Reproduce “Exploring Approaches for Detecting Memorization of Recommender System Data in Large Language Models” 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
Antonio Colacicco, Vito Guida, Dario Di Palma, Fedelucio Narducci, Tommaso Di Noia
Researchers from Politecnico di Bari evaluated jailbreaking, unsupervised, and automated methods for detecting MovieLens-1M memorization in LLaMA models. Automatic Prompt Engineering proved most effective for extracting textual item data, achieving up to 26% exact-match accuracy, while all methods struggled to retrieve numerical user and rating information.
Key insights
- Automatic Prompt Engineering (APE) emerges as the most effective and scalable strategy for detecting memorized textual dataset information in LLMs.
- LLMs exhibit a stark difference in extractability: semantic textual data (e.g., movie titles) is considerably easier to detect/extract than non-semantic numerical data (e.g., user IDs, ratings).
- Unsupervised latent knowledge probes can distinguish real textual data from fabricated entries, but their signals do not conclusively confirm explicit memorization and are ineffective for numerical data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.