Automated BIM LOD Inferecning: A Hybrid Approach Using Knowledge Graphs and Multimodal Large Language Models

Abstract: Automated Level of Development (LOD) assessment in Building Information Modeling (BIM) is hindered by semantic ambiguity in specifications and geometric complexity of objects. Existing rule-based and vision-based methods lack contextual reasoning or struggle with fragmented, occluded geometry. This study proposes a hybrid framework combining a three-layered knowledge graph—built top-down via LLM requirement extraction and bottom-up via visual population—with a geometric pipeline reassembling meshes and exposing occluded components. A multimodal LLM cross-references graph-derived criteria with multi-view geometry, producing explainable assessments. A four-condition ablation study achieves 83.47% accuracy while reducing false negatives, confirming feasibility for explainable BIM quality assurance.