LLM-based code generation and system migration in language-driven engineering

Abstract This paper illustrates the power of extending Language Driven Engineering (LDE) with Domain-Specific Natural Languages (DSNLs) through a case study on two levels. Both cases benefit from the characteristic decomposition feature of LDE, resulting in tasks tailored to the application of domain-specific languages, here with a focus on the application of DSNLs supported by LLM-based code generation. In the first case study, we show how DSNL-supported LDE facilitates the development of point-and-click adventures, whereas the second case study focuses on migration: We demonstrate how the entire LDE scenario for point-and-click adventure games can be migrated to output TypeScript instead of JavaScript using LLM-based code generation exclusively, without manually writing any code. This migration not only infers the required types, but also preserves an important property of the original LDE scenario: generated web applications can be automatically validated by design via automata learning and subsequent model checking. Even better, this property can be exploited to automatically validate the correctness of the migration by learning so-called difference automata that characterize the behavioral differences between the generated JavaScript-based and Type-Script-based applications.