Assessing the Viability of LLM Agents for Generating Reusable Compliance Checking Functions

Abstract: The complexity of building codes and variations in Building Information Modelling (BIM) practices have long challenged Automated Compliance Checking (ACC). The success of Large Language Models (LLMs) in code generation opens new opportunities for this domain. Unlike previous work relying on machine-readable rules or LLMs using hard-coded tools that limit scalability, this paper examines whether LLMs, using the Code-Act pattern, can generate scalable compliance logic for regulations up to Solihin and Eastman’s complexity class 3. We show that iterative refinement with a strong verifier evolves reusable checking functions for consistently modelled BIM data, while geometric reasoning requiring auxiliary constructions remains challenging. Source code and data available at: url{https://github.com/stefan-1992/ACC-function-generation.