PLD-4: A Multi-Task Framework for Detecting and Attributing LLM-Generated Paraphrases

ACL ARR 2025 May Submission2285 Authors

Abstract: LLMs make distinguishing human from ma- chine text challenging, particularly via para- phrasing used for evasion, impacting academic integrity, IP, and misinformation. We introduce the novel Paraphrase-based LLM Detection Framework (PLD-4), formalizing four tasks to evaluate detection in nuanced scenarios, including identifying layered AI text. Using MRPC and HLPC datasets, we employ a dual approach with feature-based and transformer models (XGBoost, DeBERTa-v3, RoBERTa). While achieving high accuracy on tasks like Sentence Pair Paraphrase Source Detection (XGBoost 96%) and Single Sentence Authorship Attribution (RoBERTa 93.9%), distinguishing original vs. paraphrased LLM output proved significantly challenging (RoBERTa 83.28%), high- lighting limitations in detecting layered AI generation. PLD-4 provides a critical foundation for developing more robust detection techniques.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Paraphrase Detection, LLM-Generated Text Detection, LLMs

Contribution Types: NLP engineering experiment, Reproduction study, Data resources, Data analysis

Languages Studied: English

Submission Number: 2285