The Dual Nature of Unlearning: Impact of Fact Salience and Model Fine-Tuning

Anna Borisiuk, Andrey Savchenko, Alexander Panchenko, Elena Tutubalina

Track: tiny / short paper (up to 4 pages)

Keywords: Machine unlearning, fact popularity, salience, supervised fine-tuning, benchmark

Abstract:

Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments
show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Anna_Borisiuk1

Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 45