Reproduce “Memorization in Language Models through the Lens of Intrinsic Dimension” 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
Stefan Arnold
This research explores the relationship between the intrinsic dimension (ID) of text sequences and their memorization likelihood in language models. Findings show that while memorization generally increases with data duplication and model size, higher intrinsic dimension correlates with less memorization in low-duplication regimes, with this effect diminishing or reversing for smaller models or under high data exposure.
Key insights
- Intrinsic dimension serves as a geometric proxy for sequence complexity, providing a deeper, latent-space understanding of memorization beyond surface-level factors.
- The influence of intrinsic dimension on memorization is dynamic, dependent on both the language model's scale and the data's exposure frequency.
- The study highlights a complex interplay between a sequence's structural complexity, the model's capacity, and its training exposure in shaping memorization behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.