计算机科学
自动推理
软件工程
自然语言理解
杠杆(统计)
人工智能
知识表示与推理
自然语言
异步通信
图形
模式(遗传算法)
可扩展性
领域(数学)
背景(考古学)
知识库
程序设计语言
水准点(测量)
语义学(计算机科学)
语义映射
作者
Yan Gao,Fuji Hu,Chengzhang Chai,Yiwei Weng,H. S. Li
标识
DOI:10.1016/j.autcon.2026.106888
摘要
Despite recent advances in large language model (LLM)-based agents, their applications to schema-guided reasoning and interaction with Building Information Modelling (BIM) data remain limited. This paper presents IFC-Agent, a tool-augmented multi-agent framework that enables natural language querying, reasoning, and modification of Industry Foundation Classes (IFC) models. The framework integrates schema-guided traversal with LLM-driven dynamic tool composition to achieve interpretable and scalable reasoning workflows. An adaptive dual-mode execution strategy combines asynchronous parallel execution with exploration–aggregation reasoning, while a dual-memory mechanism ensures operational efficiency and semantic consistency. A prototype system was developed using LangChain and validated on benchmark IFC queries. Case studies demonstrate strong performance in field querying, multi-hop reasoning, coordinate transformation, and IFC graph construction. Comparative evaluation with recent graph-based and LLM-assisted BIM frameworks highlights IFC-Agent's advantages in schema-guided reasoning and natural language interaction, establishing a practical foundation for explainable and automated BIM workflows. • Schema-guided multi-agent framework with dynamic tool composition on native IFC data. • Adaptive dual-mode execution addressing interpretability-scalability trade-off. • Dual-memory integrating procedural context and entity caches for cross-agent reuse. • LLMs leverage implicit schema awareness for autonomous multi-hop IFC traversal. • Tools extend execution capabilities but cannot replace LLM's reasoning capacity.
科研通智能强力驱动
Strongly Powered by AbleSci AI