PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Shaoyang Xu, Jingshen Zhang, Long P. Hoang, Jinyuan Li, Wenxuan Zhang

arXiv:2606.05985v1cs.CLcs.CY

TL;DR

Existing cultural evaluation measures how closely individual agents align with target cultures but does not capture whether a multicultural system preserves cultural plurality. This paper defines value diversity from dissimilarity among culturally conditioned agents’ survey responses and finds that it complements alignment, while current systems remain less diverse than human societies and social interaction further homogenizes them.

  • Problem

    Per-agent value alignment is insufficient because strongly aligned agents may still collapse toward a homogeneous value space, motivating a distinct system-level evaluation axis.

  • Method

    The framework evaluates value diversity by aggregating dissimilarities between culturally conditioned agents’ responses to a shared value survey, alongside alignment against World Values Survey majority-vote cultural references.

  • Results

    Value diversity is largely uncorrelated with alignment, and all single-backbone systems fall below the human diversity reference; mixed backbones narrow but do not close the gap, while interaction erodes diversity toward consensus.

  • Takeaways & Limitations

    Value diversity is a distinct and currently unmet evaluation challenge for multicultural multi-agent systems, with homogenization narrowing collective decision-making.

  • Takeaways & Limitations

    The interaction and participatory-budgeting experiments use simplified settings, and their low- and high-diversity configurations differ in cultural composition as well as measured diversity.

Abstract

from arXiv · show

Multicultural multi-agent systems are increasingly deployed in globally diverse settings, where different agents are grounded in different cultural backgrounds. Existing cultural evaluation focuses on value alignment: how closely a single agent matches a target culture. Yet alignment is a per-agent property and cannot reveal whether a system, taken as a whole, preserves the cultural plurality it is meant to represent. We propose value diversity as a system-level evaluation axis for multicultural agent systems, defined through the dissimilarity between culturally conditioned agents' responses on a shared value survey. Using the World Values Survey, we evaluate 19 cultures and 18 backbone models across a wide range of system configurations. We find that diversity is largely uncorrelated with alignment, indicating that the two capture complementary system properties, and that current multicultural agent systems fall substantially below human societies in value diversity. Mixed-backbone systems narrow this gap but do not close it, and the gap persists across culture compositions and agent scales. Social interaction further erodes diversity by driving agents toward consensus, and a participatory budgeting case study shows that this homogenization narrows the breadth of collective decision-making. Together, our results establish value diversity as a distinct evaluation axis for multicultural multi-agent systems and reveal a persistent homogenization tendency in current LLM-based societies. Our code and data are publicly available at https://github.com/iNLP-Lab/MultiAgent-Diversity.

1 Introduction

The paper introduces value diversity as a system-level complement to per-agent cultural alignment and evaluates it across multicultural agent systems. Current systems remain less culturally diverse than human societies, with homogenization persisting across configurations and increasing through interaction.

  • Value diversity measures dissimilarity among culturally conditioned agents’ survey responses, treating plurality as a collective system property rather than an individual attribute.The framework aggregates agent differences through pairwise or structural averaging.
  • 36.12 versus 44.07: the best single-backbone system remains below the human diversity reference.None of the 18 single-backbone systems reaches human diversity, and stronger backbone capability does not guarantee greater diversity.
  • Pearson r = −0.12 between diversity and alignment shows that the two metrics capture largely independent system properties.Several highly aligned systems remain internally homogeneous.
  • Mixed-backbone systems improve alignment and diversity, but their Pareto frontier still leaves a gap to human-level diversity.The study exhaustively examines approximately 1.89M mixed-backbone configurations.
  • Changing culture selection or agent count does not improve system-level diversity, while larger systems amplify homogenization.The result holds across the evaluated culture compositions and agent scales.
  • Social exposure slightly improves cultural fidelity but drives agents toward consensus, reducing collective plurality; further interaction rounds do not restore it.In participatory budgeting, higher-diversity systems cover broader societal priorities and produce more pluralistic resource allocations.

2 Related Work

Prior cultural evaluation studies largely measure how individual models align with cultural values or knowledge. This paper shifts evaluation to multicultural systems and addresses the open question of their cultural behavior through value diversity.

  • Cultural-value studies commonly use surveys such as WVS or Hofstede to measure alignment with a specific culture at the per-model level.
  • A separate research strand evaluates culture-specific commonsense and norms, also focusing on individual models rather than whole systems.
  • Multi-agent LLM research includes capability-oriented debate and collaboration alongside social-simulation studies of emergent agent-society behavior.
  • How to evaluate the cultural behavior of multi-agent systems remains open; this work addresses the gap through value diversity.

