The underlying mechanism by which large language models generate code is conditional probability prediction based on massive training corpora, and the model does not truly understand code logic