acceptodds
Under review as a conference paper at ICLR 2027

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

Abstract

Multi-agent LLM systems increasingly mix models from several providers, yet exposing each agent's underlying model identity to its peers significantly impairs cooperation. We show 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. We argue that the label itself causes this split, which we define as *factionalism*. We show and measure this phenomenon in two cooperative games and on a reasoning benchmark, with nine to twenty-five agents drawn from up to five open-weight model families. We further show that when the announced families are shuffled, or replaced by arbitrary labels, the factions still follow this information; when the label is removed, this behavior disappears. In strictly cooperative tasks, labeled groups spend on average % more rounds and % more tokens to reach a decision, and their success rate drops from % to %. The effect replicates across tasks, group sizes and model families. Withholding identity labels from the agents is simple and effective mitigation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent LLM systems, in-group bias, identity metadata, coordination failure

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CV8pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how explicit model-family identities affect cooperation and coordination in multi-agent LLM systems. Across tasks including Leader Election, Exclusion, and GPQA-Diamond, the authors compare labeled agents with unlabeled agents and agents assigned shuffled or misleading identities. They find that identity labels can induce factional behavior, alter agreement patterns, and reduce task success while increasing token costs. The results suggest that making model identities visible can influence multi-agent dynamics beyond agents’ underlying capabilities.

Critical strengths

The paper studies a timely and practically relevant question as multi-agent LLM systems become more common. The use of truthful, shuffled, and neutral identity conditions is a compelling experimental design that helps investigate whether behavior depends on identity labels rather than model capabilities alone. The evaluation spans multiple tasks and team sizes, and the combination of task outcomes, agreement-graph analysis, and qualitative examples provides several perspectives on how identity cues may affect agent interactions. The findings also suggest a concrete design implication: identity disclosure should be considered carefully when constructing multi-agent systems.

Critical weaknesses

The paper’s central claim about degraded cooperation is stronger than the experiments fully establish. The agreement-graph measures may conflate communication patterns, strategic voting, and genuine factional preference, while the tasks, particularly Exclusion, may make coalition formation strategically useful rather than simply dysfunctional. The experiments also do not fully disentangle the effects of an agent’s own identity label from the identities of other agents. In addition, statistical uncertainty and multiple-comparison concerns across the many experiments and outcomes deserve more attention, and the reported cost-performance trade-offs are mixed. This make it difficult to determine how broadly the findings generalize to genuinely cooperative settings.

Questions for authors
  1. How do you distinguish identity-driven factionalism from strategic coalition formation? Would the main findings persist in a task where agents have an explicitly shared objective and forming identity-based coalitions offers no strategic advantage?
  2. Can you isolate the causal effects of own identity versus others’ identities? For example, have you considered a factorial experiment that independently varies whether an agent sees its own model-family label and whether it sees other agents’ labels, while controlling for prompt length?
Reviewer_WP6rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper investigates identity-induced factionalism in heterogeneous multi-agent systems, demonstrating that when LLM agents are aware of their own and their peers' model families, they exhibit strong in-group favoritism and partition into identity-aligned factions. To quantify this structural bias across multi-agent networks, the authors introduce two graph-theoretic metrics: Edge Density Ratio (EDR), which measures homophily by comparing intra-group versus inter-group interaction agreement, and Adjusted Mutual Information (AMI), which evaluates the alignment between Louvain-detected interaction communities and ground-truth identity labels. Evaluated across two stylized cooperative games as well as a downstream GPQA-Diamond question-answering task mediated by leader election, the study reveals persistent identity-driven clustering and recommends masking unnecessary model-identity metadata from peer agents to foster unbiased collaboration while retaining it within a central orchestrator for routing and auditing.

Critical strengths
  1. The paper investigates a practically consequential design variable in heterogeneous multi-agent systems (MAS): the visibility of agent identity. Because model family names and agent profiles are routinely injected into system prompts or message headers in real-world orchestrations, examining how such cues shape emergent social dynamics is relevant. By demonstrating how locally rational agent interactions can aggregate into identity-aligned factions that degrade system-level coordination, the study highlights a structural vulnerability in collaborative MAS design.

  2. The experiment is well-constructed to isolate the causal effect of identity labels from intrinsic model traits and brand priors, bolstered by quantitative graph-theoretic operationalizations. The mislabeled counterfactual experiments, which decouple visible family labels from the underlying models, provide evidence that emergent clustering follows perceived labels rather than latent linguistic styles or capability similarities, while rotated neutral color labels rule out priors tied to brand reputation.

  3. The evaluation is comprehensive and tests robustness across varied network scales, demographic compositions, and model families. By evaluating different ensembles, the authors mitigate concerns that factionalism is merely an artifact of a specific population geometry. In addition, targeted leave-one-out and model-substitution ablations confirm that while effect magnitudes vary across backbones, the overarching phenomenon of identity-induced clustering persists across multi-agent configurations.

Critical weaknesses
  1. The paper has not established a direct causal link between factionalism and the observed degradation in collaborative performance. While the experiments demonstrate two co-occurring phenomena when model identity labels are exposed, co-occurrence does not prove that graph-level clustering is the mediating mechanism driving the performance loss. Alternative confounders remain plausible. To substantiate the claim, the authors might introduce some structural interventions to test whether suppressing factional clustering directly restores collaborative efficiency, alongside instance-level analyses correlating factionalism intensity with downstream task metrics across individual runs.

