Extractable Memorization From First Principles

arXiv cs.CL · 18 hours ago

Researchers propose a matched comparison methodology to rigorously measure extractable memorization in large language models, arguing that prior work either overstates extraction or fails to distinguish memorization from predictability. The approach requires comparing generation probabilities of training sequences against non-training sequences as a baseline, with concrete thresholds calibrated through conformal testing or census methods. This framework clarifies which prior extraction claims have validity issues and establishes more precise standards for when memorization evidence should be considered reliable.