A review of the development, training, and application scenarios of multimodal large models for construction engineering

Purpose To address fragmented research and the lack of integrated development frameworks for multimodal large models (MLMs) in construction engineering, this study employs a systematic literature review approach and proposes a structured framework focusing on the design and construction stages. Design/methodology/approach This study examined literature on large models (LMs) and MLMs for Construction Engineering indexed in the Web of Science Core Collection database from 2014 to 2026. After multiple rounds of screening and forward-and-backward snowballing, 235 valid documents were obtained. Co-citation analysis, keyword clustering and co-occurrence analysis were used to identify the knowledge base and research frontiers. The studies were further coded and appraised by data modality, model type, engineering task, evidence source, application stage, evaluation method and deployment maturity. Findings This study revealed that research on LMs in construction engineering has developed rapidly, with core literature focused on retrieval-augmented generation (RAG), BIM information retrieval, knowledge graphs, visual perception, multimodal fusion and domain adaptation. Five major toolchain components were identified: multimodal data processing and fusion, model training and domain adaptation, automated annotation, model evaluation, and engineering deployment. Four application scenarios were summarized: generative AI design, embodied construction robots, multimodal question-answering (Q&A) robots, and risk prediction systems. Key challenges include data heterogeneity, knowledge timeliness, hallucination control, interoperability, computing cost and responsibility governance. Originality/value This study innovatively integrated systematic literature review, bibliometric visualization and maturity-oriented evidence coding, proposing a development-training-application framework for MLMs for Construction Engineering. Unlike single-technology or application-focused reviews, this framework integrates data processing, model training, automated annotation, evaluation, deployment and scenario application, thereby overcoming fragmented limitations and supporting scalable, trustworthy and domain-adapted applications of MLMs for Construction Engineering.