A retrieval-augmented LoRA-Tuned LLM framework for structured software test case generation

Abstract Modern software systems are becoming increasingly complex that require scalable and reliable testing mechanisms for quality and robustness. The manual test case design remains a significant limitation as it is time-consuming, inconsistent, and prone to human error when dealing with large-scale and evolving requirements. The existing automated models struggle to generate structured, semantically coherent, and contextually relevant test cases which indicates the need for more intelligent and efficient solutions. This study proposes a novel Retrieval-Augmented and LoRA-tuned Large Language Model for automated generation of structured software test cases from textual requirements. The proposed model combines the FLAN-T5 with Retrieval-Augmented Generation to integrate the contextual grounding and Low-Rank Adaptation for parameter-efficient fine-tuning that allows the high-quality generation to maintain the computational efficiency. The proposed model is trained on a large-scale custom dataset that contain real-world test case records collected from enterprise tools such as Jira and Zephyr, and generates structured JSON-based test cases aligned with industrial standards. The enterprise-style structured test-case dataset and controlled experimental protocol used in this study. The proposed configuration achieved a BLEU score of 0.5897, ROUGE-L of 0.8095, BERTScore F1 of 0.9679, and semantic cosine similarity of 0.9816. A duplicate rate of 0.000 and a hallucination rate of 0.017 were also observed under the defined evaluation criteria. The results indicate the potential of retrieval augmentation and parameter-efficient fine-tuning for structured test-case generation within the investigated setting. The enterprise dataset and trained artifacts are restricted and findings interpreted within the scope of the evaluated task rather than as evidence of universal performance superiority. The integration of retrieval augmentation results in a mean inference latency of approximately 6·27s per test case indicates a measurable computational overhead with the improved contextual grounding. The findings show a practical performance between generation quality and inference efficiency within the evaluated enterprise-style test-case generation setting.