LLM-based Unit Test Generation for Dynamically-Typed Programs

Abstract: Automated unit test generation has been widely studied, but generating effective tests for dynamically typed programs remains a significant challenge. Existing approaches, including search-based software testing (SBST) and recent LLM-based methods, often suffer from type errors, leading to invalid inputs and assertion failures, ultimately reducing testing effectiveness. To address this, we propose TypeTest, a novel framework that enhances type correctness in test generation through a vector-based Retrieval-Augmented Generation (RAG) system. TypeTest employs call instance retrieval and feature-based retrieval to infer parameter types accurately and construct valid test inputs. Furthermore, it utilizes the call graph to extract richer contextual information, enabling more accurate assertion generation. In addition, TypeTest incorporates a repair mechanism and iterative test generation, progressively refining test cases to improve coverage. In an evaluation on 125 real-world Python modules, TypeTest achieved an average statement coverage of 86.6% and branch coverage of 76.8%, outperforming state-of-theart tools by 5.4% and 9.3%, respectively.