TestLoter: A logic-driven framework for automated unit test generation and error repair using large language models

Abstract

Automated unit test generation is a critical technique for improving software quality and development efficiency. However, traditional methods often produce test cases with poor business consistency, while large language model based approaches face two major challenges: a high error rate in generated tests and insufficient code coverage. To address these issues, this paper proposes TestLoter, a logic-driven test generation framework. The core contributions of TestLoter are twofold. First, by integrating the structured analysis capabilities of white-box testing with the functional validation characteristics of black-box testing, we design a logic-driven test generation chain-of-thought that enables deep semantic analysis of code. Second, we establish a hierarchical repair mechanism to systematically correct errors in generated test cases, significantly enhancing the correctness of the test code. Experimental results on nine open-source projects covering various domains, such as data processing and utility libraries, demonstrate that TestLoter achieves 83.6% line coverage and 78% branch coverage. Our approach outperforms both LLM-based methods and traditional search-based software testing techniques in terms of coverage, while also reducing the number of errors in the generated unit test code.