Comparing LLM-based and MDE-based code generation for agile MDE

Large Language Models (LLMs) are increasingly applied to many different software tasks, and by means of specific domain training, they can perform well in specialised areas. Code-aware LLMs can generate executable program<br/>code from natural language specifications and prompts, but there are questions about the reliability of this process, because the results are not always consistent and correct. Model-driven engineering (MDE) also provides code generation techniques, which produce code automatically based on software designs expressed in UML or OCL. Compared to the LLM solution, the code generation process in MDE is more explicit, and the transformed code is determined by the source model, and there is usually a high correctness level for such generation. The code-generation by example (CGBE) process uses symbolic machine learning to synthesise code generators for particular code generation processes in MDE. The resulting generators map from UML/OCL specifications to the<br/>target programming language. In this paper we compare the effectiveness of code generation using baseline and fine-tuned LLMs, such as DeepSeek and LLama2, with CGBE code generators.