MultiAgent-Value-Diversity

The official code of Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems.

Set up

Python 3.12.0.

Install the packages via

conda env create -f environment.yml

or

pip install -r requirements.txt

The data used in experiment are in data folder: WVS questions, WVS ground truth, and project file used in section7.

The WVS data are adopted from the CultureSPA project: https://github.com/shaoyangxu/CultureSPA

Metric Implementation

The evaluate.py implements all the metrics: Value Alignment, Pairwise Diversity and Structural Diversity.

And you can run it to evaluate systems, before this, you should choose the target model and output_dir below the if __name__ == "__main__": line.

Cultural Agent Inference (Section 5)

For systems with GPT, Claude, Gemini, Grok, Llama models as backbones, you can use the sec5_infer_api.py to do inference, which will produce the cultural agent outputs in the wvs_evaluation folder.

Below is an example (inference for all cultural agents initalized with gpt-5.4):

for c in AUS BOL BRA CAN CHN DEU ETH GBR IND KEN MEX NGA NLD NZL RUS THA UKR USA ZWE
do
    python sec5_infer_api.py --culture $c --model gpt-5.4
done

For Qwen backbones, you can simply use the sec5_infer_torch.py by specify the same input parameters (culture/model).

wvs_evaluation folder provides all the outputs, and you can just run evaluate.py on them to produce all the results in Table 1.

System Homogenization Analysis (Section 5.1/5.2/5.3)

The analysis code for each subsection is introduced as below:

• The relationship between Alignment and Diversity (Section 5.1): sec5_fig1_left.py; with per-question breakdown: sec5_fig1_right.py
• Mixed Backbones (Section 5.2): sec5_fig2.py
• Cultural Composition and Agent Count (Section 5.3): sec5_fig3.py

Multi-round Social Exposure (Section 6)

sec6_social_exposure_api.py and sec6_social_exposure_torch.py provide the first-round social exposure for GPT, Claude, Gemini, Grok, Llama Systems, and Qwen Systems respectively.

For example, you can run:

for c in BRA CHN MEX NGA NZL
do
    python sec6_social_exposure_api.py --culture $c --model gpt-5.4
done

to generate the results of gpt-5.4 in wvs_evaluation_interaction folder.

---

After the first-round interaction, you can further run the sec6_multi_turn_api.sh and sec6_multi_turn_torch.sh to conduct additional rounds of interaction, like:

bash sec6_multi_turn_api.sh <model> <max_round>

(max_round = 5 in the paper) The outputs here will be stored in wvs_evaluation_interaction_round{round_id}

---

And you can use sec6_fig4.py and sec6_fig5.py to summarize the results of first-round social exposure and the multi-round.

Collective Decision-Making (Section 7)

Run the sec7_vote.py to do the Decision-Making in WVS-Participatory Budgeting task:

for c in AUS BOL BRA CAN CHN DEU ETH GBR IND KEN MEX NGA NLD NZL RUS THA UKR USA ZWE; do
  python sec7_vote.py --culture $c --model claude-opus-4.7 --n_runs 20
done

Here, we run for all cultural agents. And the sec7_fig6.py further pick the lowest- and highest-diversity system for outcome comparision.

Others

choose_culture.py evaluates human diversity across all (18^5) cultural combinations. Based on this, we selected BRA, CHN, MEX, NGA, and NZL for the main experiments, as this combination exhibits top diversity.

Citation

If you find this repo useful, please cite:

@article{xu2026beyond,
  title={Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems},
  author={Xu, Shaoyang and Zhang, Jingshen and Hoang, Long P and Li, Jinyuan and Zhang, Wenxuan},
  journal={arXiv preprint arXiv:2606.05985},
  year={2026}
}