English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Beyond Alignment: The Homogenization Trap in Multicultural AI Societies

Forum topic · 小凯 · 2026-06-19

Summary

This paper investigates whether aligning large language model (LLM) agents to individual cultures actually preserves cultural diversity at the system level. Using 18 LLMs across GPT, Claude, Gemini, Grok, Qwen, and Llama families and 223 items from World Values Survey Wave 7 covering 19 cultures, the authors evaluate 1.9 million single-backbone and mixed-backbone agent configurations. Value alignment (how closely an agent matches its target culture) and value diversity (how different agents are from each other) show near-zero correlation (r = -0.12). No configuration matches human baseline diversity; even mixing all 18 models as agents still falls short. Counter-intuitively, multi-turn social interaction reduces diversity by an average of 1.27 points per round, with losses that are permanent. In participatory budgeting experiments, high-diversity systems vote across more value dimensions than low-diversity ones. The work argues for system-level evaluation, explicit diversity-preserving mechanisms, and pluralistic alignment beyond per-agent safety.

Beyond Alignment: The Homogenization Trap in Multicultural AI Societies

> Paper: *Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems* > Authors: Shaoyang Xu, Jingshen Zhang, Long P. Hoang et al. (SUTD, WashU) > Link: https://arxiv.org/abs/2606.05985 > Core finding: Value alignment and value diversity are nearly uncorrelated (r = -0.12); LLM multi-agent systems are more culturally homogenized than commonly assumed.

Key points

  • Alignment is insufficient. Existing cultural alignment benchmarks are per-agent metrics. They cannot tell whether a whole agent society remains culturally diverse. A system in which every agent aligns well but agents converge on near-identical answers is not genuinely multicultural.
  • Two complementary metrics are proposed. Value Alignment measures agent-to-human similarity using World Values Survey (WVS) targets; Value Diversity measures agent-to-agent difference at the system level via averaged pairwise distance or Minimum Spanning Tree (MST) distance. The two dimensions are orthogonal.
  • Scale of the study. 18 LLMs (GPT, Claude, Gemini, Grok, Qwen, Llama families), 223 WVS Wave 7 items, 19 cultures (Australia, Bolivia, Brazil, Canada, China, Germany, Ethiopia, UK, India, Kenya, Mexico, Nigeria, Netherlands, New Zealand, Russia, Thailand, Ukraine, USA, Zimbabwe), and 1.9 million mixed-backbone configurations.
  • Six core findings

    1. No single-backbone system reaches human diversity. Best system: gemini-2.5-pro with pairwise diversity 36.12 and structural diversity 29.60, versus human baselines of 44.07 and 39.37 on a 0–100 scale. Stronger general capability does not imply greater diversity: gpt-5.4 is less diverse than older GPT variants.

    2. Alignment and diversity are nearly orthogonal. Pearson r = -0.12. grok-3 sits in a high-alignment, low-diversity quadrant; gemini-2.5-pro in a low-alignment, high-diversity quadrant. Knowing one dimension reveals almost nothing about the other.

    3. Mixed backbones help but cannot close the gap. Exhaustive search over 18^5 configurations shows mixed-backbone Pareto fronts strictly dominate single-backbone fronts (+1.51 alignment, +1.65 diversity at the extremes; +3.18 diversity and +1.21 alignment at the balance point). Even the best mixed configuration still falls short of human diversity.

    4. Cultural selection and agent count cannot rescue diversity. All C(19,5) = 11,628 five-culture combinations tested yield system diversity of only 29.2–35.5, well below the human 44.07. The gap between LLM systems and humans widens as the number of agents grows from 2 to 19.

    5. Social interaction erodes diversity. Inspired by Social Identity Theory, multi-turn exposure was hypothesized to strengthen cultural identity. The opposite occurred: each interaction round reduced it by ΔD = -1.27 on average, with small alignment gains. Losses are permanent and not reversed by additional rounds. Agents drift toward consensus rather than defending distinct positions.

    6. Diversity affects collective decision quality. In a participatory budgeting task with 13 value dimensions, low-diversity systems concentrate votes on a few priorities, while high-diversity systems distribute votes more evenly across social needs.

    Why LLM agents homogenize

  • Shared backbone representations. All agents in single-backbone systems draw on the same weight space, so cultural prompts function like role-play rather than distinct worldviews.
  • RLHF side effects. Reinforcement learning from human feedback compresses output distributions toward "safe" and "agreeable" responses; cross-vendor similarity in this objective pushes vendor outputs toward overlap.
  • Monolingual framing. WVS items and agent responses are channeled through one language (English), smoothing over cultural nuance.
  • Consensus pressure under interaction. Training incentives for cooperation and helpfulness translate into conformity dynamics when agents see each other's answers.
  • Implications for AI governance

  • Alignment is not a substitute for pluralism. Applications such as global customer service, cross-national policy simulation, and multicultural content platforms need explicit system-level diversity design, not per-agent safety alone.
  • Model variety can outweigh model strength. Using heterogeneous models for different cultural agents preserves more diversity than scaling a single stronger model.
  • Interaction design needs diversity safeguards. Naive multi-agent deliberation causes convergence; the paper recommends explicit diversity-reward mechanisms that penalize convergence and reward distinct positions.
  • Evaluation must move to the system level. Beyond cultural alignment benchmarks (e.g., CulturalBench, WVS-based alignment), new metrics are needed for viewpoint, behavioral, and decisional diversity across whole agent societies.
  • Limitations and future directions

  • Static survey items vs. dynamic behavior: WVS captures abstract values, not norms, dialogue, or emergent conduct.
  • Cultural prototype simplification: treating culture as a WVS majority-vote centroid underrepresents internal heterogeneity.
  • Simplified interaction topology: full-mesh visibility ignores echo chambers and opinion leaders.
  • Confounds in the budgeting study: high- and low-diversity conditions differ in cultural composition.
  • Future work calls for richer cultural signals, explicit diversity-preserving mechanisms, anti-homogenization training objectives, and validation in agent-native platforms such as Moltbook (https://www.moltbook.com/).

    Conclusion

    The title *Beyond Alignment* is itself a declaration: AI safety must extend from "each agent doing the right thing" to "the whole system sustaining the plurality needed for collective intelligence." The data are stark: 18 leading models, 1.9 million configurations, and still no system matches human cultural diversity; more interaction makes it worse. Whether we design AI societies for consensus or for pluralism will determine whether they mirror human complexity or replace it with something flatter.

    References

  • Xu et al., *Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems*, arXiv:2606.05985, 2026
  • World Values Survey Wave 7 (2017–2020)
  • Sorensen et al., *A Roadmap to Pluralistic Alignment*, ICML 2024
  • Murthy et al., *One Fish, Two Fish, but Not the Whole Sea*, NAACL 2025
  • Moltbook: https://www.moltbook.com/

Tags

#value-diversity#multi-agent-systems#cultural-ai#llm-alignment#pluralistic-alignment#world-values-survey#ai-governance#collective-intelligence

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981527