Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking
A new training-free framework, Source-Conditioned Description-Length Gain (SCDG), is introduced for generative plagiarism detection. It contrasts a frozen language model's description length of a suspicious document with and without a candidate source, yielding token-level log-likelihood gains. Evaluated on PAN at CLEF benchmarks, SCDG achieves 0.92 Precision, 0.97 Recall, and 0.94 F1 on a PAN 2025-derived pairwise benchmark, outperforming all baselines. On PAN 2026's multi-source retrieval task, it reaches 0.83 nDCG@10 and 0.96 Recall@100, surpassing all baselines.
Researchers propose SCDG, a training-free method for detecting generative plagiarism by measuring how much a candidate source reduces the description length of a suspicious document under a frozen language model. On PAN benchmarks, it achieves 0.94 F1 in pairwise detection and 0.83 nDCG@10 in multi-source retrieval, outperforming existing approaches.
SCDG leverages the description-length view of probabilistic prediction, using token-level log-likelihood gains to quantify incremental predictive evidence from a candidate source. This directional, training-free approach contrasts with similarity-based methods and LLM-generated-text detection, focusing on source reuse rather than AI involvement.
The method addresses a critical gap in academic integrity tools by detecting source reuse in LLM-generated text, even after extensive rewriting or multi-source synthesis. Its training-free nature and strong benchmark performance suggest potential for integration into plagiarism detection systems and peer review workflows.
SCDG could enhance commercial plagiarism detection software, offering a more robust solution for institutions and publishers facing challenges from LLM-generated content. Its training-free design reduces computational overhead, making it cost-effective for large-scale deployment.
Next signals include further validation on diverse datasets, exploration of real-world deployment in academic publishing, and potential adaptation for other domains like journalism or legal document analysis. The approach may also inspire new evaluation metrics for generative models.