ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models

Researchers introduced ZeroUnlearn, a few-shot unlearning framework for large language models.

It maps sensitive inputs to a neutral state and removes their original representations with a closed-form multiplicative update.

A gradient-based extension handles multiple samples, aiming to preserve general utility while deleting private or harmful knowledge.

The approach could be a more efficient alternative to retraining or heavy fine-tuning for model editing and privacy protection.