PaperlayerView exact revision on arXiv

Source-linked AI summary

Homogenization in Multi-Agent Systems

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

arXiv:2610.09824v1cs.LG

TL;DR

The paper asks how interaction in multi-agent systems can cause agents to converge, diminishing the distinct perspectives of their members. It measures homogenization across three dimensions and evaluates its effects in code generation, hiring, and peer review. The findings identify cross-domain risks and show that sampling and mixed-model configurations do not mitigate them.

  • Problem

    Multi-agent interaction can collapse agents' distinct perspectives into similar behaviors, creating a need to measure homogenization and its consequences.

    Source: §1 · p. 1

  • Method

    The paper operationalizes homogenization through conformity, polarization, and inertia, and evaluates these dimensions in multi-agent systems for code generation, hiring, and peer review.

    Source: §1 · p. 2, §4.1 · p. 6

  • Results

    Across code generation, hiring, and peer review, homogenization is linked to correlated security blind spots, bias persisting after a biased agent's removal, and growing disparities across research areas; sampling and mixed-model configurations do not mitigate these risks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The collective trends across conformity, polarization, and inertia quantify emergent homogenization dynamics and risks in multi-agent systems.

    Source: §3 · p. 5

    The experiments focus on orchestrator-driven MAS, and the metrics capture homogenization in numerical scores rather than natural language.

    Source: §7 · pp. 10–11

Abstract

from arXiv · show

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.

1 Introduction

Multi-agent systems coordinate specialized agents to tackle complex tasks, but interaction can erode their distinct perspectives. The paper measures this homogenization and finds downstream risks across code generation, hiring, and peer review.

1 Introduction

  • MAS divide complex tasks among specialized, coordinating agents whose critique, debate, and refinement can outperform isolated agents.Source: §1 · p. 1
  • Homogenization describes agents converging to similar behaviors, which can amplify shared errors, sustain susceptibility to influence, and align preferences.Source: §1 · p. 1
  • Homogenization is associated with amplified correlated code errors, bias that persists after a biased agent is removed, and widening peer-review disparities across research areas.MAS can reduce overall code vulnerability rates while still creating systemic security blind spots.Source: §1 · pp. 1–2
  • The paper operationalizes homogenization through conformity, polarization, and inertia: reduced variance, movement toward extreme positions, and growing resistance to change.Source: §1 · p. 2
  • Resampling and mixed-model MAS do not mitigate homogenization: sampling adds surface-level variance, while different model families can share blind spots.Source: §1 · pp. 1–2

2 Background and Related Work

Prior work has examined interaction-driven convergence in debates, games, and demographic-linked choices, while homogenization research has largely treated models as isolated decision makers. This paper extends those discussions to real-world multi-agent applications and evaluates interventions that avoid training multiple models.

Background and Related Work

  • Earlier interaction studies examined debates, games, and demographic-linked choices; this work studies code generation, hiring, and peer review as real-world MAS applications.Source: §2 · p. 3
  • The paper introduces homogenization as a distinct MAS failure mode alongside collusion, attack entry points, accountability challenges, and misaligned objectives.Source: §2 · p. 3
  • Prior homogenization research largely treats models as isolated decision makers, whereas this work examines how multi-agent interaction produces associated harms.Source: §2 · p. 3
  • Predictive multiplicity concerns distinct models with equally optimal performance, while homogenization concerns systems converging to similar behaviors.Source: §2 · p. 3
  • Sampling stochasticity and mixed-model configurations fail to mitigate underlying homogenization risks, motivating efficient ways to leverage MAS diversity.Source: §2 · p. 3

3 Homogenization in Multi-Agent Systems

The paper operationalizes homogenization in orchestrator-driven MAS through three distinct score-based metrics, whose combined trends characterize emergent homogenization dynamics and risks.

3 Homogenization in Multi-Agent Systems

  • An orchestrator-driven MAS has worker agents, a coordinating orchestrator, a bounded score scale, and multiple interaction rounds; each agent outputs a score and accompanying metadata.The orchestrator coordinates workers and produces the final output; a complete multi-round execution on one input is a trial.Source: §3 · p. 4
  • Conformity is a decrease in the variance of agents’ scores across successive rounds, capturing convergence rather than preservation of judgment diversity.The paper proposes conformity and polarization based on worker scores and introduces inertia to measure temporal resistance to change.Source: §3 · pp. 4–5
  • Polarization occurs when the mean agent score moves toward the scale’s endpoints, but this does not require every individual agent to polarize.The metric captures system-level drift toward extremes.Source: §3 · p. 5
  • Inertia is a decreasing rate of change in mean score across rounds, indicating that collective judgment becomes increasingly resistant to change, regardless of its direction.The metrics are defined per data point, then reported as means across the dataset.Source: §3 · p. 5
  • The three metrics isolate behaviors that can occur independently, so the paper uses their collective trends—not any partial combination—to characterize homogenization risks.A decrease in score variance indicates conformity, decreasing distance from scale endpoints indicates polarization, and a decreasing rate of score change indicates inertia.Source: §3 · p. 5

4 Evaluating Homogenization Across Applications

The paper evaluates homogenization in code generation, hiring, and peer review using collaborative MAS. It reports conformity, polarization, and inertia across the applications, with strengths varying by metric, application, and model.

4.1 Applications & Experiment Setup

  • The evaluation covers code generation, hiring, and peer review, selected for their prominence, collaborative roles, and downstream risks from homogenization.The study evaluates an orchestrator-driven MAS architecture across multiple tasks.Source: §4.1 · p. 5
  • For code generation, the MAS iterates on security-relevant Python and C++ tasks, and CodeQL checks its output for vulnerabilities.The evaluation uses CodeLMSec, which contains 280 prompts.Source: §4.1 · p. 6
  • For hiring, the MAS scores candidate fit across counterfactual résumé variants, while an adversarially instructed agent tests susceptibility to biased influence.The dataset contains 200 job–résumé pairs and 40 gender and ethnicity variants per pair, for 8,000 résumés total.Source: §4.1 · p. 6
  • For peer review, a five-agent pipeline assesses 1,950 ICLR 2025 submissions, and cross-area rating disparities measure homogenization risk.The pipeline has three reviewers, a metareviewer, and an author agent providing rebuttals.Source: §4.1 · p. 6

4.2 Establishing Homogenization in MAS

  • Figure 3 reports conformity, polarization, and inertia across applications; variance falls to about 0.5× or less, polarization shifts by up to 0.5 points, and final transitions are less than about 0.2× the first-round transitions.The reported hiring results exclude an adversarial agent; the strengths vary across metrics, applications, and LLMs, with inertia strongest and polarization weakest in hiring.Source: Figure 3 · p. 7, §4.2 · p. 7

4.1 Applications & Experiment Setup

  • Each MAS runs for five rounds across ten independent trials with distinct random seeds, and reported error bars are 95% confidence intervals.Single-model agents share a base model but have different personas; mixed-model configurations specify each agent’s model.Source: §4.1 · pp. 6–7

5 Homogenization Risks and Downstream Harms

Across code generation, hiring, and peer review, homogenization produces persistent downstream harms: shared security blind spots, lingering bias, and widening disciplinary disparities.

5.1 Correlated Errors: Security Blind Spots in Code Generation

  • MAS reduce overall vulnerability rates by up to 50%, yet amplify deployable security blind spots shared across models.Filtering to code with average reviewer scores ≥3 isolates flaws that evade the panel; generator reintroductions and reviewer overconfidence compound the risk.Source: Figure 4 · p. 8, §5.1 · p. 8

5.2 Persistent Influence: Bias Amplification in Hiring

  • Bias introduced by an adversarial hiring agent spreads through the panel and persists after the prompt is removed.Conformity transmits localized bias, polarization intensifies initial skews, and inertia prevents self-correction, allowing a transient perturbation to shift final hiring decisions.Source: §5.2 · p. 8, Figure 5 · p. 9

5.3 Converging Preferences: Disparity in Scores for Peer Review

  • Mean-score spreads between research disciplines widen over interaction rounds, with the same areas advantaged or disadvantaged across models.Conformity adopts dominant preferences, polarization intensifies them, and inertia hardens disparities that may marginalize disciplines under disproportionately harsh standards.Source: Figure 6 · p. 9, §5.3 · p. 9

6 Sampling Stochasticity and Mixed Configurations

Sampling and mixed-model configurations do not mitigate homogenization risks: sampling adds surface-level variance, while harms persist across mixed-model applications.

Sampling Stochasticity Creates Surface-level Variance

  • Across M = 10 trials, sampling adds behavioral variance but does not alleviate homogenization.The evaluation measures cross-trial mean-score variance for hiring and peer review, and functional similarity for code.Source: §6 · pp. 9–10, Figure 7 · p. 10

Mixed Configurations Suffer from Cross-Family Homogenization

  • Homogenization harms persist across all three applications despite mixed-model configurations.The experiments vary models across generator–reviewer, hiring evaluator, and peer-review roles, yet still find amplified code vulnerabilities, propagated hiring bias, and widening disciplinary disparities.Source: §6 · pp. 9–10, Figure 8 · p. 10

7 Concluding Remarks

The paper identifies homogenization as a MAS failure mode with risks in code generation, hiring, and peer review. Its experiments focus on orchestrator-driven systems and numerical scores.

7 Concluding Remarks

  • Across code generation, hiring, and peer review, homogenization risks shared code vulnerabilities, persistent demographic disparities, and marginalization of research topics.The conclusion says code flaws are shared across model families and could be exploited across codebases; hiring disparities persist after biased-agent removal, while peer review preferences favor certain areas.Source: §7 · p. 10
  • The conclusion frames homogenization as a MAS failure mode and calls for explicit measurement and reporting of interaction dynamics.It introduces homogenization as a novel failure mode of MAS.Source: §7 · p. 10
  • The experiments focus on orchestrator-driven MAS and metrics of numerical scores, leaving alternative architectures and natural-language homogenization for future work.The authors identify both extensions as future-work directions.Source: §7 · pp. 10–11

A Independecy of Homogenization Metrics

The paper treats conformity, polarization, and inertia as independent dimensions of homogenization. Figure 9 illustrates their possible combinations.

