Abstract: The rapid advancement of large code models has blurred the boundaries between traditional programming languages, enabling models to comprehend, generate, and seamlessly translate code across diverse language paradigms. This evolution raises a profound question: much like human languages, do programming languages possess “family” relationships based on structural and semantic proximity? Furthermore, is there a universal “lingua franca” among them? If we could map the genealogical spectrum of programming languages—akin to linguists charting the Indo-European language family—could we leverage this topology to optimize the training of large language models for greater intelligence and efficiency? Our recent research provides a definitive answer. For the first time, we employ a data-driven approach to automatically mine the underlying “family tree” of 19 mainstream programming languages. By integrating these genealogical relationships to guide LLM training, our proposed method demonstrates significant performance improvements across four critical downstream tasks: code summarization, code search, code generation, and code translation.