Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearning
This addresses the critical safety issue of unlearning hazardous facts in LLMs for AI security, representing a novel method rather than an incremental improvement.
The paper tackled the problem of removing dangerous knowledge from language models without harming general performance, achieving over 30x greater reduction in post-attack accuracy and 30x less disruption compared to the best baseline while using less than 3 GPU-seconds per fact.
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.