Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Xavier Del Giudice, Alessio Palma, Matteo Migliarini, Fabio Galasso, Indro Spinelli
cs.MA, cs.AI
2026-09-28
LLM agents that see each other's model family form label-aligned factions, spend 55% more tokens and 30% more rounds, and drop from 96% to 81% success; removing labels erases it.
Multi-agent LLM systems increasingly mix models from several providers, and identity leaks everywhere: system prompts, agent names, and message headers often carry each agent's model family. The metadata is meant for operators, not agents, yet it lands in the agents' context, and nobody had systematically tested whether agents act on it. The paper asks one narrow question: does merely telling agents each other's model family degrade how well the group cooperates?
Two no-stakes coordination games, with 9 to 25 agents drawn from up to five open-weight families (Gemma, GPT-OSS, Nemotron, Qwen, GLM), 50 runs per configuration. Each round, every agent casts one ballot, for a leader or for a peer to remove, then messages exactly two chosen peers. The tasks have no correct answer, no reward for favoring same-family peers, and no personas; forming factions earns nothing. The only manipulated variable is the visible label:
Each run becomes an agreement graph whose edge weight is the rate at which a recipient's next-round commitment matches the sender's. A within-group to between-group edge-density ratio measures homophily, and Louvain community detection scored with Adjusted Mutual Information (AMI) measures whether communities split along labels. The mislabeled condition is the sharpest design choice: if clustering were driven by architectural similarity, communities would align with true families; if the label is the cause, they follow the announced tags.
All 24 labeled or mislabeled experiments were factional; none of the 12 unlabeled ones was.
| Setting | AMI (community-label alignment) | Outcome |
| Exclusion, labeled | 0.55-0.76 | all significant |
| Leader Election, labeled | 0.24-0.53 | all significant |
| Mislabeled, vs visible labels | 0.18-0.69 | all significant |
| Mislabeled, vs true families | mostly n.s. | near chance |
| Unlabeled | 0.02-0.08 | none significant |
With family names swapped for colors, factions still form along the visible colors, and after shuffling the colors the split follows the colors, not the architectures. The label itself, not recognition of a peer's writing style, drives the division.
The costs land on the ledger. In strictly cooperative tasks, labeled groups spend on average 30% more rounds and 55% more tokens, and success drops from 96% to 81%; mislabeled groups also slow down and cost more. Leader Election takes 1.4x rounds and 1.8x tokens; in Exclusion at N=12 the success ratio falls to 0.61; tokens per successful run rise 1.5x to 2.2x, growing with roster size. At N=25 in Exclusion, 8 of 100 unlabeled runs converge and zero labeled runs do.
Two log-probability probes localize the mechanism. A stylistic score on raw peer messages tracks true architecture but is tiny (at most 0.07 nats/char), so it cannot explain the split. An adoption score on peers' actual votes inside the receiving agent's real context follows the visible label. Asked which family they belong to, models given a false label repeat it 78% of the time and name the true one in 11%; they cannot detect the lie. The pattern replicates on GPQA-Diamond, where answers have objective ground truth: 1.31x more rounds at N=9, with success down from 81.8% to 75.0%.
The engineering takeaway is nearly free: do not expose model identity to agents. An orchestrator can keep provider metadata for routing, calibration, and auditing without letting it enter the interaction context. Mixed-provider pipelines are common now, and embedding identity in agent names is a widespread habit; by this paper's numbers, that habit burns tokens.
The phenomenon also maps cleanly onto the minimal group paradigm from social psychology, where humans favor arbitrary in-groups. LLM agent populations reproduce it with model-family metadata as the arbitrary tag, in a controllable and measurable setting.
The authors scope the claims deliberately: at most five open-weight families, text-only consensus games built for this study, a matched-sampling check on four families. Closed models, larger rosters, other decoding regimes, and richer agentic workflows are untested, and GPQA-Diamond reuses the Leader Election protocol.
Two concerns stand out. The consensus games are self-referential with no external answer key, and the effect shrinks on GPQA-Diamond (a 7-8 point success gap, roughly 10% more tokens), so stronger task signals in production systems could drown out the label effect. The tested checkpoints are mid-scale open-weight models around 20B-36B; whether frontier closed models factionalize the same way is unanswered.