Latest News

• [ 2025.05.08 ] 🔄 We have updated the dataset to unify the label "MGT" to "LGT" to match the documentation. No other changes were made. The original MGT in the training set is equivalent to LGT (label=2). We apologize for any inconvenience and confusion caused.
• [ 2025.04.15 ] 💡 We have released the detailed task guidelines and training data ~
• [ 2025.03.20 ] 🔥 Registration for NLPCC-2026 Task 6 is now open. Welcome to join us!

Introduction

The rapid development of large language models (LLMs) has given rise to a series of challenges, including the generation of disinformation, the spread of harmful content, and various forms of misuse. Against this backdrop, the efficient discrimination between LLM-generated text and human-written text has become an urgent and critical research issue in the field of natural language processing (NLP). While remarkable progress has been made, relevant research has largely focused on English, systematic and technical exploration for the Chinese remain scarce. This shared task aims to fill this gap, build more robust Chinese LLM-generated text detectors, and advance research and real-world applications in this field within the Chinese.

Following the success of the 1st Shared Task on LLM-Generated Text Detection (NLPCC 2025), the 2nd Shared Task on LLM-Generated Text Detection in 2026 features significant upgrades: the task formulation has been expanded from binary to ternary classification. Specifically, in addition to distinguishing between human-written text and LLM-generated text, a new category for identifying LLM-refined text has been introduced, which better aligns with real-world application scenarios of LLMs. Participating teams are required to design and implement text detection algorithms based on the training data provided to achieve accurate classification and will undergo rigorous stress testing.

Dataset

The training set for this task is primarily sampled and adapted from the CUDRT (TIST 2026) dataset (Chinese subset: Complete 25 ratio). The test set will be released in the later stages of the competition for teams to verify the effectiveness of their methods and evaluate final performance. The evaluation dataset is extended and constructed based on the DetectRL benchmark (NeurIPS 2024) framework, containing multiple generation models and domain data to ensure the authenticity and challenge of the evaluation scenarios.

Training Set

Contains data from 4 types of LLMs and 2 domains. Specifically, data sources include news and academic writing, and generation models include GPT-4, Qwen, ChatGLM, and Baichuan. The training set contains a total of 20,370 pairs of samples. Each pair contains four fields: "ID", "HWT", "LGT", and "HLT". "ID" identifies the sample number, while "HWT", "LGT", and "HLT" represent "Human-written text", "LLM-generated text", and "LLM-refined text" respectively.

Data Download

The training data can be found at the following links:

• Github: Data Folder

Data Restrictions

To support the development of detection systems, participants are allowed to perform data processing and augmentation based on the provided training data, but no external additional new datasets may be introduced.

Evaluation Metrics

The official evaluation metric for this task is the macro-averaged F1-Score.

Submission and Evaluation

The submission platform for this evaluation task (tentatively Alibaba Cloud Tianchi) will be launched one week before the release of the test data for Phase 1 evaluation (open evaluation, leaderboard verification); and subsequently Phase 2 evaluation (closed evaluation, final ranking). Specific details of the test result text are as follows:

1. Test Result File

• Your test result file must be a JSON file containing all samples.
• Please ensure that the text and ID fields remain unchanged.
• Each sample in the JSON file should contain the following fields: "ID", "text", and "label" (HWT: 0, LGT: 1, HLT: 2).

2. Code and Data

• The code folder should contain all the code required for data augmentation, data processing, model training, and model inference, as well as the complete dataset used to train the detector.
• Please include a simple README.md and an environment configuration file (e.g., requirements.txt).

3. Technical Report

The technical report should describe in detail the methods used to solve the task. All solutions will be handled confidentially and deleted uniformly after the NLPCC 2026 conference.

Tentative Schedule

Organizers

This shared task is jointly organized by the Natural Language Processing & Portuguese-Chinese Machine Translation Laboratory (NLP2CT) at the University of Macau, Central China Normal University, and Alibaba Cloud.

• Junchao Wu | University of Macau （Contact）nlp2ct.junchao@gmail.com | Homepage
• Derek Fai Wong | University of Macau | Homepage
• Runzhe Zhan | University of Macau | Homepage
• Zeyu Wu | University of Macau
• Zhiwen Xie | University of Macau / Central China Normal University
• Yichao Du | Alibaba Cloud
• Longyue Wang | Alibaba Cloud | Homepage

FAQ

Q: Where can I register for this shared task? Can team registration be updated?