3 System-Level Value Diversity

The framework extends cultural evaluation from individual-agent alignment to system-level value diversity, measuring how dissimilar culturally conditioned agents’ survey responses are. It uses pairwise and structural diversity metrics to characterize collective plurality.

  • Evaluation framework: Each culturally assigned agent answers World Values Survey questions, producing a response vector used for alignment and diversity evaluation.Cultural references are based on majority-vote population responses, representing each culture’s prototypical value orientation.
  • Conceptual distinction: Value alignment measures agent-to-human similarity, whereas value diversity measures agent-to-agent dissimilarity within a multicultural system.These properties can vary independently: agents may remain highly aligned while collapsing toward homogeneous responses.
  • Pairwise Diversity: Pairwise Diversity averages dissimilarity across all agent pairs.The metric is defined from the response vectors of two agents and aggregates their pairwise distances across the system.
  • Structural Diversity: Structural Diversity averages the N −1 distances in the minimum spanning tree of the pairwise distance graph.This discounts geometrically redundant inter-agent relations and provides a sharper characterization of global system-level spread.
  • Human reference: Human reference diversity is computed by applying the same diversity metrics to majority-vote response vectors for each culture.This creates a comparable human benchmark for the culturally grounded agent system.

4 Experimental Setup

The experiments use World Values Survey data from 57 countries, select 19 culturally diverse countries, and evaluate 18 LLM backbones. Systems contain five culturally assigned agents under either shared or mixed backbone configurations.

  • Survey: The study retains 223 of 260 Wave 7 World Values Survey questions after excluding 37 daily-life-specific items.Wave 7 collected human responses from 57 countries between 2017 and 2020.
  • Cultures: The evaluation selects 19 cultures spanning multiple continents from the 57 countries represented in WVS Wave 7.The selected cultures include countries across Africa, Asia, Europe, North America, South America, and Oceania.
  • Backbone models: The experiments use 18 LLMs from the GPT, Claude, Gemini, Grok, Qwen, and Llama families as backbone models.API-based models use temperature 1.0 and top-p 1.0, while Qwen and Llama models use temperature 0.6.
  • System configuration: Main systems contain N = 5 agents assigned BRA, CHN, MEX, NGA, and NZL cultural identities.These cultures were selected for substantial real-world differences, while other cultures support further analysis.
  • System configuration: The study compares single-backbone systems, where all agents share one model, with mixed-backbone systems using different models across agents.This design tests whether backbone heterogeneity changes multicultural system behavior.

5 Main Diversity Results

Single-backbone multicultural systems remain less diverse than human societies, while diversity captures information largely independent of alignment. Mixed backbones improve the diversity–alignment frontier, but culture selection and agent scaling do not close the diversity gap.

  • Single-backbone systems: 36.12 versus 44.07: the most diverse single-backbone system remains below the human diversity reference.Across all 18 single-backbone systems, none reaches the human level.
  • Diversity and alignment: Pearson r = −0.12: diversity and alignment show almost no correlation across systems.Systems can be highly aligned yet internally homogeneous, or diverse yet less aligned with humans.
  • Mixed backbones: Mixed-backbone systems strictly dominate the single-backbone Pareto frontier across the diversity–alignment plane.At the alignment-optimal end, mixed backbones improve alignment by ∆A = +1.51; at the diversity-optimal end, they improve diversity by ∆D = +1.65.
  • Agent count: As agent count increases, the disparity between LLM-based systems and human societies becomes increasingly pronounced.Culture selection alone provides limited gains, while increasing agent count amplifies initial homogenization.

6 Towards Dynamic Interaction

Social exposure reduces system-level diversity rather than reinforcing culturally distinctive positions. Across multiple rounds, homogenization persists and no system returns to its static diversity baseline.

  • One-round exposure: Every system loses diversity after one round of social exposure.Across representative backbones from six model families, the average decrease is ∆D = −1.27, while alignment generally rises by a much smaller margin.
  • One-round exposure: Agents drift toward consensus under social exposure, reducing the system’s collective plurality.Social exposure slightly improves per-agent cultural fidelity but does not reinforce culturally distinctive positions.
  • Multi-round interaction: No system recovers its round-0 diversity after multiple interaction rounds.Some systems stabilize after the first round, while many continue losing diversity with moderate fluctuations.
  • Multi-round interaction: Static diversity measurements may overestimate diversity in realistic interactive deployments.This follows from the persistent homogenization induced by social exposure.

7 How Diversity Shapes Collective Decision-Making

The participatory budgeting case study links system-level value diversity to collective prioritization. Low-diversity systems concentrate approvals on a few societal dimensions, whereas high-diversity systems cover substantially broader priorities.

