Safety Degradation and Preservation During Benign Business Fine-Tuning

While most Large Language Models (LLMs) are designed to be safe, they often need business fine-tuning, which may change the model's inherent safety behaviors and weaken the safety guardrails already put in place, something that can't be fully managed just by tweaking the training data. Thus, it becomes an important problem to preserve the safety properties of a model during the fine-tuning process. This paper summarizes approaches of how a model can maintain safety capabilities while learning new tasks, highlighting the interactions between training data, changes in the internal model, protective measures and final evaluation. It reviews the research literature along multiple dimensions of factors that exacerbate safety, internal model change, preservation of safety and evaluation approaches, and it highlights specific research gaps in the context of the available literature. Existing methods have defined specific protective measures for pre-, fine-tuning and post-training stages, have studied the mechanisms of safety degradation, and have made progress in research and validation from the attack and defense sides. Due to the variety and idealized conditions of experiments, however, additional research is required to account for the applicability of the experiments, the prediction of the degradation phenomena in safety, and safety maintenance in multiple adapter combinations.