An Adaptive LLM-Based Agent Framework with Self-Extending Tool Library for Automated Building Code Compliance Checking
Year: 2025
Type: article
Abstract: Building Information Modeling (BIM)-based automated compliance checking has been extensively studied in the Architecture, Engineering, and Construction (AEC) industry. However, most existing approaches rely on static, rule-based workflows that assume regulatory interpretation and BIM data alignment are fully resolved prior to execution, limiting their applicability across heterogeneous Industry Foundation Classes (IFC) models and diverse regulations. This thesis proposes a Large Language Model (LLM)-based agentic framework that treats compliance checking as an iterative, evidence-driven reasoning process. Based on the Reasoning and Acting (ReAct) paradigm, the system enables an autonomous agent to reason directly over regulatory text and incrementally explore IFC models through tool-based interactions. By supporting adaptive planning, flexible access to heterogeneous model representations, and controlled extension of computational capabilities, the framework enables regulatory ambiguity and imperfect BIM data to be handled within the checking process. The framework was evaluated using IFC models with diverse modeling practices and data quality, together with regulatory clauses of varying complexity. The results demonstrate that the system achieves high performance, reaching a weighted-average F1 score of 0.976, while maintaining stable execution under realistic and imperfect data conditions. Overall, this work shows that agentic reasoning provides a viable foundation for more generic and adaptable BIM-based automated compliance checking.
Author: Li, Y.
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7137466442