Verbatim Forgetting, Knowledge Forgetting, and Privacy Leakage as Predictors of Retained Utility Following Model Unlearning
Year: 2026
Type: article
Abstract: This analysis examines whether model unlearning reconciles information removal with useful performance, addressing international debates on accountable language model deployment. A secondary analysis uses published benchmark results from 2024, comprising eight unlearning methods evaluated in news and books, yielding 16 analytical observations. Descriptive statistics, Spearman correlations and exploratory individual method exclusions characterize performance relationships. Verbatim residual memorization correlates positively with retained utility in news and books, with coefficients of 0.904 and 0.944; corresponding knowledge residual correlations are 0.880 and 0.919. These directions persist across individual method exclusions. Absolute privacy leakage exhibits contrasting associations with utility, reaching 0.252 in news and 0.994 in books, while the news association changes sign under exclusions. Four observations combine zero residual memorization with zero retained utility, demonstrating that suppressed recall can accompany broader performance deterioration. These descriptive findings refine selective forgetting evaluation by distinguishing consistent memorization relationships from setting dependent privacy patterns. Deployment assessments should evaluate removal, disclosure and retained capability separately before accepting an unlearning intervention.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225462712