Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
It is argued that when agents are aware of each other's model family, the group splits into clusters, where agents prefer interacting with others carrying their same label, although nothing in the task rewards or asks for such a split.