meituan-longcat/DPT-Agent

meituan-longcat/DPT-Agent

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

Language: Python

License: MIT

Stars: 5

Forks: 0

Open issues: 0

Created: 2025-11-07T04:27:07Z

Pushed: 2025-11-07T04:27:22Z

Default branch: main

Fork: no

Archived: no

README:

DPT-Agent

This is the official implementation of paper "Leveraging Dual Process Theory in Language Agent Framework for Simultaneous Human-AI Collaboration" accepted by ACL 2025 Main.


📄 Paper | 🌐 Website | 📘 机器之心 | 🧪 AGI-Eval

🔥News

• [2025/05/16] The DPT-agent paper has been accepted by ACL 2025!
• [2025/03/18] Our work is featured by 机器之心 on Wechat!
• [2025/03/06] We have established a partnership with AGI-Eval platform. The benchmark results of Overcooked Challenge are now available on AGI-Eval-Overcooked Challenge.

Overcooked Challenge


DPT-Agent Framework


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

@inproceedings{zhang-etal-2025-leveraging,
title = "Leveraging Dual Process Theory in Language Agent Framework for Real-time Simultaneous Human-{AI} Collaboration",
author = "Zhang, Shao and Wang, Xihuai and Zhang, Wenhao and Li, Chaoran and Song, Junru and Li, Tingyu and Qiu, Lin and Cao, Xuezhi and Cai, Xunliang and Yao, Wen and Zhang, Weinan and Wang, Xinbing and Wen, Ying",
editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-long.206/",
doi = "10.18653/v1/2025.acl-long.206",
pages = "4081--4108",
ISBN = "979-8-89176-251-0",
}