Low-Rank

A training-free method that robustifies LLM safety alignment against fine-tuning by extrapolating low-rank safety subspaces, significantly reducing attack success rates while preserving model utility.