GRACE:Gradient-guided Coreset Selection for LLM Unlearning
Artificial IntelligenceMachine Learning
Summary
The authors address a problem in machine unlearning for large language models, where it's hard to know exactly which data to forget or keep based on only a few examples. They propose a method called GRACE that cleverly picks small sets of data to remove or retain by looking at how the model's gradients (which guide learning) change. GRACE uses special math techniques to select these data points in a way that keeps the model useful while making it forget the undesired behavior. They tested GRACE in several situations and found it maintains good forgetting while improving how well the model works compared to other methods.
Authors
Praveen Bushipaka, Andrea D'Angelo, Lucia Passaro, Tommaso Cucinotta
Abstract
Machine Unlearning methods for Large Language Models typically assume pre-specified forget and retain sets. In realistic settings, however, requests may provide only a few examples of undesired behavior, requiring forget and retain sets to be inferred from heterogeneous corpora. We study this data-selection problem and propose GRACE , a gradient-guided coreset selection method that constructs both forget and retain sets for LLM unlearning. GRACE first computes a forget direction from seed examples that elicit the undesired behavior, then selects a compact forget coreset whose gradients approximate this direction using non-negative orthogonal matching pursuit. To preserve model utility, it selects retain examples after projecting out the forget direction and applying clustered orthogonal matching pursuit in the remaining gradient space. Across two target domains, two model families, and four unlearning algorithms, GRACE improves model utility while maintaining comparable forget quality, with particularly consistent gains over prior gradient-based selection methods.