Does Localization Inform Unlearning? A Rigorous Examination of Local Parameter Attribution for Knowledge Unlearning in Language Models

TL;DR

This paper critically examines localized unlearning in language models, revealing that effective unlearning does not necessarily depend on strictly localized parameter updates, challenging core assumptions of current methods.

Contribution

It provides a rigorous evaluation of localized unlearning approaches, questioning the link between parameter locality and unlearning effectiveness.

Findings

Local parameter updates are not always causally responsible for unlearning.

The set of parameters needing modification for unlearning is not strictly localized.

Localized unlearning assumptions may not hold universally.

Abstract

Large language models often retain unintended content, prompting growing interest in knowledge unlearning. Recent approaches emphasize localized unlearning, restricting parameter updates to specific regions in an effort to remove target knowledge while preserving unrelated general knowledge. However, their effectiveness remains uncertain due to the lack of robust and thorough evaluation of the trade-off between the competing goals of unlearning. In this paper, we begin by revisiting existing localized unlearning approaches. We then conduct controlled experiments to rigorously evaluate whether local parameter updates causally contribute to unlearning. Our findings reveal that the set of parameters that must be modified for effective unlearning is not strictly determined, challenging the core assumption of localized unlearning that parameter locality is inherently indicative of effective knowledge removal.