Back to Feed
Benchmarks & Evals / Efficiency & Inference

Detecting AI Plagiarism via Predictive Compression

Original: Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking
By Peijia Guo, Wenxuan Xie, ZiGuang Li, Ming Li Published 2026-08-04

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • The method uses cross-attention-like logic to compare unconditional codelength against source-conditioned codelength to detect dependencies.
  • It achieves high accuracy on benchmarks, with 0.94 F1 score on the PAN 2025 derived dataset.
  • The approach is computationally intensive compared to traditional text similarity metrics.
  • It provides evidence of content reliance rather than definitive proof of intent or provenance.

Summary & Methodology Analysis

The methodology relies on information theory to measure incremental predictive evidence. The system tokenizes both the suspicious document and the candidate source, then feeds these into a frozen autoregressive language model, which predicts the next token in a sequence based on prior tokens. By calculating the codelength of the suspicious document in two states, one without the source in the context window and one with it, the system derives the Source-Conditioned Description-Length Gain (SCDG). This scalar score represents the compression gain, effectively quantifying how much the source text resolves the uncertainty of the suspicious text. Token-level log-likelihood gains are tracked to allow for localized evidence tracing within the document.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary purpose of this research?

The research aims to detect generative plagiarism where LLMs are used to rewrite or synthesize content from existing sources, making it hard to identify via surface-level similarity metrics.

Q2. How does this method differ from standard plagiarism detection?

Unlike traditional metrics that look for surface-level word overlap, this method measures how well a source document explains the content of a suspicious document through the lens of a language model.

Q3. Is this tool intended to prove copyright infringement?

No, the paper notes that the method provides evidence of source dependence but does not serve as absolute proof of intent or provenance.

Q4. What models were used to implement the SCDG approach?

The implementation utilized Qwen3-8B-Base, Llama-3.1-8B, and Ministral-3-8B-Base-2512.

Q5. How does the system select candidates for comparison?

It constructs a DAAC indicator using retrieval ranks from BM25, sentence-level dense embeddings, and abstract-level sources to filter candidate sets before scoring them with SCDG.

Q6. What are the performance metrics on the PAN 2026 task?

The method reached 0.83 nDCG at 10 and 0.96 Recall at 100.

Q7. What is a major trade-off when using this method?

SCDG is computationally more expensive than traditional similarity metrics.

Q8. Has this been evaluated on languages other than English?

The paper does not specify evaluation on other languages, as testing was primarily restricted to English scientific documents and a specific news-domain set.

Q9. What was the result on the Multi-News test?

The calibrated SCDG classifier produced positive decisions for only 0.125 percent of the article pairs.