Study reveals LLMs naturally gravitate toward similar actions in multi-agent coordination but struggle to maintain diversity when rewarded for it

Researchers distinguish between primary algorithmic monoculture (inherent action similarity in AI agents) and strategic monoculture (agents adjusting similarity based on incentives)

Experiments comparing human and large language model subjects show LLMs exhibit high baseline similarity in their actions

Both humans and LLMs respond to coordination incentives by regulating their similarity levels, but LLMs lag significantly behind humans when incentivized to sustain diversity

LLMs demonstrate strong coordination abilities when agents use similar actions, indicating their monoculture tendency is largely predetermined rather than strategically adaptive