New Framework SCDG Improves Generative Plagiarism Detection
The new SCDG framework offers a training-free approach to detecting generative plagiarism by comparing the description length of a document generated by a frozen language model against and without a candidate source. This method provides token-level log-likelihood gains to assess predictive evidence and addresses the challenges posed by large language models (LLMs) to academic integrity and peer review. Evaluations were conducted using the PAN at CLEF benchmarks. The research, which highlights the limitations of previous detection methods that focused solely on AI involvement, can be accessed on arXiv under identifier 2608.03859.
Key facts
- SCDG is a training-free framework for generative plagiarism detection.
- It contrasts a frozen language model's description length of a document with and without a candidate source.
- The method yields token-level log-likelihood gains to measure predictive evidence.
- Evaluation is performed on PAN at CLEF benchmarks.
- The paper is available on arXiv with identifier 2608.03859.
- The work addresses challenges LLMs pose to academic integrity and peer review.
- Prior detection methods target AI involvement, not source reuse.
- Similarity-based methods struggle with extensive rewriting and multi-source synthesis.
Entities
Institutions
- arXiv
- PAN at CLEF