Reproduce “Revisiting Non-Verbatim Memorization in Large Language Models: The Role of Entity Surface Forms” on OpenResearch

About this paper

Yuto Nishida, Naoki Shikoda, Yosuke Kishinami, Ryo Fujii, Makoto Morishita, Hidetaka Kamigaito, Taro Watanabe

Research from Nara Institute of Science and Technology introduces RedirectQA, a new dataset to systematically analyze how different surface forms of an entity impact Large Language Models' (LLMs) factual knowledge access. The study reveals that access to memorized facts is partially dependent on the specific surface form used, with consistency varying significantly across categories like spelling variants and abbreviations.

Key insights

• Access to memorized factual knowledge in LLMs is partially dependent on the specific linguistic surface form of an entity used in a query, leading to frequent prediction inconsistencies.
• The degree of factual access consistency varies systematically across different categories of surface-form variations; models are more robust to minor orthographic changes (Spelling Variants) but less so to larger lexical variations (Alternative Names and Abbreviations).
• Factual memorization exhibits 'cross-surface coupling,' where accuracy for a specific surface form is influenced by both its own frequency in pretraining data and the aggregate frequency of the underlying entity.

Summary by alphaXiv.