HERMES

A toolkit for the multi-agent simulation of emergent communication and languages.

HERMES provides the building blocks for running language games: populations of agents, worlds that supply the entities agents talk about, interaction scripts that define who says what to whom, alignment strategies that update agents after each interaction, and monitors that record how the emerging language develops over time.

Installation

pip install hermes-toolkit

The core framework is pure Python. Experiments with neurally grounded agents additionally need PyTorch, and the experiment logger needs pandas:

pip install "hermes-toolkit[all]"      # torch + pandas
pip install "hermes-toolkit[torch]"    # torch only

The distribution is named hermes-toolkit, but the package is imported as hermes:

import hermes
from hermes.experiment import Experiment

Getting started

Worked examples live in docs/source/tutorials:

• hermes_naming_game.ipynb — a basic naming game
• hermes_grounded_naming_game.ipynb — a naming game grounded in images
• hermes-MARL-concept-learning.ipynb — concept learning with neural agents
• hermes_naming_game_pyfcg.ipynb — integration with FCG

Core concepts

Experiments are configured with a dictionary; each component is given as a (class_name, module_name) pair, so your own subclasses drop in without changing the framework.

Citation

If you use HERMES in academic work, please cite it. A CITATION.cff file will be added with the first tagged release.

Contributing

The main branch is protected — please create a feature branch for your contribution and open a merge request against main.

License

Apache License 2.0 — see LICENSE.