Technical Summary: TSM-BENCH: Detecting LLM-Generated Text in Real-World Wikipedia Editing Practices

Problem Statement

The integrity of user-generated content (UGC) platforms like Wikipedia is threatened by the proliferation of machine-generated text (MGT). While existing benchmarks for MGT detection primarily rely on generic, unconstrained prompts (e.g., "Write an article about machine learning"), real-world editors employ Large Language Models (LLMs) for specific, contextually constrained tasks such as summarization, paragraph continuation, and style transfer. The authors argue that these task-specific instances of MGT differ significantly from generic MGT; due to their constrained formulation and contextual conditioning, they tend to resemble human-written text (HWT) more closely in style and semantics. Consequently, current State-of-the-Art (SOTA) detectors, which are often trained or evaluated on generic data, may fail to identify MGT in realistic editing scenarios, potentially leading to an overestimation of detector reliability and risking the contamination of training data for future AI models.

Methodology

The paper introduces TSM-BENCH, a comprehensive benchmark designed to evaluate MGT detectors on task-specific, real-world editing practices. The methodology involves four primary stages:

1. Task Definition: Based on empirical research into how Wikipedia editors use LLMs, the authors define four specific editing tasks:

   • Introductory Paragraph: Writing the opening paragraph of a new section.
   • Paragraph Continuation: Extending an existing, incomplete human-written paragraph.
   • Summarisation: Generating a lead section based on the article body.
   • Text Style Transfer (TST): Neutralizing biased sentences to adhere to Wikipedia's Neutral Point of View (NPOV) policy.

2. Benchmark Construction:

   • Data: The corpus comprises 152,910 parallel texts (human-written and machine-generated) across three languages (English, Portuguese, Vietnamese) representing high, medium, and low-resource levels.
   • Generators: MGT is generated using six different models, ranging from large LLMs (GPT-4o, GPT-4o-mini, Gemini 2.0, DeepSeek) to smaller language models (SLMs: Qwen 2.5-7B, Mistral-7B).
   • Prompt Engineering: To ensure high-quality, realistic MGT, the authors adapt prompts from natural language generation literature. They automatically evaluate these prompts against a minimal baseline using metrics like BLEU, ROUGE, BERTScore, and QAFactEval. The highest-scoring strategies (e.g., Retrieval-Augmented Generation for paragraph writing, Few-shot for TST) are selected for data generation.

3. Experimental Design: Five experiments are conducted to benchmark 12 SOTA detectors (including off-the-shelf, zero-shot, and supervised models):

   • Experiment 1: Evaluating off-the-shelf detectors on generic vs. task-specific data.
   • Experiment 2: Assessing zero-shot and supervised detectors across all task-language-generator configurations.
   • Experiment 3: Testing out-of-domain generalisation (training on task-specific vs. generic data and testing across Wikipedia, news, and social review domains).
   • Experiment 4: Feature analysis using SHAP values to determine what linguistic cues models rely on when trained on generic vs. task-specific data.
   • Experiment 5: Analyzing cross-task generalisation (training on one task and testing on others).

Key Contributions

• TSM-BENCH: The introduction of a multilingual, multi-generator, and multi-task benchmark specifically for task-specific MGT detection, containing over 150,000 samples.
• Empirical Evidence of Performance Drop: Demonstration that detection accuracy significantly degrades when moving from generic to task-specific scenarios.
• Generalisation Asymmetry: Identification of a critical asymmetry where models fine-tuned on task-specific data generalize well to generic data, but models trained on generic data fail to generalize to task-specific data.
• Feature Analysis: Evidence that models trained on generic data overfit to superficial artifacts (e.g., section formatting markers like "=="), whereas models trained on task-specific data rely more on semantically meaningful tokens.

Results

The experiments yield several critical findings:

• Significant Accuracy Decline: Compared to evaluations on generic data, average detection accuracy drops by 10–40% on task-specific data. Specifically, accuracy drops by up to 32% for zero-shot detectors, 20% for off-the-shelf detectors, and 10% for supervised detectors.
• Detector Failure in Real-World Scenarios: Off-the-shelf detectors, which achieve >93% accuracy on generic data, drop to between 47% and 73% on task-specific tasks. Supervised models perform better (79.7%–91.8% average accuracy) but still struggle significantly with tasks like Paragraph Continuation and sentence-level TST.
• Generalisation Asymmetry: Models trained on task-specific data achieve high accuracy (often >90%) when tested on generic data across domains. Conversely, models trained on generic data fail to detect task-specific MGT, often performing near random chance (<60%) even within the same domain.
• Overfitting to Artifacts: SHAP analysis reveals that models trained on generic data heavily rely on superficial formatting cues (e.g., "==", "#") which are absent or different in task-specific editing. Models trained on task-specific data focus on semantic tokens, indicating a more robust learning of MGT patterns.
• Cross-Task Limitations: Detection performance does not transfer well across different writing tasks, suggesting that future detectors must be trained on a diverse combination of tasks rather than a single generation type.

Significance

The paper concludes that prior benchmarks relying on generic generation settings likely overestimate the performance of MGT detectors. The findings suggest that most current detectors are unreliable for automated detection in real-world contexts such as UGC platforms, where users employ LLMs for specific, constrained tasks.

The significance of TSM-BENCH lies in its ability to provide a more realistic foundation for developing and evaluating future models. By highlighting the limitations of current detectors and the specific nature of task-specific MGT, the authors argue that the field must move beyond generic benchmarks. They recommend that future models be trained on diverse combinations of task-specific data to ensure robustness in real-world applications, thereby safeguarding the knowledge integrity of platforms like Wikipedia and preventing the degradation of AI training data.