Prompt-Optimized LLMs for Requirements Architecture Construction: Empowering Scalable Requirement Analysis in Autonomous Transportation Systems

This research explores leveraging Large Language Models (LLMs) for automated requirements architecture generation in Autonomous Transportation Systems (ATS). Facing challenges like cross-domain complexity, frequent changes, and high error rates with traditional methods, the study proposes a systematic LLM-based framework. The methodology defines a hierarchical Requirements Architecture (N-SR-FR hierarchies with traceability links) based on ARC-IT, utilizes prompt engineering for generation, develops a quality analysis model (Content Analysis + Structure Analysis), and implements prompt optimization for iterative improvement. Experiments confirm LLMs can generate structurally complete drafts with basic traceability within minutes, achieving 100% success rate, showcasing significant efficiency gains. However, generation quality faces a bottleneck: LLMs prioritize semantic coherence over domain precision, leading to generalized descriptions, and rely on surface-level associations for traceability links, causing misalignments. This results in outputs characterized as “Non-standard but Workable” – possessing baseline completeness and core semantic coverage but lacking perfect engineering compliance. This study’s innovations include the first quantitative evaluation of LLMs in ARC-IT, novel evaluation methods, and a prompt optimization pathway achieving up to 39.5% overall quality improvement. It advocates a hybrid workflow: using LLMs to rapidly generate draft prototypes as efficient starting points, followed by domain engineer refinement for precision, logical repair, and domain knowledge infusion, providing a scalable solution for dynamic ATS development.