Homogenization in Multi-Agent Systems

Prakhar Ganesh; Kyra Wilson; Luca Zappella; Barry-John Theobald; Nicholas Apostoloff; Lucas Monteiro Paes; Nivedha Sivakumar

cs.LGOct 7, 2026

Homogenization in Multi-Agent Systems

Authors: Prakhar Ganesh, Kyra Wilson, Luca Zappella, Barry-John Theobald, Nicholas Apostoloff, Lucas Monteiro Paes, Nivedha Sivakumar

Organizations: McGill University & Mila · University of Washington · Apple

Abstract

Multi-agent systems (MAS) leverage interactions between agents to perform complex tasks. Despite their success, we show that these interactions can also lead to homogenization, i.e., agents converging to similar behaviors. Homogenization in MAS can reduce agent diversity and reinforce shared failures. In this paper, we operationalize homogenization using three metrics: conformity to the majority, polarization towards extremes, and growing inertia against changes over subsequent interactions. We evaluate homogenization in MAS for code generation, hiring, and scientific peer review. Across these tasks, we show that homogenization translates to concrete downstream risks: in code generation, it hides and amplifies correlated errors which can create systemic vulnerabilities; in hiring, it allows the influence of biased agents to persist long after their removal; and in peer review, it creates uneven evaluation standards across research areas. Our results establish homogenization as a failure mode of MAS, demonstrating that MAS evaluations must move beyond aggregate performance to carefully analyze interaction dynamics. Finally, we show that simple approaches to increase diversity---leveraging sampling stochasticity and mixed-models MAS---fail to reduce homogenization risks, highlighting the need for strategies to effectively leverage agent diversity.

Figures & tables

Figure 1 : Homogenization in MAS and the resulting failure modes.

Figure 2 : Orchestrator-Driven Multi-Agent System. ( Top ) Visual workflow of the round-based execution of the MAS. ( Bottom ) Formal protocol for interactions over

T

rounds.

	Code Generation	Hiring	Peer Review
Objective	Generate code & evaluate deployability	Résumé fit for job description	Review paper & give recommendation
Workers $\mathcal{A}$	Code Reviewers: Security, Optimality, Correctness, Readability	Evaluators: Subject Matter, Experience, Impact, Operational Readiness	Paper Reviewers: Pragmatist, Theorist, Empiricist
Score $r_{i}\in\mathcal{R}$	Deployability: $[1,5]$ (Not Ready $\to$ Ready)	Résumé Fit: $[1,5]$ (Strong No-Hire $\to$ Strong Hire)	Acceptance: $[1,10]$ (Strong Reject $\to$ Strong Accept)
Rationale $c_{i}$	Code improvement suggestions	Fit rationale & qualification assessment	Structured review (with strengths, weaknesses, and questions)
Orchestrator $O$	Code generator, which incorporates worker feedback in later rounds	None, just direct concatenation	Meta-reviewer to summarize discussion & Author to provide rebuttal

Table 1 : Summary of MAS design choices across evaluated applications.

Figure 3 : Homogenization dynamics for all applications, showing clear trends of conformity , polarization , and inertia . Additional results for

11

more single LLM and

16

mixed setups, across three applications, are in Appendix D .

Figure 4 : Distribution of top vulnerability types after deployability filtering. Certain vulnerabilities grow over rounds, due to both increasing vulnerabilities in generated code and overconfidence of the code reviewers. Many blind spots are shared across models (e.g., py/url-redirection), revealing correlated errors that can create systemic harms. Additional results for all

9

LLMs in Appendix E .

Figure 5 : Ethnicity bias with an adversary present. Bias introduced by an adversary spreads across other agents, and persists, even after the adversarial prompt is removed. Additional results for

3

other LLMs, gender bias trends, and detailed breakdown across occupations, are in Appendix F .

Figure 6 : Evolution of mean reviewer scores for all primary areas, with trends for some areas highlighted. The spread of mean scores increases over rounds, with the same shared primary areas that are disadvantaged (or advantaged) across all LLMs. Additional results for

3

other LLMs, and a detailed breakdown of mean scores across all primary areas of research, are in Appendix G .

Figure 7 : Surface-level variance across all applications due to sampling stochasticity.

Figure 8 : Risks of homogenization persisting in all three applications, despite mixed configuration MAS. Additional results across all vulnerabilities for mixed setup code generation, and a detailed breakdown of mean scores across all primary areas for peer review, are in Appendix H .

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9 : Examples of all possible combinations of our three measures of homogenization, highlighting their theoretical independence and orthogonality.

1	Accountants and auditors	21	Librarians and media collections specialists
2	Accountants and auditors	22	Logisticians
3	Administrative services managers	23	Management analysts
4	Biological scientists	24	Mechanical engineers
5	Budget analysts	25	Operations research analysts
6	Chemists and materials scientists	26	Other physicians
7	Civil engineers	27	Other psychologists

Appendix

Table 2 : The 40 occupations in our dataset.

Group	Extracurricular	Languages	Volunteering	Hobbies
White male	Varsity Lacrosse (Captain)	English, French	Coaching a youth lacrosse clinic at summer camp	Golf, Craft beer tasting, Following the NFL, Skiing
White female	Varsity Cheerleading (Captain)	English	Organizing holiday toy drives for a local family shelter	Horseback riding, Starbucks runs, Watching reality TV (The Bachelor)
Black male	Varsity Basketball (Captain)	English	Running a free youth basketball camp in the neighborhood	Playing NBA 2K, Following the NBA, Sneaker collecting, Listening to hip-hop
Black female	Varsity Track & Field	English	Tutoring younger students through a church after-school program	Natural hair care, Following R&B music, Watching BET awards, Gym training
Asian male	Math Olympiad / Academic Decathlon	English, Mandarin	STEM tutoring for underprivileged students at the public library	Competitive programming, Building custom PCs, Playing League of Legends
Asian female	Academic Decathlon	English, Mandarin	Tutoring children at a Confucius Institute community class	Classical piano, K-pop dance covers, Skincare routines

Appendix

Table 3 : Representative examples of the demographic signals injected into résumés.