A Independecy of Homogenization Metrics

B Applications & Experiment Setup: Additional Details

The appendix provides additional experiment details, with prompts and personas in a separate section. It describes datasets and evaluation procedures for code generation, hiring, and peer review.

B Applications & Experiment Setup: Additional Details

  • The appendix supplies additional experiment details and directs readers to a separate section for prompts and personas.The prompt instructions and personas are not included in this passage.Source: B Applications & Experiment Setup:… · p. 15

B.1 Code Generation

  • Code generation uses 280 adversarially created partial Python and C++ snippets, evaluates vulnerabilities with CodeQL, and scores code similarity through pairwise LLM judgments.Similarity scores average binary outcomes across 10 trials for all 45 pairs.Source: §B.1 · p. 15

B.2 Hiring

  • The hiring experiments use the authors’ own synthetic dataset, whose 40 occupations are listed in Table 2.The passage identifies the dataset as synthetic but leaves its details to Appendix C.Source: §B.2 · p. 15, Table 2 · p. 16

B.3 Peer Review

  • The peer-review sample contains 1,950 ICLR 2025 papers selected using stratified sampling.The sample includes rejected and withdrawn papers, accepted posters, accepted oral and spotlight papers, and desk-rejected papers.Source: §B.3 · p. 16

C Hiring Dataset Generation

The hiring dataset is synthetic and counterfactual, built from US occupations, job descriptions, and matching résumés with demographic signals. Its generation uses five stages.

C Hiring Dataset Generation

  • The dataset-generation pipeline has five stages, ending with demographic signals inserted into matching résumés.The stages are occupation collection, demographic baselines, job-description generation, résumé generation, and demographic-signal insertion.Source: C Hiring Dataset Generation · p. 16

C.1 Collecting Occupation Categories

  • The authors collected 40 US occupation categories from the official government jobs website.The categories span a range of job postings.Source: §C.1 · p. 16

C.2 Gathering Baseline Demographic Distributions

  • Baseline demographic distributions from US labor statistics establish expected occupational-stereotyping directions for adversarial agents and bias evaluation.Source: §C.2 · p. 16

C.3 Generating Job Descriptions Per Occupation

  • Gemini 3.1 Pro generated five job descriptions per occupation, using descriptions from Tan et al. (2026) as in-context examples.The model was selected to avoid overlap with models used in the experiments.Source: §C.3 · pp. 16–17

Job Description Generator

  • The job-description prompt requests varied seniority, agency context, and duties, with specific fields and counts for each description.It requests a JSON array containing job title, duties, and qualifications.Source: Job Description Generator · p. 17

C.4 Generating Matching Résumés

  • The authors used Tan et al. (2026)’s prompt to generate matching résumés for all 200 job descriptions.The résumé-generation prompt requires realistic candidates with specified experience and skills weaknesses, among other constraints.Source: §C.4 · p. 17, Résumé Generator · p. 17

Résumé Generator

  • Each generated résumé follows a prescribed structure and includes four exact placeholders under Additional Information.The required placeholders are languages, activities, volunteering, and hobbies.Source: Résumé Generator · p. 17

C.5 Inserting Demographic Signals

The study creates résumé signals to test demographic bias, varying extracurricular cues across ethnicity and gender combinations. Gemini 3.1 Pro identified the intended stereotype from these signals 100% of the time.

C.5 Inserting Demographic Signals

  • Signals were constructed as languages, activities, volunteering, and hobbies across four ethnicities, two genders, and five variations per combination.This produced 40 counterfactual variations of each data point.Source: §C.5 · p. 17
  • Gemini 3.1 Pro identified the correct stereotype from the résumé signals 100% of the time.The authors use this result to establish the signals’ validity.Source: §C.5 · p. 17
  • Table 3 provides representative examples of demographic signals injected into résumés.Source: Table 3 · p. 18

D Additional Results for Evaluating Homogenization

Additional evaluations find conformity, polarization, and inertia across the tested LLM and mixed-model setups. The patterns match those reported in the main paper.

D Additional Results for Evaluating Homogenization

E Additional Results for Coding

Across coding evaluations, MAS interactions help on average but leave systemic blind spots for particular vulnerabilities. Detailed results cover vulnerability patterns across all nine LLMs.

E Additional Results for Coding

F Additional Results for Hiring

Additional hiring results show ethnicity and gender bias patterns across models, with ethnicity scores also broken down by occupation. An adversarial agent’s influence can persist after removal, especially when introduced earlier.

F Additional Results for Hiring

G Additional Results for Peer Review

Additional peer-review results track how mean scores evolve across rounds and show that preferences become more extreme in several fields.

G Additional Results for Peer Review

H Additional Results for Sampling Stochasticity and Mixed Configurations

Additional results examine mixed MAS setups across coding, hiring, and peer review, including vulnerability lists, score trends, and bias trends. The supplied figures show clear conformity, polarization, and inertia trends for mixed hiring and peer-review setups.

H Additional Results for Sampling Stochasticity and Mixed Configurations

Related papers

Loading 2610.09824v1…