TSM-Bench: Detecting LLM-Generated Text in Real-World Wikipedia Editing Practices

This repository contains the official implementation of our ICLR 2026 paper TSM-Bench: Detecting LLM-Generated Text in Real-World Wikipedia Editing Practices.

Please find our data on HuggingFace. This includes the data introduced in our WikiNLP @ ACL 2025 paper WETBench: A Benchmark for Detecting Task-Specific Machine-Generated Text on Wikipedia, as well as the machine-generated text introduced with TSM-Bench.

---

TSM-Bench

Automatically detecting machine-generated text (MGT) is critical to maintain- ing the knowledge integrity of user-generated content (UGC) platforms such as Wikipedia. Existing detection benchmarks primarily focus on generic text gen- eration tasks (e.g., “Write an article about machine learning.”). However, editors frequently employ LLMs for specific writing tasks (e.g., summarisation). These task-specific MGT instances tend to resemble human-written text more closely due to their constrained task formulation and contextual conditioning. In this work, we show that a range of MGT detectors struggle to identify task-specific MGT reflecting real-world editing on Wikipedia. We introduce TSM-BENCH, a multilingual, multi-generator, and multi-task benchmark for evaluating MGT de- tectors on common, real-world Wikipedia editing tasks. Our findings demonstrate that (i) average detection accuracy drops by 10–40% compared to prior bench- marks, and (ii) a generalisation asymmetry exists: fine-tuning on task-specific data enables generalisation to generic data—even across domains—but not vice versa. We demonstrate that models fine-tuned exclusively on generic MGT overfit to superficial artefacts of machine generation. Our results suggest that, in contrast to prior benchmarks, most detectors remain unreliable for automated detection in real-world contexts such as UGC platforms. TSM-BENCH therefore provides a crucial foundation for developing and evaluating future models.

---

Requirements

1. Clone the Repository

git clone tbd
cd TSM-Bench

2. Download Pretrained Models

We host the pre-trained models for Experiment 4 (generalisation) on Google Drive. You can also run the models yourself using the script generalise/code/train_hp_g.sh.
The following script will download and unzip the models into generalise/code/hp_len.
The download is ~18GB, unzipped size is ~25GB.

bash download_models.sh

Note: This script requires gdown. Install it via:

pip install gdown

---

3. Set Up the Python Environment

We recommend using a virtual environment to manage dependencies.

python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
export HF_HOME="" # optimal to manage hf cache

---

Main Results

Note:

• Experiments 1-2 were run on either a single NVIDIA A100 80GB or two NVIDIA A100 40GB GPUs. ÷
• Experiments 3-5 were run on a single NVIDIA A100 40GB.
• We strongly recommend using GPUs to replicate results.

Experiment 1: Off-the shelf detectos

Run the following script to reproduce the results:

bash run_ots.sh

Experiment 2: Supervised and zero-shot detectors

To run black-box detectors, provide your OpenAI API key. If you skip this, only local models will be evaluated.

Note:
Zero-shot evaluations may take up to 1.5 days. We recommend splitting scripts across HPC jobs.
Supervised detectors run much faster. To run only those:

bash detect_train_hp.sh

To run all:

export OPENAI_API_KEY=sk-...
bash run_detection.sh

---

Experiment 3: Out-of-domain generalisation

This will populate generalise/data/detect with files named: trainFile_2_testFile_model_language.jsonl

bash run_generalisation.sh

---

Experiment 4: Feature analysis

To generate the SHAP plot run:

bash run_shap_vals.sh

Experiment 5: Cross-task generalisation

This will populate generalise/data/detect with files named: trainFile_2_testFile_model_language.jsonl

bash tbd shortly!

---

Other Results

Linguistic Analysis

Run the files in linguistic_analysis/la.sh with a GPU.

Prompt Selection

You can run this without QAFactEval if it causes issues.

To replicate our prompt selection evaluation:

1. Create a Conda Environment for QAFactEval

We recommend using Conda: Clone and install QAFactEval into the current directory:

conda env create -f environment_qafe.yml
pip install -r requirements_qafe.txt

Clone and install QAFactEval into the current directory. Follow setup instructions at: https://github.com/salesforce/QAFactEval. Don't forget to add the model_folder in scorers/qafe.py.

2. Download Style Classifiers

Same procedure as above. Ensure gdown is installed.

bash download_sc.sh

3. Run the Evaluation

This example runs the evaluation for Vietnamese. Adjust the language as needed.

bash run_prompt_eval.sh

---

Contributing

Valuable contributions include:

• Implementing robust data cleaning with mwparserfromhtml
• Extending to more languages, adding generators, and expanding task coverage

---

Citation

If you use this work, please cite:

TSM-Bench

@inproceedings{
quaremba2026tsmbench,
title={{TSM}-Bench: Detecting {LLM}-Generated Text in Real-World Wikipedia Editing Practices},
author={Gerrit Quaremba and Denny Vrandecic and Elizabeth Black and Elena Simperl},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=zimuL7ZmIi}
}

WETBench

@inproceedings{quaremba-etal-2025-wetbench,
    title = "{WETB}ench: A Benchmark for Detecting Task-Specific Machine-Generated Text on {W}ikipedia",
    author = "Quaremba, Gerrit  and
      Black, Elizabeth  and
      Vrandecic, Denny  and
      Simperl, Elena",
    booktitle = "Proceedings of the 2nd Workshop on Advancing Natural Language Processing for Wikipedia (WikiNLP 2025)",
    month = aug,
    year = "2025",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.wikinlp-1.6/"
}