LLM-based Automated Test Case Generation Framework for Software Quality Assurance

Abstract

Automated test case generation remains a critical challenge in software engineering, with existing approaches achieving limited branch coverage and mutation scores. Current techniques based on search-based software testing fail to adequately capture semantic information from source code, resulting in test suites that exercise code paths without verifying behavioral correctness. This work addresses these limitations by introducing AutoTest, a framework integrating large language model (LLM) capabilities with coverage-guided test synthesis. The proposed approach employs a multi-stage pipeline comprising repository analysis, semantic summarization, and iterative test refinement driven by an adaptive gradient-based optimization algorithm. Experiments were conducted across five benchmark datasets containing 147 open-source Java projects. AutoTest achieved 87.6% branch coverage and 72.4% mutation score, representing improvements of 11.8 and 10.7 percentage points over the strongest baseline. Convergence was attained within 95 iterations, compared to 247 iterations for the nearest competitor. Statistical analysis confirmed significance at p <0.001 (Cohen's d= 1.17). The results demonstrate that LLM-guided semantic understanding substantially improves automated test generation effectiveness.