SAEs Can Improve Unlearning: Dynamic Sparse Autoencoder Guardrails for Precision Unlearning in LLMs

Aashiq Muhamed, Jacopo Bonato, Mona T. Diab, Virginia Smith

Keywords: unlearning, sparse autoencoders, representation editing, steering vectors, relearning

TL;DR: We introduce Dynamic SAE Guardrails, a computationally efficient method that leverages sparse autoencoders for precision unlearning in LLMs, outperforming existing approaches while maintaining model utility.

Abstract: Machine unlearning is a promising approach to improve LLM safety by removing unwanted knowledge from the model. However, prevailing gradient-based unlearning methods suffer from issues such as high computational costs, hyperparameter instability, poor sequential unlearning capability, vulnerability to relearning attacks, low data efficiency, and lack of interpretability. While Sparse Autoencoders are well-suited to improve these aspects by enabling targeted activation-based unlearning, prior approaches underperformed gradient-based methods. This work demonstrates that, contrary to these earlier findings, SAEs can significantly improve unlearning when employed dynamically. We introduce Dynamic SAE Guardrails (DSG), a novel method for precision unlearning that leverages principled feature selection and a dynamic classifier. Our experiments show DSG substantially outperforms leading unlearning methods, achieving superior forgetting-utility balance. DSG directly addresses the core drawbacks of gradient-based approaches, offering enhanced computational efficiency and stability, robust performance in sequential unlearning, stronger resistance to relearning attacks, better data efficiency, including zero-shot settings, and more interpretable unlearning.

Submission Number: 3