GroupGPT: A Token-efficient and Privacy-preserving Agentic Framework for Multi-User Chat Assistant

News

• 🔥We have released a curated list of Awesome Multi-User Agents resources
• 🔥We have released the weights for GroupGPT’s components.
• 🔥We have released GroupGPT: A Token-efficient and Privacy-preserving Agentic Framework for Multi-User Chat Assistant. Check out the paper.

Overall Framework

GroupGPT adopts a small–large model collaborative architecture to decouple intervention timing from response generation, enabling efficient and accurate decision-making.

Contents

• Install
• Models
• MUIR Dataset
• Data Curation
• Training
• Inference

Install

git clone https://github.com/Eliot-Shen/GroupGPT.git
cd GroupGPT
conda create -n groupgpt python=3.10 -y
conda activate groupgpt
pip install -r requirements.txt

Models

You can found model weights via HuggingFace: 🤗 EliotShen/qwen-3-4b-intervention and 🤗 EliotShen/llama-3.2-3B-privacy.

MUIR Dataset

Due to privacy considerations, the MUIR dataset is not publicly downloadable. To request access, please fill out the application form.

After review, the dataset will be sent to your email address.

Data Curation

The data_curate.py script demonstrates our data curation pipeline for constructing training data.

• You need to replace the LLM interface in the script with your own API.
• Convert raw multi-user chat logs into JSON format as input.
• The pipeline will automatically process the data and generate the final training dataset.

Training

We provide training scripts for the key components of GroupGPT:

• Intervention Judge Model

  This model determines whether and when the system should intervene in multi-user conversations.

  sh train.sh

• Privacy Transcriber Model This module is responsible for privacy-aware rewriting of user inputs.

  sh privacy_train.sh

Make sure to properly configure the dataset paths and training hyperparameters before running the scripts.

Inference

Intervention Judge Model

The Intervention Judge Model determines whether the system should intervene in a multi-user group chat and identifies the most appropriate intervention reason.

Input Format

The model expects a JSON-formatted group chat context as input.

Example:

[
  {
    ...
  },
  {
    "user": "Alice",
    "message": "I studied for the exam all night and still feel like I failed..."
  },
  {
    "Intervention": "Emotional Support",
    "Reason": "Users are expressing frustration about the exam."
  },
  {
    "user": "Bob",
    "message": "Same here, that test was brutal."
  },
  {
    "user": "Charlie",
    "message": "Yeah the questions were way harder than the homework."
  }
]

This JSON context is inserted into the prompt and sent to the model.

Output Format

The model outputs a JSON decision indicating whether an intervention is needed.

Example:

{
  "choice": "Emotional Support",
  "reason": "Multiple users express frustration and discouragement about the exam."
}

If no intervention is necessary:

{
  "choice": "Stay Silent"
}

BibTeX

@article{shen2026groupgpt,
  title={GroupGPT: A Token-efficient and Privacy-preserving Agentic Framework for Multi-User Chat Assistant},
  author={Shen, Zhuokang and Wang, Yifan and Chen, Hanyu and Shen, Yunhang and Huang, Wenxuan and He, Gaoqi and Xie, Jiao and Ji, Rongrong and Lin, Shaohui},
  journal={arXiv preprint arXiv:2603.01059},
  year={2026}
}