{"ID":2870872,"CreatedAt":"2026-06-01T04:54:23.091178241Z","UpdatedAt":"2026-06-01T04:54:23.091178241Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2509.11816","arxiv_id":"2509.11816","title":"Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearning","abstract":"Current unlearning and safety training methods consistently fail to remove dangerous knowledge from language models. We identify the root cause - unlearning targets representations which are too general - and develop a highly selective technique that unlearns robustly while preserving general performance. Our method performs PCA on activations and module-output gradients to identify subspaces containing common representations, then collapses these subspaces before computing unlearning updates, a technique we term Collapse of Irrelevant Representations (CIR). This avoids unlearning general knowledge and targets only representations specific to the facts being unlearned. When unlearning bio- and cyber-hazardous facts from Llama-3.1-8B, we achieve over 30x greater reduction in post-attack accuracy than the best baseline (Circuit Breakers), while disrupting general performance 30x less, and using less than 3 GPU-seconds per fact. Thus, by disentangling harmful and benign capabilities at the level of representations, CIR enables robust and non-disruptive unlearning.","short_abstract":"Current unlearning and safety training methods consistently fail to remove dangerous knowledge from language models. We identify the root cause - unlearning targets representations which are too general - and develop a highly selective technique that unlearns robustly while preserving general performance. Our method pe...","url_abs":"https://arxiv.org/abs/2509.11816","url_pdf":"https://arxiv.org/pdf/2509.11816v2","authors":"[\"Filip Sondej\",\"Yushi Yang\"]","published":"2025-09-15T11:55:10Z","proceeding":"cs.LG","tasks":"[\"cs.LG\",\"cs.AI\",\"cs.CL\"]","methods":"[\"Large Language Model\",\"Language Model\"]","has_code":false}