Applying sparse autoencoders to unlearn knowledge in language models
This addresses the challenge of knowledge unlearning in language models for AI safety, but it is incremental as it builds on existing methods like Representation Misdirection.
The paper tackled the problem of removing specific knowledge from language models using sparse autoencoders (SAEs), finding that negative scaling of feature activations can unlearn biology-related questions with minimal side-effects, but SAE-based methods currently underperform compared to fine-tuning techniques.
We investigate whether sparse autoencoders (SAEs) can be used to remove knowledge from language models. We use the biology subset of the Weapons of Mass Destruction Proxy dataset and test on the gemma-2b-it and gemma-2-2b-it language models. We demonstrate that individual interpretable biology-related SAE features can be used to unlearn a subset of WMDP-Bio questions with minimal side-effects in domains other than biology. Our results suggest that negative scaling of feature activations is necessary and that zero ablating features is ineffective. We find that intervening using multiple SAE features simultaneously can unlearn multiple different topics, but with similar or larger unwanted side-effects than the existing Representation Misdirection for Unlearning technique. Current SAE quality or intervention techniques would need to improve to make SAE-based unlearning comparable to the existing fine-tuning based techniques.