Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning
This addresses the need for safer and more efficient knowledge removal in LLMs, though it is incremental by focusing on training stage differences.
The paper tackles the problem of machine unlearning in LLMs by showing that facts from pretraining vs. fine-tuning are not equally forgettable, with SFT-based unlearning achieving 10-50% higher retention and more stable results.
Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information.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.