Knowledge assessment and enhancement pathways for large language model in architectural design

The rapid advancement of large language models (LLMs) is creating significant opportunities for integrating artificial intelligence into architectural design. This study evaluates six representative foundation LLMs on their proficiency in architectural knowledge. Results show that LLMs excel in handling basic architectural concepts in pure text form and demonstrate strong single-modality reasoning. In cross-modal tasks, they effectively recognize and reason about concrete architectural images but struggle significantly with abstract diagrammatic symbols – often producing structurally coherent yet occasionally hallucinated responses. Core limitations center on three areas: cross-modal cognition, depth of domain-specific knowledge, and the orchestration of external design tools. To address these gaps, the paper outlines four technical enhancement pathways: prompt engineering, domain-specific model training, retrieval-augmented generation, and multi-agent collaboration, analyzing the applicability, strengths, and constraints of each. Building on this analysis, the study identifies critical challenges in deploying LLMs for real-world design support and proposes future research directions: embedding cross-modal understanding of architectural graphics and abstract concepts, integrating diverse tools into complex workflows, constructing high-quality, architecture-specific datasets, and exploring LLMs’ potential for creative reasoning beyond conventional aesthetic norms. This study aims to support future research on large language models for architectural design.