To Forget or Not? Towards Practical Knowledge Unlearning for Large Language Models
This work addresses the challenge of practical knowledge unlearning for LLMs to protect privacy and copyright, but it is incremental as it builds on existing unlearning paradigms with a new method for better precision.
The paper tackles the problem of large language models (LLMs) inadvertently erasing essential knowledge during unlearning of sensitive data, and finds that existing methods often over-unlearn; it proposes MemFlex, a method that uses gradient information to precisely target sensitive parameters, showing superior performance in both precise unlearning and knowledge retention.
Large Language Models (LLMs) trained on extensive corpora inevitably retain sensitive data, such as personal privacy information and copyrighted material. Recent advancements in knowledge unlearning involve updating LLM parameters to erase specific knowledge. However, current unlearning paradigms are mired in vague forgetting boundaries, often erasing knowledge indiscriminately. In this work, we introduce KnowUnDo, a benchmark containing copyrighted content and user privacy domains to evaluate if the unlearning process inadvertently erases essential knowledge. Our findings indicate that existing unlearning methods often suffer from excessive unlearning. To address this, we propose a simple yet effective method, MemFlex, which utilizes gradient information to precisely target and unlearn sensitive parameters. Experimental results show that MemFlex is superior to existing methods in both precise knowledge unlearning and general knowledge retaining of LLMs. Code and dataset are released at https://github.com/zjunlp/KnowUnDo.