Fast Unlearning at Scale via Margin Self-Correction

Arxiv CS.LG (machine learning)

Proposes a margin self-correction method for efficient LLM unlearning, avoiding costly retraining. Preserves model utility while removing selected training data. Demonstrated at scale with improved performance.

llm

Arxiv CS.LG (machine learning)·Federico Di GennaroAlexander ShevchenkoFanny Yang

8 days ago