LLM 기반 BIM 모델링 평가 및 피드백 프레임워크의 설계와 실증적 효과

Empirical Validation of wise‑BIM: A Large‑Language‑Model- Driven Framework for BIM Modeling Evaluation and Feedback

This study proposes wise-BIM, a BIM modeling evaluation and feedback framework that embeds large language models (LLMs), and empirically validates its effectiveness in civil-engineering design and checking scenarios. The framework comprises three stages: (A) structuring decision nodes, (B) defining purpose-centered elements, and (C) a RAG-prompt conversational feedback loop. By integrating static resources with dynamic reasoning, wise-BIM provides contextualized guidance to modelers without training on project files. In a minimum-viable expert study across representative tasks, the LLM-RAG condition, compared with manual work, improved goal refinement, reduced error-detection time, decreased rework cycles, lowered cognitive workload (NASA-TLX), reduced clarification requests (Clarify count), and increased usability (SUS). We also observed model-specific trade-offs between communication efficiency and the depth of goal elaboration, informing practical model-selection strategies. These findings indicate that conversational feedback by LLMs can mitigate the adaptability and explainability limitations of rule-based automation in BIM and provide a human-in-the-loop pathway for quality control during modeling.