LLM-Based Adaptive Exploration for BIM Information Extraction
A novel approach known as adaptive exploration employs an LLM-based agent to retrieve data from BIM models by iteratively running code and identifying structure during execution, thus addressing the challenges posed by the heterogeneity of BIM that static query generation cannot manage. This method is assessed using ifc-bench v2, a newly introduced open-source benchmark for BIM question-answering, which includes 1,027 tasks derived from 37 IFC models across 21 projects. A factorial ablation study involving two levels of LLM capabilities and four augmentation strategies reveals that adaptive exploration consistently surpasses static query generation in all tested configurations.
Key facts
- Adaptive exploration uses an LLM-based agent to extract information from BIM models.
- The agent iteratively executes code to discover structure at runtime.
- Static query generation assumes a fixed data organization, which BIM heterogeneity invalidates.
- ifc-bench v2 is an open-source BIM question-answering benchmark introduced alongside this work.
- The benchmark comprises 1,027 tasks across 37 IFC models from 21 projects.
- A factorial ablation across two LLM capability levels and four augmentation strategies was conducted.
- Adaptive exploration significantly outperforms static query generation across all configurations.
Entities
—