A: Fill out the following link to register. Before the registration deadline, you can fill out this link repeatedly to update team registration information.

Q: Is there any restriction on the number of model parameters or data augmentation methods? If an ensemble approach is used, are there any constraints on the total number of parameters?

A: NLPCC 2026 Share Task 6 places no restrictions on the number of parameters for a single model, nor on the total number of parameters after ensemble. There are also no restrictions on data augmentation methods or models used. You may use open-source LLMs or closed-source APIs for text paraphrasing, rewriting, and other data-augmentation operations; these are fully allowed under the competition rules.

Q: The training set is sourced from news and academic writing, using four generation models. Will the subsequent test data also consist of these two sources and these four models, or will it be expanded? What is the generation method for test data? Will it contain formatting artifacts (e.g., "\n") or incomplete texts?

A: As stated in the README under the Dataset section, the test set will be constructed using out-of-distribution data from another dataset (DetectRL benchmark, NeurIPS 2024). It will no longer be limited to the news and academic domains or the four generation models covered in the training set. The test data may introduce texts from unknown domains and unknown generation models, including different data generation schemes, to conduct multi-dimensional stress testing and comprehensively evaluate the detector's actual performance in real-world application scenarios. Test-set LGT text may not use the same front-25%-token continuation as the training set; even if continuation is used, the prefix tokens will be cleaned, resulting in higher overall quality. The test set undergoes strict preprocessing and will not retain redundant formatting symbols such as "\n". It will also avoid abrupt text truncation, though it will not force every text to end with a period, preserving natural forms such as byline signatures to test real-world robustness.

Q: Training data quality issues: some samples contain pure English text labeled as HLT; some end with "123" with preceding text identical to the human text; and the README mentions LGT while the dataset uses MGT. How should these be understood?

A: The pure English samples labeled as HLT and the ~2,000 samples ending with "123" (with preceding text identical to the human text) are both native noise from the CUDRT dataset; no additional cleaning was performed. Teams are expected to design their own data-cleaning strategies. Regarding the label inconsistency: in the current training set, MGT is equivalent to LGT (label=2). We will update the dataset in a subsequent release to uniformly correct MGT to LGT, along with an official correction notice.

Q: Can we use pseudo-labeling on the test set data (i.e., running the test data through our model to obtain pseudo-labels and using them as expanded training data)?

A: No, it is NOT allowed to use test set data for pseudo-labeling. You cannot add any test set samples with predicted pseudo-labels into your training data for model optimization. This practice is not applicable in real-world application scenarios, so we do not permit such operations in the competition. The official test set is strictly for final-submission evaluation only. Any form of using test set data for training, fine-tuning, pseudo-label construction, or data augmentation is considered a violation of the competition rules. You may only use the officially provided training set. Methods to expand the training set are limited to the announced legal data augmentation methods, and no new data sources may be explicitly introduced.

References

If your research uses relevant datasets, or if this task is helpful to you, please consider citing the following literature:

@article{wu2024detectrl,
  title={Detectrl: Benchmarking llm-generated text detection in real-world scenarios},
  author={Wu, Junchao and Zhan, Runzhe and Wong, Derek F and Yang, Shu and Yang, Xinyi and Yuan, Yulin and Chao, Lidia S},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={100369--100401},
  year={2024}
}

@inproceedings{wu2025overview,
  title={Overview of the NLPCC 2025 Shared Task 1: LLM-Generated Text Detection},
  author={Wu, Junchao and Zhan, Runzhe and Wang, Qianli and Yuan, Yulin and Chao, Lidia S and Wong, Derek F},
  booktitle={CCF International Conference on Natural Language Processing and Chinese Computing},
  pages={263--274},
  year={2025},
  organization={Springer}
}

@article{tao2026toward,
  title={Toward Reliable Detection of LLM-Generated Texts: A Comprehensive Evaluation Framework with CUDRT},
  author={Tao, Zhen and Chen, Yanfang and Xi, Dinghao and Li, Zhiyu and Xu, Wei},
  journal={ACM Transactions on Intelligent Systems and Technology},
  volume={17},
  number={2},
  pages={1--35},
  year={2026}
}