Reproduce “GRU: Mitigating the Trade-off between Unlearning and Retention for LLMs” on OpenResearch

About this paper

Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, Bo Han

The authors introduce gradient rectification for unlearning, adjusting an update when it conflicts with a gradient from retained examples. In TOFU’s synthetic benchmark, the method improves forgetting quality (FQ) and model utility (MU) across tested setups. For Llama-2-7B with gradient ascent in TOFU’s 5% setup, FQ rises from −16.9281 to −3.5161 and MU from 0.0000 to 0.5190; these scores do not prove complete erasure.

Key insights

• The trade-off between unlearning and retention in LLMs is often caused by conflicting gradient directions, a conflict that can be empirically observed and mathematically mitigated.
• Gradient projection techniques, previously used in other learning paradigms, can be successfully adapted to constrain unlearning updates, effectively preserving general model utility.
• Effective unlearning is achievable even without explicit retain data by strategically using subsets of the unlearning dataset itself as internal retention references.

Summary by alphaXiv.