kandpal-2023-non-verbatim-memorization-distinction
IN premise — summaries/2026/08/24/kandpal-2023-long-tail-knowledge-s5-related-work.md
Created 2026-08-25T02:58:07+00:00
Kandpal et al. (2023) distinguish their findings from verbatim memorization (privacy literature, e.g., Carlini et al. 2019/2021), focusing instead on non-verbatim factual knowledge stored in paraphrased form and generated at test time.
Summary
This research targets a different problem than the well-known privacy memorization literature: instead of asking whether a model can spit back exact training sentences, it examines whether models hold facts in a transformed, paraphrased form they can regenerate in new words. The distinction matters because that kind of semantic knowledge retention is qualitatively different from string-level memory, and it implies that safeguards designed to block verbatim recall may say nothing about how much factual understanding a model actually carries.