Reproduce “Automated Detection of Pre-training Text in Black-box LLMs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ruihan Hu, Yu-Ming Shang, Jiankun Peng, Wei Luo, Yazhe Wang, Xi Zhang

Researchers at Beijing University of Posts and Telecommunications developed VeilProbe, an automated framework that detects if a text was part of a black-box LLM's pre-training data. The framework achieves state-of-the-art accuracy on various benchmarks and LLMs, including proprietary models like ChatGPT and Claude, without requiring human intervention or internal model access.

Key insights

  • Sophisticated inference of latent membership features from LLM input-output behaviors is more effective than simple indicators for black-box pre-training data detection.
  • Proxy open-source LLMs can be utilized to infer crucial internal model characteristics, such as key token importance, from black-box target models.
  • Combining a prototype-based classifier with Information Bottleneck denoising allows for robust and accurate detection even when only a small number of ground-truth labeled samples are available.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.