  2. The experimental tasks rely heavily on stylized social-choice games that diverge substantially from realistic multi-agent collaboration workflows. The primary benchmarks reduce collaboration to symmetric voting and consensus formation, where identity metadata has no functional utility. Even the downstream GPQA-Diamond evaluation reuses the exact same Leader Election protocol to select a single responder, meaning the study evaluates a remarkably narrow range of interaction dynamics across all three settings. In realistic heterogeneous MAS deployments, collaboration typically involves asymmetric role specialization, task decomposition, iterative verification, and the exchange of complementary domain evidence (e.g., coding versus mathematical reasoning). The authors should incorporate at least one substantive collaborative workflow involving division of labor and shared artifact generation to demonstrate that identity masking remains beneficial when model heterogeneity is relevant.

  3. The evaluation protocol on GPQA-Diamond conflates social coordination failure with epistemic reasoning degradation, obscuring the true source of the reported performance drop. Because the system only submits an answer if the ensemble successfully elects a leader within the allotted rounds, an unsuccessful trial aggregates two distinct failure modes: failing to reach electoral consensus ("No Leader Elected") versus electing a leader whose substantive answer is wrong ("Incorrect Answer"). Moreover, because AMI on GPQA is still computed over electoral voting commitments, both the structural and performance metrics reflect a repeated test of the voting bottleneck rather than scientific reasoning.

  4. The neutral color-label experiments omit downstream task performance metrics. While Table 4 and Appendix Tables 8&9 demonstrate that arbitrary labels are sufficient to induce structural homophily and community clustering, the paper did not report whether this minimal-group factionalism translates into the same task failure rates, convergence delays, and token overheads observed under real model-family labels (Table 2). Arbitrary color labels might induce preferential in-group interaction while still permitting cross-group consensus when task incentives require it, whereas real model names might trigger stubborn capability-related priors that deadlock coordination.

Questions for authors

See weaknesses

Reviewer_cFKQOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces the concept of factionalism, a type of LLM multi-agent system miscoordination error in which the interaction graph is partitioned according to an assigned, visible label. It explores the phenomenon on two coordination tasks, Leader Election and Exclusion, where there is no incentive for group formation. They further extend the experiments to GPQA-Diamond, a more realistic task that maintains the same interaction protocol as Leader Election but where success requires submitting the correct answer to a GPQA question.

The authors show factionalism emerges in all configurations where visible identity labels are provided, and disappears in the label-free condition. Factionalism emerges both when agents are assigned their true model families as labels, as well as when they are mislabeled. These results are reproduced in an ablation study where model identity labels are replaced by neutral colors.

Lastly, the authors show how factionalism impairs coordination, generally reducing task success, increasing the number of rounds to coordinate, and increasing the number of reasoning tokens.

Critical strengths
  • The authors present a clear measurement framework, along with an operational definition of factionalism that hinges on homophily and community structure, and a decision rule to detect this phenomenon. This framework is applicable to any MAS that stores interaction logs and shares the same interaction and voting structure.
  • The authors validate the connection between visible labels and factionalism by mislabeling the agents’ identities, showing community formation with respect to the assigned labels rather than the true model identities. They further support this result by computing the stylistic and adoption scores, as well as by running an ablation where identity labels are replaced by neutral colors.
Critical weaknesses
  • The paper presents factionalism as a newly identified phenomenon. However, the neutral color and mislabeled experiments show that the assigned label (i.e. assigned group) is what drives the effect. This phenomenon is in-group bias under minimal group conditions, and its existence in LLM agents was already established by [1] in a minimal-group allocation task. This is a framing issue rather than a lack of contribution given that the manuscript differs substantially with respect to [1] (heterogeneous populations, no incentives apart from task success, direct messaging).
  • For the Leader Election and GPQA-Diamond environments, there is no visible public information about the votes. All information about the other players’ actions is obtained via the payloads i.e. depends on who the agents decide to communicate with. There is, however, no analysis of the relationship between the communication structure and the assigned labels. Additionally, the EDR and AMI computations do not distinguish between two agents that never communicate, and a pair in which one repeatedly messages the other without ever converging to a common vote. Without this analysis, it is impossible to differentiate whether agents match the votes of same-label peers more often, or simply match the votes of whoever communicated with them.

[1] Zongwei Wang, Bincheng Gu, Hongyu Yu, Junliang Yu, Tao He, Jiayin Feng, Chenghua Lin, and Min Gao. When agents see humans as the outgroup: Belief-dependent bias in llm-powered agents, 2026. URL https://arxiv.org/abs/2601.00240.

Questions for authors
  • For the Leader Election and GPQA-Diamond environments, could the authors provide the AMI between the detected communities and the assigned labels for a communication graph where edges count the messages being sent (without requiring a vote match in the next round)? Could the authors also report the agreement rate among contacted pairs, splitting by same and different labels?
  • The Exclusion game’s AMI scores are consistently higher than the Leader Election ones. Could the authors provide any intuition as to why that is the case? For example, is it related to the Exclusion game including a public tally, the task semantics, or the longer run lengths?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.