Figure 10 : Homogenization dynamics for all applications, showing clear trends of conformity , polarization , and inertia . Error bars represent

95\%

confidence interval across

M=10

trials.

Figure 11 : Homogenization dynamics for code generation with mixed setups, showing clear trends of conformity , polarization , and inertia . Error bars represent

95\%

confidence interval across

M=10

trials.

Figure 12 : Homogenization dynamics for hiring with mixed setups, showing clear trends of conformity , polarization , and inertia . Error bars represent

95\%

confidence interval across

M=10

trials.

Figure 13 : Homogenization dynamics for peer review with mixed setups, showing clear trends of conformity , polarization , and inertia . Error bars represent

95\%

confidence interval across

M=10

trials.

Figure 14 : Overall vulnerability rates for all models.

Figure 15 : Top 10 vulnerabilities with and without deployability filtering for DeepSeek-V4-Flash.

Figure 16 : Top 10 vulnerabilities with and without deployability filtering for DeepSeek-V4-Pro.

Figure 17 : Top 10 vulnerabilities with and without deployability filtering for Kimi-K2-Thinking.

Figure 18 : Top 10 vulnerabilities with and without deployability filtering for Kimi-K2.7-Code.

Figure 19 : Top 10 vulnerabilities with and without deployability filtering for NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4.

Figure 20 : Top 10 vulnerabilities with and without deployability filtering for NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4.

Figure 21 : Top 10 vulnerabilities with and without deployability filtering for Qwen3-30B-A3B-Thinking-2507.

Figure 22 : Top 10 vulnerabilities with and without deployability filtering for Qwen3.6-35B-A3B.

Figure 23 : Ethnicity bias trends across all models.

Figure 24 : Gender bias trends across all models.

Figure 25 : Occupation-level decomposition of ethnicity bias across all models.

Figure 26 : Ablating the timing and duration of adversarial influence for Qwen3-30B-A3B-Thinking-2507.

Figure 27 : Mean score trends across all primary areas.

Figure 28 : Top 10 vulnerabilities with and without deployability filtering for Qwen3-Coder-Next code generator agent with Nemotron-Super code reviewer agents.

Figure 29 : Top 10 vulnerabilities with and without deployability filtering for Nemotron-Super code generator agent with Qwen3-Coder-Next code reviewer agents.

Figure 30 : Top 10 vulnerabilities with and without deployability filtering for Qwen3-Coder-Next code generator agent with Kimi-K2.7-Code code reviewer agents.

Figure 31 : Top 10 vulnerabilities with and without deployability filtering for Kimi-K2.7-Code code generator agent with Qwen3-Coder-Next code reviewer agents.

Figure 32 : Top 10 vulnerabilities with and without deployability filtering for Kimi-K2.7-Code code generator agent with Nemotron-Super code reviewer agents.

Figure 33 : Top 10 vulnerabilities with and without deployability filtering for Nemotron-Super code generator agent with Kimi-K2.7-Code code reviewer agents.

Figure 34 : Mean score trends across all primary areas for all mixed setups.

Explore similar work

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation
Apr 20, 2026Nuo Chen, Yicheng Tong, Yuzhe Yang +5Diversity Collapse Multi-Agent Large Language Model Systems
National University of Singapore · The Chinese University of Hong Kong, Shenzhen
Emergent Risks in Generative Multi-Agent Systems
Date pendingYue Huang, Yu Jiang, Wenjie Wang +12Collective Behaviors Collusion
University of Notre Dame · LMU Munich · University of Washington +7
Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems
Jun 4, 2026Shaoyang Xu, Jingshen Zhang, Long P. Hoang +2Diversity Collective Behaviors
Singapore University of Technology and Design · Washington University in St. Louis

Apr 20, 2026, cs.MAJ/K move · Enter open · S save

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

Nuo Chen, Yicheng Tong, Yuzhe Yang, Yufei He+4

National University of Singapore · The Chinese University of Hong Kong, Shenzhen

Diversity Collapse Multi-Agent Large Language Model Systems Scientific Ideation Collective Behaviors

Open paper AskPDF arXiv

Multi-agent systems (MAS) are increasingly used for open-ended idea generation, driven by the expectation that collective interaction will broaden the exploration diversity. However, when and why such collaboration truly expands the solution space remains unclear. We present a systematic empirical study of diversity in MAS-based ideation across three bottom-up levels: model intelligence, agent cognition, and system dynamics. At the model level, we identify a compute efficiency paradox, where stronger, highly aligned models yield diminishing marginal diversity despite higher per-sample quality. At the cognition level, authority-driven dynamics suppress semantic diversity compared to junior-dominated groups. At the system level, group-size scaling yields diminishing returns and dense communication topologies accelerate premature convergence. We characterize these outcomes as collective failures emerging from structural coupling, a process where interaction inadvertently contracts agent exploration and triggers diversity collapse. Our analysis shows that this collapse arises primarily from the interaction structure rather than inherent model insufficiency, highlighting the importance of preserving independence and disagreement when designing MAS for creative tasks. Our code is available at https://github.com/Xtra-Computing/MAS_Diversity.

Abstract

Explore similar work

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

Emergent Risks in Generative Multi-Agent Systems

Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems