Sovereign 8s.is mirror of standardgalactic/llm_opinion_dynamics
  • nesC 99.8%
  • Python 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
VincentBrockers 7defbe3875
Revise data references and repository instructions
Updated README to reflect changes in data storage and access links.
2026-07-23 23:22:25 +02:00
data removed large file 2026-07-23 14:32:50 +02:00
envs Align code and figures with revised manuscript 2026-05-27 10:39:36 +02:00
figures improved discussion flowchart 2026-07-23 23:06:41 +02:00
src aligned figure numbering with revised manuscript 2026-07-23 23:04:56 +02:00
.gitignore Initial commit 2025-08-29 10:58:18 +02:00
LICENSE Initial commit 2025-08-29 10:58:18 +02:00
README.md Revise data references and repository instructions 2026-07-23 23:22:25 +02:00

LLM Opinion Dynamics

Code and data for the manuscript:

Disentangling Interaction and Bias Effects in Opinion Dynamics of Large Language Models
Vincent C. Brockers, David A. Ehrlich, Viola Priesemann, 2026

This repository contains the simulation, Bayesian inference, parameter recovery, and plotting code used to analyse opinion dynamics in interacting large language model agents. It is intended to support inspection and reproduction of the analyses reported in the manuscript.

Repository structure

llm_opinion_dynamics/
├── data/
│   ├── mixtral_tuned_clim_data.csv
│   ├── recovery/
│   │   ├── recovery_data/
│   │   └── recovery_reports/
│   ├── traces/
│   │   ├── full_model/
│   │   └── idv_model/
│   ├── unprompted_opinion_priors/
│   └── variance_explained/
│       ├── loo_r2_values/
│       └── nc_values/
├── envs/
│   ├── bayes_env.yml
│   └── vllm_env.yml
├── figures/
│   ├── figure_1.pdf
│   └── ...
├── src/
│   ├── inference.py
│   ├── inference_idv.py
│   ├── simulate-api.py
│   ├── simulate-vllm.py
│   ├── plotting/
│   └── recovery/
├── LICENSE
└── README.md

The data required for full functionality of this repository are archived separately at: https://doi.org/10.25625/MFVLDG

This record contains both the processed inference data used for the Bayesian analyses and figure generation, and the full simulated discussion outputs with associated metadata.

To reproduce the analyses, download the dataset and place or extract its contents into the repository’s data/ directory. Posterior traces are stored in data/traces/. Exported manuscript figures are provided in figures/, and figure-generation scripts are in src/plotting/.

Large fine-tuned model checkpoints are available at: https://doi.org/10.25625/4EKOMS

Installation

Clone the repository and keep the directory structure unchanged, as several scripts use relative paths.

git clone https://github.com/Priesemann-Group/llm_opinion_dynamics.git
cd llm_opinion_dynamics

Create the conda environments:

conda env create -f envs/bayes_env.yml
conda env create -f envs/vllm_env.yml

The environment names defined in the files are:

conda activate pymc   # Bayesian inference, recovery, plotting
conda activate vllm   # local vLLM simulations and API simulations

Reproducing the analyses

Bayesian inference

The main hierarchical Bayesian model can be run from src/:

conda activate pymc
cd src

python inference.py --llm Llama-3.1-8B-Instruct

The script accepts additional sampler and model-ablation arguments. Inspect available options with:

python inference.py --help

The climate-only model with individual attractors is implemented in:

src/inference_idv.py

Parameter recovery

Parameter-recovery scripts and configuration files are in:

src/recovery/

Corresponding recovery data and reports are stored in:

data/recovery/

Plotting

Manuscript figure PDFs are stored in figures/. Plotting scripts are in src/plotting/:

conda activate pymc
cd src/plotting

python figure_3.py

Analogous scripts are provided for the remaining plotted figures.

Running new simulations

Two simulation backends are provided.

API-based simulations

src/simulate-api.py supports OpenAI-compatible APIs. Set the required API key before running:

export OPENAI_API_KEY=<your_key>
# or, for xAI:
export XAI_API_KEY=<your_key>

Example:

conda activate vllm
cd src

python simulate-api.py \
  --backend openai \
  --model gpt-4o-mini \
  --topic "Climate Change" \
  --samples 25 \
  --length 5 \
  --output ../data/example_api_run.csv

Local vLLM simulations

For local Hugging Face / vLLM inference, use:

conda activate vllm
cd src

python simulate-vllm.py \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --topic "Climate Change" \
  --samples 25 \
  --length 5 \
  --gpu 0 \
  --output ../data/example_vllm_run.csv

Local simulations require a CUDA-compatible GPU. Depending on model size and hardware, full sequential simulations can take several hours.

Notes on reproducibility

The repository includes processed data, posterior traces, recovery outputs, and final figure PDFs, so the reported analyses can be inspected without rerunning all LLM simulations. Rerunning simulations may produce small numerical differences because LLM sampling and remote API backends can be stochastic or version-dependent.

License

This repository is released under the BSD 3-Clause License.