simulation_streams

Simulation Streams Tech Report

A development platform for simulations with large language models. It comes with an Entity-Component-Systems approach and graphical editor.

The platform is a flask app most easily used locally in a venv, which can be started with source setup.sh (on linux). This sets up all dependencies and ends with instructions for the different commands for launching the editor or command line running. The editor is displayed in a browser.

A library of simulation configs can be found under configs, including a market economy, a social simulation and 6 tasks from the classical reinforcement learning literature.

Usage examples

1 source setup.sh

2a To create the local server that runs the web app:

```
python app.py configs/social_catch_game.py --metrics=configs/metrics_social_catch_game.txt --web --model='gemini-2.0-flash-exp' --api_key='your_key'
```

2b To run a number of steps from the command line and return the metrics:

```
python app.py configs/social_catch_game.py --metrics=configs/metrics_social_catch_game.txt --steps=10 --model='gemini-2.0-flash-exp' --api_key='your_key'
```

2c To open the editor with an empty config:

```
python app.py --web --model='gemini-2.0-flash-exp' --api_key='your_key'
```

Citing Simulation Streams

If you use Simulation Streams in your work, please cite the accompanying article:

@article{sunehag2025simulation,
  title={Simulation Streams: A Programming Paradigm for Controlling Large Language Models and Building Complex Systems with Generative AI.},
  author={Sunehag, Peter and Leibo, Joel Z},
  journal={arXiv preprint arXiv:2501.18668},
  year={2025}
}

Disclaimer

This is not an officially supported Google product.