Zhanhao Hu

Safety Offset Learning: When Benign Fine-Tuning Misaligns LLMs

According to the abstract, Jesson Wang, Zhanhao Hu, Minghao Lin and colleagues argue that benign fine-tuning can erode safety by teaching a refusal-to-compliance transition. Fine-tuning on benign requests a model tends to refuse increased harmful compliance, but only when the discrepancy is large and has no narrow shortcut.