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

When we build AI systems with different cultural identities, we usually check if they act like real people from those cultures. This study shows that even if they act correctly individually, they often lose their unique cultural differences when they interact. This leads to a "boring" consensus that lacks the variety found in human societies.

The rise of large language model (LLM) agents has led to the emergence of multi-agent systems. These systems simulate social platforms or collaborative environments. As these systems are deployed in globally diverse settings, researchers have focused heavily on "value alignment." This means ensuring a single agent’s responses match the cultural norms of a specific population. However, this focus overlooks a critical systemic property. It ignores whether a collection of diverse agents maintains its plurality or collapses into a monolithic, homogenized viewpoint.

The blind spot in cultural alignment

Current evaluation paradigms treat culture as a per-agent attribute. Researchers typically measure how closely an agent's response to a social survey matches a target culture's majority view. While this captures "local cultural fidelity," it fails to capture the health of the system as a whole. A system could consist of agents that are all perfectly aligned with their respective cultures. Yet, if those agents all gravitate toward the same "average" answers, the system effectively loses its diversity.

The authors argue that alignment and diversity are not the same thing. As shown in, the two metrics are largely uncorrelated, with a Pearson correlation of only $r = -0.12$.

This means a system can achieve high alignment with human values while remaining internally homogeneous. Such a system is essentially a collection of "correct" but identical voices. Relying solely on alignment creates a blind spot. Developers might believe they have built a multicultural society, when they have actually built a monoculture of well-behaved clones.

Quantifying the spread of ideas

To solve this, the authors propose "value diversity" as a new system-level evaluation axis. They define diversity by how dissimilar the agents are from one another. They implement this through two distinct mathematical lenses:

1. Pairwise Diversity ($Diversity_P$): This calculates the average dissimilarity across every possible pair of agents. It is akin to measuring the average distance between every house in a neighborhood to see how spread out the population is.
2. Structural Diversity ($Diversity_S$): To avoid counting redundant connections, the authors use a Minimum Spanning Tree (MST). An MST is a graph theory concept that connects all points in a network using the shortest possible total edge length. This provides a sharper characterization of the system's global spread.

The researchers applied this framework to 19 cultures and 18 backbone models. They used data from the World Values Survey. By assigning agents specific cultural identities through system prompts, they created a controlled environment. This allowed them to observe how different model architectures influence the collective "value space" of the system.

Homogenization is the default state

The study's findings reveal a significant gap between artificial and human societies. The authors report that none of the 18 single-backbone systems tested reached human levels of diversity. Even the most diverse single-backbone system, gemini-2.5-pro, achieved a $Diversity_P$ of 36.12. This is notably lower than the human reference of 44.07 [Table 1].

Crucially, the authors find that increasing the number of agents does not help. As the agent count grows, the homogenization of these systems becomes increasingly amplified [Figure 3b]. The problem worsens when agents begin to interact. Through "social exposure"—where agents observe the responses of their peers before answering—the authors report an average decrease in diversity ($\Delta D = -1.27$) .

Instead of strengthening their unique cultural identities, the agents drift toward a central consensus. This trend persists over multiple rounds of interaction .

Social interaction in current LLM societies acts as a powerful homogenizing force.

Limits of the simulation

While the study provides a rigorous new metric, several caveats remain. First, the interaction model is simplified. "Social exposure" involves agents observing text responses. This is quite different from the nuanced and long-term social cues that shape human culture. Second, the authors use the majority vote of the World Values Survey as the "cultural prototype." This treats culture as a single, fixed point. It may ignore the internal diversity present within actual human populations.

Furthermore, the study focuses on explicit value statements. It does not explore whether these homogenization tendencies extend to more subtle cultural signals. These include linguistic dialects or emergent social norms in unscripted dialogue. For a practitioner building a complex social simulation, these results suggest a warning. The "consensus" observed in an LLM-based society may be an artifact of the model architecture. It may not reflect true social dynamics.

A verdict for multicultural design

The verdict is clear. If you want to build a truly pluralistic multi-agent system, do not rely on a single model. The authors demonstrate that "mixed-backbone" systems outperform single-backbone systems. These systems use different LLMs for different agents. They perform better on the Pareto frontier of both alignment and diversity .

This suggests that the inherent biases of different model families can help preserve the breadth of a collective system.

Current LLM-based societies are prone to a persistent homogenization. This narrows the scope of collective decision-making .

For engineers moving toward agent-native social platforms, the goal must shift. You must move from ensuring individual agents are "correct" to ensuring the system remains sufficiently diverse. The system must represent the complexity of the real world. Code and data for this framework are available at https://github.com/iNLP-Lab/MultiAgent-Diversity.

Figures from the paper