Researchers found that exposing model identities to each other in multi-agent LLM systems can cause factionalism, where agents prefer interacting with others carrying the same label, leading to decreased cooperation. Withholding identity labels can mitigate this issue.