Applying Genetic Improvement Techniques for Automated Program Repair of Transpiled Code

We use Genetic Improvement (GI)-based Automated Program Repair (APR) techniques for syntax correction on transpiled code produced by both large language models (LLMs) and rule-based translators. A three-stage pipeline was developed, combining rule- based test generation, an optional LLM-driven preprocessing stage for syntax correction, and GI-based repair strategies. LLM-assisted Type Change Operator and Boolean Value Change Operator were added to the MAGPIE GI framework, which reduced transpilation bugs from Python to Java by 33% (LLM) and by 18% on rule-based translations. A comprehensive taxonomy of common transpilation bugs was developed, mapping faults to mutation operators, alongside an evaluation of the effectiveness of secondary LLM interventions.

In International Workshop on Genetic Improvement@ ICSE