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

## The Claim

Multicultural multi-agent systems—where different AI agents represent different cultures—are failing to preserve cultural plurality. The authors show that current systems are far less diverse in values than actual human societies, and that diversity is a fundamentally different property from the standard metric of "alignment" (how well an agent matches its target culture). This gap matters: when agents interact, they converge toward consensus, which narrows the range of viewpoints available for collective decisions.

## Why Alignment Isn't Enough

Most research evaluates multicultural systems by checking how closely each individual agent matches its assigned culture—a per-agent property. This misses a crucial system-level question: does the *collection* of agents, working together, actually represent the cultural diversity they're supposed to embody?

The authors introduce **value diversity** as a distinct metric: measure how dissimilar agents' responses are on shared questions about values (drawn from the World Values Survey, a real-world dataset covering 19 cultures). High diversity means agents give meaningfully different answers; low diversity means they're saying the same things despite coming from different cultural backgrounds.

The key empirical finding: diversity and alignment are largely uncorrelated. You can have an agent that perfectly matches its target culture (high alignment) in a system that is nonetheless culturally homogeneous. The two properties measure different aspects of system behavior.

## What the Evidence Shows

The authors tested this across 18 language models and various system configurations. Their findings:

- **The diversity gap is large.** Current multicultural agent systems fall "substantially below human societies" in value diversity—they quantify this, though the abstract doesn't specify the exact margin. - **Mixing backbones helps, but not enough.** Using different model architectures for different cultures narrows the gap but doesn't close it. The problem persists regardless of how many cultures you include or how many agents you deploy. - **Interaction erodes diversity.** When agents participate in social interactions (e.g., discussion, voting), they converge toward consensus. This sounds reasonable in isolation, but it's a problem: homogenization shrinks the set of perspectives that influence collective decisions. - **Real-world consequences.** A case study using participatory budgeting—a decision-making process where stakeholders vote on resource allocation—shows that this diversity loss narrows *what kinds of projects get considered*. Fewer viewpoints means fewer kinds of needs get represented.

## Why It Matters

Deploying multicultural systems without tracking diversity could be worse than having no diversity at all. If stakeholders believe their system represents multiple cultures, but agents actually converge to a single consensus perspective, the system legitimizes decisions that may exclude important viewpoints—while creating the false appearance of inclusivity.

The paper establishes value diversity as a measurable, important metric separate from alignment. It also identifies a systematic problem: LLM-based agents, regardless of their training or initial cultural grounding, tend to homogenize when put in interaction. This isn't a bug in any single model; it's a structural tendency of current systems.

The authors have released code and data, making the evaluation framework available for future work on solving the homogenization problem.