Reproduce “Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shengyuan Hu, Yiwei Fu, Zhiwei Steven Wu, Virginia Smith

Researchers at Carnegie Mellon University demonstrated that existing approximate unlearning methods for Large Language Models (LLMs) are vulnerable to "benign relearning attacks," where models can be prompted to regenerate supposedly unlearned harmful or copyrighted content using only loosely related public information. This work reveals that unlearning often results in obfuscation of knowledge rather than its robust erasure, challenging the foundational assumptions of current techniques.

Key insights

  • Current approximate unlearning methods tend to obfuscate or suppress the output of unlearned information rather than robustly erasing its underlying representation from the model's memory.
  • Knowledge within LLMs appears deeply embedded and interconnected; loosely related benign data can reactivate latent associations to recover supposedly forgotten content.
  • Parameter-Efficient Fine-Tuning (PEFT) techniques like LoRA, often used for unlearning, can make models even more susceptible to benign relearning attacks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.