Abstract: In this paper, we consider the model merging process for large language models (LLMs) under a two-stage optimization framework. Traditional merging methods usually apply fixed blending rates ...