Abstract

BIM design workflows involve complex, multi-step interactions with professional software, which limit the direct applicability of large language models (LLMs). This paper presents a BIM design assistant driven by LLM agents, enabling natural-language-based assistance for BIM design and visualization tasks through reliable tool invocation. First, a two-layer Revit interface function library is designed to bridge LLM agents and BIM software, providing secure, extensible, and fine-grained model manipulation capabilities. Based on the ReAct paradigm, an agent framework is developed that integrates dialogue history management, structured prompt engineering, and step-by-step reasoning to support long-horizon and multi-turn BIM tasks. To evaluate the proposed system, an agent-oriented BIM task benchmark is constructed, covering information querying, model visualization, component modification, and deletion, with tasks further categorized by difficulty. Experimental results with three representative LLMs show that high-performance general-purpose models can reliably act as BIM design assistants within the proposed framework, achieving up to 90% task accuracy. These results demonstrate the feasibility and reliability of LLM-agent-driven BIM design assistance in practical scenarios.

Download paper here

Download preprint here

The authors are grateful for the financial support received from the National Natural Science Foundation of China (No. 52378306).

Financial Sources: