arxiv:2606.05985

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

Published on Jun 4

· Submitted by

xushaoyang on Jun 18

iNLP Lab @ SUTD

Upvote

Authors:

Shaoyang Xu ,

Abstract

Multicultural multi-agent systems exhibit limited value diversity despite cultural alignment, with social interaction reducing diversity and compromising collective decision-making breadth.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

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.

View arXiv page View PDF GitHub 2 Add to collection

Community

beiweixiaoxu

Paper author Paper submitter about 9 hours ago

Multi-agent systems are inherently multicultural as well. This is especially apparent given the recent popularity of Moltbook and the subsequent line of analytical work (e.g., #MoltNet). These build on the premise that different agents represent different users interacting with one another, so the agent system embodies value diversity.

Building on existing work in cultural-alignment evaluation (on single model), we define cultural diversity as an (new) evaluation dimension at the level of the agent system (P2).

It assumes a system of N agents, each representing a different culture. After collecting these agents' responses to the WVS (World Values Survey) questionnaire, we compute the pairwise differences among them to obtain the system's diversity.

Our preliminary evaluation, treating a single backbone as the system (N=5), shows that models fall far short of the value diversity found in human society (P3).

We also evaluated the degree of cultural alignment of these systems and found that alignment and diversity are only weakly correlated, indicating that cultural diversity offers a distinct evaluation perspective (P4).

Beyond single-backbone systems, the N agents are in fact more likely to be initialized from different backbones. We swept through all configurations and found that mixing backbones can (to some extent) improve both alignment and diversity (P5).

By varying the cultural composition of the system (P6, top) and the number of cultural agents (P6, bottom), we found that these changes do not produce meaningful gains in cultural diversity.

More seriously, once agents begin to interact—we tried the simplest form of interaction, in which a cultural agent sees other cultural agents' answers before responding—system diversity drops further (P7, left), and additional rounds of interaction fail to recover this decline (P7, right).

Finally, we initialized two agent systems—one low-diversity and one high-diversity—and had each participate in a democratic decision-making task. We found that the high-diversity system produced a more balanced and equitable voting distribution in terms of building social value (P8).