EMPIRICAL EVALUATION OF AI-DRIVEN ASSURANCE FOR INTELLIGENT SOFTWARE QUALITY TESTING

Abstract

Recent advancements in large language models (LLMs) and generative AI have created new opportunities for intelligent software quality assurance (SQA). While conceptual frameworks such as the AI-Driven Assurance Framework (AIDAF) define adaptive and explainable assurance architectures, few studies have experimentally validated their performance using real AI models. This paper presents an empirical evaluation of AIDAF, implemented with GPT-4o, CodeBERT, and Gemini 1.5 Pro for automated test generation, semantic validation, and defect prediction. The experimental setup employs Python-based integration of public APIs and transformer pipelines to execute test cases for representative web and mobile application modules. Comparative analysis demonstrates that GPT-4o achieved the highest assurance accuracy with precision = 0.91, recall = 0.87, and F1-score = 0.89, outperforming CodeBERT (F1 = 0.83) and Gemini 1.5 Pro (F1 = 0.86). These results confirm that AI-driven assurance can achieve human-comparable quality evaluation with significantly reduced manual effort. The study highlights key strengths, limitations, and future enhancement areas for integrating multi-model intelligence into continuous quality pipelines.