Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems. By Xavier Del Giudice and four others, from Sapienza University of Rome. Published on the twenty-eighth of September. No code repository is linked to it.

Multi-agent LLM systems increasingly mix models from several providers. The claim here is that telling each agent which model family its peers come from significantly impairs cooperation. Once agents know each other's family, the group splits into clusters, with agents preferring to interact with others that carry their own label. Nothing in the task rewards that split, and nothing asks for it. The agents are supposed to be on one team, working toward one outcome, and yet knowing which model family their peers come from is enough to sort them into camps. The authors argue that the label itself causes the split, and they give it a name: factionalism.

The phenomenon is shown and measured in two cooperative games and on a reasoning benchmark, with groups of nine to twenty-five agents drawn from up to five open-weight model families. The decisive tests manipulate the label. When the announced families are shuffled, or swapped for arbitrary labels, the factions still follow the label. When the label is removed, the behavior disappears. That pattern is what points at the label as the cause. If agents were grouping by something real about their peers, scrambling or replacing the names should not have mattered, and it did not stop the factions forming.

The costs are concrete. In strictly cooperative tasks, labeled groups take on average thirty percent more rounds and fifty-five percent more tokens to reach a decision. Their success rate falls from ninety-six percent to eighty-one percent. Slower, more expensive and less successful, all from a label the task never needed.

The effect replicates across tasks, group sizes and model families, but those are the ones tested: open-weight models, up to five families, groups of up to twenty-five, playing cooperative games and a reasoning benchmark. The cost figures come from the strictly cooperative tasks. What the shuffling tests establish is that agents respond to the announced label, not to anything they detect about their peers.

A model's name, dropped into a prompt, turns out to be enough to divide a group that has every reason to work together. The remedy the authors offer is correspondingly plain: withhold identity labels from the agents, which they report is simple and effective.