Algorithmic Prompt Generation for Diverse Human-like Teaming and Communication with Large Language Models

This repository is the official implementation of "Algorithmic Prompt Generation for Diverse Human-like Teaming and Communication with Large Language Models".

For more information, refer to the following links:

• arXiv

Contents

• Manifest
• Installation
  • Repository Setup
  • LLM Setup
• Instructions
  • Logging Directory Manifest
  • Running an Experiment
  • Reloading
• Reproducing Paper Results
• License

Manifest

• config/: hydra configuration files.
• src/: Python implementation and related tools.
• overcooked_ai/: Overcooked (https://github.com/HumanCompatibleAI/overcooked_ai) submodule

Installation

Repository Setup

Create a virtual environment in python 3.9 and download dependencies from requirements.txt.

Next, create an editable package for the overcooked environment

cd overcooked_ai
pip install -e .

LLM Setup

We provide some additional details on how to setup the repository to work with the following language model frameworks:

• HuggingFace: Ensure that you set the environment variable of HF_TOKEN to the appropriate authentication token. See http://huggingface.co for more information.
• OpenAI: Set the OPENAI_API_KEY environment variable similar to above. See https://platform.openai.com/playground for more information.
• SGLang (Recommended): Please see https://github.com/sgl-project/sglang for details on how to start the server locally.

Instructions

Logging Directory Manifest

Regardless of where the script is run, the log files and results are placed in a logging directory in outputs/. The directory's name is of the form %Y-%m-%d/%H-%M-%S/, e.g. 2025-03-24/17-12-17. Inside each directory are the following files:

- .hydra/  # Hydra configuration files
- archive/  # QD Archive results
- archive_history.pkl  # Entire archive history
- reload.pkl  # Reload file to restart failed experiments
- qdllm.log  # Log file for QD-LLM experiments

Running an Experiment

To run an experiment, use:

python -m src.qdllm

Reloading

While the experiment is running, its state is saved to reload.pkl in the logging directory. If the experiment fails, e.g. due to memory limits, time limits, or network connection issues, reload.pkl may be used to continue the experiment. To do so, execute the same command as before, but append the path to the logging directory of the failed experiment.

python -m src.qdllm --experimental-rerun <path_to_dir>/.hydra/config.pickle

Reproducing Paper Results

The config/ directory contains the config files required to run the experiments shown in the paper. Below is a brief description of each config:

config/
    manager/
        agents/
            llm.yaml: LLM agent configuration
        communication/
            without_communication: *Without* communication condition
            with_communication: *With* communication condition
        env/
            module_cfg/
                grid/: Contains open, ring, hallway, and forced coordination hallway layouts:
                prompt/
                    steakhouse_prompt: LLM agent steakhouse prompt framework
            steakhouse.yaml: Steakhouse configuration file
        model/
            quantization_config: Contains 4bit and 8bit quantization config files for local huggingface models
            llama_8b_hf.yaml: Meta Llama 3 8B Instruct (see https://huggingface.co/models)
            llama_70b_hf.yaml: Meta Llama 3 70B Instruct (see https://huggingface.co/models)
            llama_70b_sglang.yaml: Meta Llama 3 70B Instruct using SGLang (see https://github.com/sgl-project/sglang)
        qd/
            measures/: All 9 workload measure files used for QD
            x0/x0: Initial prompt for QD experiments
            planqd.yaml: PLAN-QD configuration
            random.yaml: Random mutation configuration
        default.yaml: Configuration script

To reproduce the results from the paper, modify the configuration in default.yaml for manager/communication, manager/qd. To change the layout, change the value in manager.env.steakhouse.grid.

For example:

python -m src.qdllm \
  manager/env/module_cfg/grid=open \
  manager/qd=planqd \
  manager/communication=with_communication

will run PLAN-QD on the open layout with communication. Additionally:

python -m src.qdllm \
  manager/env/module_cfg/grid=hallway \
  manager/qd=random \
  manager/communication=without_communication

will run the random mutation baseline on the hallway layout without communication.

License

This code is released under the MIT License, with the following exceptions:

• overcooked_ai is adapted from the Overcooked AI Framework, which is released under the MIT License.

• src/envs/steakhouse is adapted from the Steakhouse Framework proposed in this paper, which is released under the MIT License.