Scholarlink

As building regulations grow increasingly complex, large language models (LLMs) offer the potential to streamline compliance processes. However, questions remain about their accuracy, reliability, and contextual understanding. This study evaluates three LLM-driven tools: STRUCK, BR18.Chat, and ChatGPT, by benchmarking their outputs against 143 expert-validated regulatory questions derived from BR18. Using a mixed-methods approach combining quantitative analysis with expert evaluation and case studies, we measure accuracy, interpretability, and practical applicability. Results reveal interpretive inconsistencies and varying degrees of regulatory precision. Our findings support the integration of LLMs into compliance workflows, provided professional oversight is maintained, and suggest design improvements for domain-specific AI solutions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating LLM Accuracy in Referencing Danish Building Regulations

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dorthe Holmberg Lauritzen,
  • Peter Nørkjær Gade,
  • Judith Fauth

摘要

As building regulations grow increasingly complex, large language models (LLMs) offer the potential to streamline compliance processes. However, questions remain about their accuracy, reliability, and contextual understanding. This study evaluates three LLM-driven tools: STRUCK, BR18.Chat, and ChatGPT, by benchmarking their outputs against 143 expert-validated regulatory questions derived from BR18. Using a mixed-methods approach combining quantitative analysis with expert evaluation and case studies, we measure accuracy, interpretability, and practical applicability. Results reveal interpretive inconsistencies and varying degrees of regulatory precision. Our findings support the integration of LLMs into compliance workflows, provided professional oversight is maintained, and suggest design improvements for domain-specific AI solutions.