ISO-Grounded NFR Specification in LLM Code Generation: A Comparative Study

A recent study from arXiv (2608.13742) investigates whether grounding Non-Functional Requirements (NFRs) in the ISO/IEC 25010 Quality Model improves code generation by large language models (LLMs). The research compares three specification styles: a terse one-line baseline (NL-simple), rich natural-language prose (NL-rich), and structured JSON (Structured). Using the HumanEval and HumanEval-ET benchmarks, the study evaluates four NFRs—performance, error handling, code smell, and readability—across ten prompt variations per condition under a fixed model snapshot, with paired non-parametric analysis. The primary finding indicates that ISO-grounded enrichment improves static quality proxies, such as reducing unreadability density across all four NFRs (e.g., performance unreadability drops from 0.88 to 0.69 for NL-rich), and reduces sensitivity to prompt wording. However, it does not reliably improve functional correctness; notably, for error handling, extended-test pass rate decreases, suggesting a tension between defensive coding patterns and exact test expectations. The study was announced as a cross-type on arXiv, indicating its relevance to both AI and software engineering communities.