Research·arXiv cs.CL·14h ago

New method detects plagiarism in heavily rewritten LLM outputs

Researchers propose Source-Conditioned Description-Length Gain, a training-free method to detect when LLM outputs plagiarize specific sources even after heavy rewriting or multi-source blending. Unlike existing similarity-based approaches that fail on paraphrased content, SCDG measures how much a frozen language model's encoding cost drops when given a candidate source, revealing hidden dependencies. The technique addresses a critical gap in academic integrity as generative AI proliferates: distinguishing permissible AI use from source theft. This directional framework could reshape plagiarism detection infrastructure for publishers, universities, and peer review systems relying on current similarity metrics.

Modelwire context

Explainer

The paper's core insight is that similarity-based plagiarism detection fails on paraphrased or multi-source blended content because it measures surface overlap, not dependency. SCDG sidesteps this by measuring how much a language model's compression improves when it sees a candidate source, which reveals whether the model actually relied on that source during generation, regardless of textual resemblance.

This connects directly to the retrieval and reranking work from the past few days. SciRet (August 4) showed that domain-specific reranking matters because off-the-shelf models fail on specialized corpora, and the Georgia Tech eRisk submission (August 4) demonstrated that staged retrieval pipelines combining multiple ranking methods outperform single approaches. SCDG is solving the inverse problem: not ranking sources to retrieve, but ranking candidate sources after the fact to identify which ones actually influenced the output. The underlying principle is the same: one-size-fits-all metrics (similarity, general-purpose rerankers, single retrieval modalities) break down when you need to detect hidden dependencies rather than surface matches.

If SCDG maintains detection accuracy above 80% on a held-out benchmark of adversarially paraphrased academic papers (using techniques like back-translation or synonym substitution) that fool existing plagiarism tools, the method is production-ready for university systems. If performance drops below 60% on such adversarial sets, the encoding-cost signal is too brittle for institutional deployment.

Coverage we drew on

SciRet: A Compute-Aware Empirical Study of Retrieval and Reranking for Scientific RAG · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsLarge language models · Source-Conditioned Description-Length Gain

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.