  • Participatory Budgeting: Low-diversity systems collapse approvals onto only a few societal dimensions.The comparison uses systems with the same claude-opus-4.7 backbone and five agents.
  • Participatory Budgeting: High-diversity systems produce substantially broader societal coverage in collective decisions.Each agent votes for four of 13 projects, with repeated samples aggregated into vote-frequency distributions.
  • Participatory Budgeting: The case study suggests that value diversity may shape downstream collective behavior, not merely characterize system composition.The observed behavioral difference is broader societal-priority coverage by the more diverse system.

8 Conclusion

The paper proposes value diversity as a system-level complement to per-agent alignment for multicultural multi-agent systems. Experiments show persistent shortfalls from human diversity, further erosion through interaction, and narrower collective decisions under homogenization.

  • Conclusion: Value diversity is proposed as a system-level evaluation axis complementing per-agent value alignment.The paper evaluates this axis across 19 cultures, 18 backbones, and millions of system configurations.
  • Conclusion: All single-backbone systems fall below the human diversity reference, while mixed backbones narrow but do not close the gap.The gap persists across culture selection and agent scaling.
  • Conclusion: Social interaction erodes diversity by driving agents toward consensus, narrowing collective decision-making.The participatory budgeting case study connects this homogenization with reduced breadth of societal priorities.
  • Conclusion: Value diversity remains a distinct and currently unmet challenge for LLM-based societies.The conclusion frames this as a persistent homogenization tendency in increasingly realistic multicultural multi-agent systems.

Limitations

The study’s evaluation is constrained by simplified interaction and decision-making settings, prototype-based cultural references, and a minimal abstraction whose generalizability remains open.

  • Participatory budgeting and social-exposure experiments use simplified settings that do not fully represent realistic interactions, networks, or policy environments.The low- and high-diversity budgeting configurations also differ in cultural composition, so composition and measured diversity are not fully disentangled.
  • Cultural comparisons treat WVS majority votes and culturally conditioned agent responses as prototypical value orientations.This reference choice compares population aggregates with model-generated cultural responses.
  • Single LLM samples imperfectly approximate cultural prototypes, while WVS majority votes aggregate responses from large human populations.This asymmetry may affect absolute comparability between human and model diversity scores.
  • WVS-7 responses were collected between 2017 and 2020, so scores represent a temporal snapshot rather than a timeless cultural reference.
  • The framework’s minimal abstraction leaves open whether reported homogenization patterns generalize to richer cultural signals and emergent agent behaviors.The paper identifies everyday dialogue, normative reasoning, and agent-native platforms as future settings.
  • The minimal prompt omits detailed demographic profiles, behavioral descriptions, and culture-specific stereotypes, primarily activating the backbone’s internal cultural representation.

A.2 Implementation of Diversity Metrics

The paper implements pairwise diversity by normalized response distances and structural diversity by aggregating those distances through a minimum spanning tree.

  • Structural Diversity: Structural Diversity is computed from the pairwise distance dictionary and agent list using minimum spanning tree extraction.The implementation uses SciPy’s graph library for MST extraction.
  • Pairwise Diversity: Pairwise Diversity takes agent response vectors and WVS question metadata, compares each agent pair on shared valid questions, and returns normalized distances.Questions absent from the reference or offering no variation are skipped.
  • Structural Diversity: The implementation records pairwise distances in a dictionary before constructing the structural-diversity calculation.

C Additional Results on Diversity–Alignment Relationship

Additional results further examine diversity–alignment dynamics, interaction, and collective decisions across metrics, backbones, and task configurations.

  • Diversity–Alignment Relationship: The correlation between Alignment and Structural Diversity remains weak, supporting their distinction as system properties.Figure 7 reports this relationship across single-backbone multicultural agent systems.
  • Diversity–Alignment Relationship: Structural Diversity more sharply separates genuine structural plurality from redundant pairwise disagreement.Several highly aligned systems show lower structural diversity despite reasonable cultural alignment.
  • Multi-turn Alignment Dynamics: Social exposure produces no systematic alignment trend across rounds, while diversity consistently declines.Most systems fluctuate near round-0 alignment, and the diversity loss is larger and more consistent.
  • Multi-turn Alignment Dynamics: During interaction, homogenization changes collective plurality more than per-agent cultural fidelity.Alignment-based evaluation alone would not detect this system-level effect.
  • Participatory Budgeting: Participatory budgeting uses 13 projects derived from WVS value dimensions, with agents selecting exactly four projects for funding.The projects translate abstract value dimensions into concrete public-policy initiatives.
  • Participatory Budgeting: High-diversity systems distribute support across substantially broader societal dimensions in both gpt-5.4 and gemini-3.1-flash-lite-preview configurations.Both backbone-specific results are presented as consistent with the main reported trend.
Loading 2606.05985v1…