LLM-AQuA-DiVeR: LLM-Assisted Quality Assurance Through Dialogues on Verifiable Specification with Requirement Owners

Quality Assurance (QA) is important for verifying software compliance with stakeholder requirements. QA faces a fundamental challenge of requirement interpretation ambiguity, which can result in insufficient software verification and failure in achieving the stakeholders' intended quality. The interpre-tation challenge intensifies in software development driven by Large Language Models (LLMs), where over-reliance can lead to missed quality-critical alternatives. However, existing works have paid limited attention to stakeholder involvement. We propose an LLM-assisted QA framework extending conventional LLM-driven development to enable stakeholder engagement in software verification. Our framework employs formal methods and rigorous testing to meet diverse quality demands, though this comprehensive verification introduces technical complexity affecting stakeholder engagement and verification costs. Our framework addresses these challenges through two key LLM roles: 1) an explanation assistant for stakeholder understanding, 2) a refinement assistant for incorporating stakeholder feedback while maintaining feasible verification costs. Our initial evaluation empirically demonstrates the framework's effectiveness through participant assessment scores, showing improved quality risk comprehension and efficient feedback incorporation in the verification process.