[STAI-X 2026] Segmenting Human-LLM Co-authored Text via Change Point Detection ✨

This repository contains the official implementation of Segmenting Human-LLM Co-authored Text via Change Point Detection (accepted at STAI-X 2026).

Our method provides a change point detection viewpoint on segmenting the human-LLM co-authored text (Figure 1 🧠). We leverage the hetergeneity among segments to design our method, achieves robust and state-of-the-art performance across a wide range of settings.

_{Figure 1. 🧠 AUCs of various detectors on the WikiQA and Story datasets with varying lengths of input texts. RoBERTa and AdaDetectGPT are two training-based detectors, the others are zero-shot detectors.}

🧭 Table of Contents

• [STAI-X 2026] Segmenting Human-LLM Co-authored Text via Change Point Detection ✨
  • 🧭 Table of Contents
  • 🛠️ Installation
    • Requirements
    • Setup
  • 🚀 Usage
  • 🎁 Additional Resources
    • Implemented baselines
    • Reproducibility guide
  • 📖 Citation

🛠️ Installation

Requirements

• Python 3.10.8
• PyTorch 2.11.0
• CUDA-compatible GPU (experiments conducted on A100 with 40GB memory)

Setup

./setup.sh

We also attach a environment.yml file to show our detailed configuration on the conda environment.

🚀 Usage

Recommended for off-the-shelf usage:

python scripts/detector_value_cp.py \
  --from_pretrained 'scripts/FineTune/ckpt/' \
  --phi NFT \
  --cp_method 'DP' \
  --power 2.0 \
  --cp_selection 'aic' \
  --weight_type 'invar' \
  --eval_dataset ${eval_data_path} \
  --output_file ${eval_result_path}

Make sure that your data is a .json file named xxx.raw_data.json with the following structure:

{
  "sampled_sentence": [
    ["Sentence 1 of Document 1", "Sentence 2 of Document 1", ...], 
    ["Sentence 1 of Document 2", "Sentence 2 of Document 2", ...], 
    ...
  ],
  "source_label": [
    ["H", "H", ....], 
    ["L", "L", ...], 
    ...
  ], 
}

• sampled_sentence include a list of document whose each element is a list containing a series of sentences
• source_label include a list of document whose each element is a list containing a series of label of sentences, either H (human) or L (LLM)
• The two lists should typically be aligned in length (one-vs-one)

🎁 Additional Resources

The scripts/ directory contains implementations of various LLM segmentation location methods from the literature. These implementations are designted to provide:

• consistent input/output formats
• simplified method comparison

Implemented baselines

Method	Script File	Paper/Website
SenPred	detect_naive_sp.py	arXiv:2510.01268
Voting	detect_voting_sp.py	EMNLP-main.463
PaLD	detect_pald.py	arXiv:2401.12070
TextTilling	detect_texttilling.py	arXiv:1908.09203

Reproducibility guide

• exp_longdocs.sh: generate results of Figure 1, then run python exp_longdocs/plot_length.py to get Figure 1
• exp_single_cp.sh: generate results of Table 1, then run python exp_single_cp/1table.py to get Figure 1
• exp_vary_paralen.sh: generate results of Table 2, then run python exp_vary_paralen/show_results.py to get Table 2

📖 Citation

If you find this work useful, please consider citing our paper:

@inproceedings{li2026segmenting,
  title={Segmenting Human-LLM Co-authored Text via Change Point Detection},
  author={Mengchu Li and Jin Zhu and Jinglai Li and Chengchun Shi},
  booktitle={Statistics and Trustworthy AI for Cross (X)-Domain Acceleration},
  abbr={STAI-X},
  year={2026},
}

If you have any questions, please feel free to open an issue.