# Revisiting Who's Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective

**Authors:** Yujian Liu, Yang Zhang, Tommi Jaakkola, Shiyu Chang  
**arXiv:** [2407.16997](https://arxiv.org/abs/2407.16997)  
**Date:** 2024-07-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.16997v1/

## Key points

- Authors investigate targeted unlearning within LLMs
- Study conducted in two main steps
- Introduce novel task of targeted unlearning
- Goal to remove information about a specific target from documents
- Criteria for successful unlearning established
- Proposed framework for achieving targeted unlearning
- Simple algorithm derived from the framework
- Comprehensive evaluations designed to assess efficacy of targeted unlearning
- Experiments on existing and new datasets demonstrate effectiveness without explicit optimization for predefined criteria
- Research contributes to advancing understanding and application of targeted unlearning within LLMs from a causal intervention perspective

## Summary

In this paper, the authors investigate , specifically focusing on . The study is conducted in two main steps. Firstly, they introduce a novel task of targeted unlearning, where the goal is to remove information about a specific target (e.g., a person) from a set of documents. They establish criteria for successful unlearning and propose a  for achieving it. This framework not only justifies but also extends their method by deriving a simple algorithm that encompasses it as a special case. Furthermore, comprehensive  are designed to assess the efficacy of targeted unlearning. Experiments on both existing and new datasets demonstrate its effectiveness without explicitly optimizing for predefined criteria. The authors provide their  for further exploration. Overall, this research contributes to advancing understanding and application of targeted unlearning within LLMs from a causal intervention perspective.