Researchers have proposed a new method called RepSelect for robustly removing unwanted knowledge and tendencies from large language models (LLMs). This is done by selectively editing specific representations in the model's weights to prevent an adversary from recovering the removed information. The authors tested RepSelect on five datasets covering various types of unwanted content, including biohazards, cyber threats, and abusive language. Compared to existing unlearning methods, RepSelect achieved a 4-40x larger drop in post-relearning answer probability, indicating that it is more effective at erasing unwanted information without harming the model's general capabilities. --- Why it matters: This matters because current LLM unlearning methods are often shallow and can be easily reversed by an adversary. RepSelect addresses this issue by providing a more robust approach to unlearning, which is essential for applications where sensitive or confidential information needs to be removed from models. Source: https://arxiv.org/abs/2606.17168