans, strategically increase it when coordination pays — but unlike humans, they cannot reduce it when differentiation is needed. Critical finding for multi-agent deployments: LLM agent swarms may converge on identical strategies even when diversity would produce better outcomes.