CoevolveSim

Code and simulation data for Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models:

@misc{savcisens2026coevolvesim,
      title={Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models}, 
      author={Germans Savcisens and Samantha Dies and Courtney Maynard and Tina Eliassi-Rad},
      year={2026},
      doi={10.48550/arXiv.2607.27512},
      publisher={arXiv},
      url={https://arxiv.org/abs/2607.27512}, 
}

CoevolveSim is a framework for studying belief diffusion within networked LLM populations. Generalist and specialist LLM agents are placed on a social network (Erdős–Rényi or Watts–Strogatz) and exchange beliefs about medical-indication statements over several rounds, each agent revising its belief after observing a summary of its neighbors' beliefs. Across four scenarios:

• I. baseline generalists,
• II. generalists with random social roles,
• III. specialists with random roles,
• IV. specialists with roles matched to their domain.

These simulations isolate the effects of persona-style role assignment, domain specialization (model heterogeneity), and role–specialization alignment on individual belief revision and population-level consensus. A hierarchy of classical opinion-dynamics surrogate models (M1–M4) is then fit to test which mechanisms (persistence, social belief composition, agent identity) are needed to reproduce the observed dynamics.

Project structure

1. src: simulation framework, agent/network configs, and analysis code (installed as an editable package).
2. tests: unit tests.
3. data: raw simulation output (data/outputs/) and derived analysis tables/figures (data/analysis/).
4. notebooks: cleaned, documented notebooks that reproduce every table and figure in the paper; see Notebooks below.

Notebooks

notebooks contains the documented, reproducible pipeline behind the paper's results. Each notebook's own intro cell states exactly which figures/tables/sections it produces:

1. maximin_selection.ipynb: reproduces the maximin selection of the 16 network realizations and 20 discussion statements used across all runs.
2. sanity_check.ipynb: check that all data is in place
3. X1_data.ipynb: turns raw per-run simulation output (data/outputs/runs/) into the two aggregated tables every later notebook builds on (agent_level_data.parquet, run_level_data.parquet).
4. X2_agent_analysis.ipynb: agent-level analysis (§What drives belief revisions?): estimated marginal means and planned contrasts for plasticity, directedness, and outgoing influence across the four scenarios and two network types, plus the variance-decomposition/ICC analysis behind opinion leaders and followers.
5. X3_run_analysis.ipynb: population-level analysis: estimated marginal means and planned contrasts for consensus change across the four scenarios and two network types (Fig. 2C), plus per-scenario convergence rates.
6. X4_manuscript_plots.ipynb: combines the X2/X3 outputs into the manuscript-ready tables and the combined contrast-forest figure.
7. X5_surrogates.ipynb: fits/pools the M1–M4 surrogate opinion-dynamics models (persistence; +global belief composition; +local neighborhood composition; +agent identity) and produces the final-state MCC and consensus-fidelity figures/tables (§Can classical opinion-dynamics models explain these dynamics?).

Python environment

This project uses uv for Python environment management. Run the setup script for initial configuration:

./setup.sh          # Install dependencies and configure environment

If you have uv installed, just run uv sync from this directory.

Run Experiments

Canonical entrypoint:

uv run python src/experiment.py

Example run with Hydra overrides:

uv run python src/experiment.py \
    catalog=random_roles \
    prompt=wR_L \
    network=erdos-renyi \
    network.params.n=48 \
    network.params.p=0.3 \
    statement=false_0 \
    probe=zeroshot \
    seed=814183 \
    experiment.max_rounds=10

Note

• Some model configurations require a Hugging Face access token in src/configs/model/*.yaml.
• A subset of analysis notebooks relies on R/rpy2 tooling (see notebooks/r_utils.py and notebook comments for details).

📃 Licenses

Important

This code is licensed under the MIT License. See LICENSE for more information.

Warning

1. This is research software. While we strive for correctness and reproducibility, please verify results for your specific use case.
2. GitHub Copilot and Claude Code contributed to code annotations, docstrings, and formatting. All algorithmic logic, methodological design, and scientific claims were developed and reviewed by the authors.

Correspondence: g.savcisens@northeastern.edu