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

TL;DR

This paper investigates how Large Language Models respond to unlearning different types of knowledge, introducing DUET benchmark and revealing distinct behaviors between pretrained and fine-tuned models.

Contribution

It introduces DUET, a comprehensive benchmark for evaluating unlearning, and highlights the differing responses of pretrained and fine-tuned models to unlearning procedures.

Findings

Fine-tuning on forget data improves stability and retention.

Pretrained models are more prone to instability and catastrophic forgetting.

SFT step yields 10-50% higher retention of knowledge.

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Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information.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 DUET (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.