DPT-Agent

This is the official implementation of paper "Leveraging Dual Process Theory in Language Agent Framework for Simultaneous Human-AI Collaboration."

Results

Usage

Installation

Create a new environment

conda create -n dptagent python=3.10 -y   # support python<=3.10
conda activate dptagent


# Install pytorch
# We do not need gpu
pip install torch torchvision torchaudio

# Installing dependencies

bash ./install.sh

Running Instructions

Set up LLMs

litellm -c llms/litellm/config.yml --port 40000

# check health
curl --location 'http://127.0.0.1:40000/health' -H "Authorization: Bearer sk-1234"

Run LLM as Indenpendent System 1 and System 2 Experiments

For single agent exp, first change the config/envs/overcooked.yaml

mode: burger_exp1
...
num_agents: 1

Then run:

sh scripts/exp0/openai.sh

Run Single Agent Experiments

For single agent exp, first change the config/envs/overcooked.yaml

mode: burger_exp1
...
num_agents: 1

Then run:

sh scripts/exp1/openai.sh

Run Collaboration Experiments with Rule-baed Agents

For collaboration exp, first change the config/envs/overcooked.yaml

mode: burger_exp1
...
num_agents: 2

Then run:

sh scripts/exp2/openai.sh

Run Human Experiment

For use map 2, change the config/envs/overcooked.yaml

mode: burger_aa_new
...
num_agents: 2

Then run:

sh scripts/overcooked/human_llm_app.sh

Then open the website http://localhost:5001

Help

For more information, please run

python llm_agent_run_act.py --help

Developing Guide

We recommend using pre-commit to unify the format before commit.

# Init
pre-commit install

# You can run manually (maybe multiple times)
pre-commit run --all-files

# or it will automatically run when you try to commit, which is slow and seems stuck.

Cite

@misc{zhang2025ldpt,
      title={Leveraging Dual Process Theory in Language Agent Framework for Real-time Simultaneous Human-AI Collaboration}, 
      author={Shao Zhang and Xihuai Wang and Wenhao Zhang and Chaoran Li and Junru Song and Tingyu Li and Lin Qiu and Xuezhi Cao and Xunliang Cai and Wen Yao and Weinan Zhang and Xinbing Wang and Ying Wen},
      year={2025},
      eprint={2502.11882},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2502.11882}, 
}