Scholarlink

This chapter considers the use of attributes of unlearning data to enhance unlearning performance. It studies a framework called “SPlit, UNlearn, MerGE” (Spunge), which splits unlearning data into subsets based on the values of a selected attribute, unlearns each subset separately, and merges the unlearned models. Spunge can be used with any unlearning method to amplify its effectiveness. We empirically demonstrate such improvements for two recent unlearning methods, reducing undesirable behaviors and hazardous knowledge in two popular LLMs.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Split, Unlearn, Merge: Leveraging Data Attributes for More Effective Unlearning in LLMs

  • 章节

  • 发表日期：

    2026年

查看全文
  • Swanand Ravindra Kadhe,
  • Farhan Ahmed,
  • Dennis Wei,
  • Nathalie Baracaldo,
  • Inkit Padhi

摘要

This chapter considers the use of attributes of unlearning data to enhance unlearning performance. It studies a framework called “SPlit, UNlearn, MerGE” (Spunge), which splits unlearning data into subsets based on the values of a selected attribute, unlearns each subset separately, and merges the unlearned models. Spunge can be used with any unlearning method to amplify its effectiveness. We empirically demonstrate such improvements for two recent unlearning methods, reducing undesirable behaviors and hazardous knowledge in two popular LLMs.