进度管理是施工管理的重要组成部分, 包含监控实际施工进度并与施工计划进行比较。传统的施工进度更新方法依赖施工人员的手动记录, 滞后性高, 更新频率慢且易出错。为解决该问题, 该文提出了一种基于建筑信息模型(BIM)和大语言模型的施工进度更新方法, 使施工人员能够通过口头上报施工进度的方式来更新建筑三维模型。通过解析工业基础类(IFC)格式的BIM文件和施工计划, 从中提取并关联构件的ID、位置信息及计划施工时间等内容, 进而建立建筑构件数据库; 通过对大语言模型进行提示词微调, 使其能够根据自然语言输入在数据库中检索构件信息和判断进度状态, 并指导Blender中的模型进行动态更新。测试结果显示：经过提示词微调后, 大语言模型在进度判断和更新指令生成方面的平均准确率达到96%, 一致性达到87%, 验证了所提出的更新方法的有效性和可行性。

Objective: Progress management is an important part of construction management, which helps effectively reduce the risk of project delay. Its main objective is to monitor the actual construction progress and compare it with the construction plan. The traditional method of progress updating relies on manual checking and recording, which not only lags behind but also is prone to recording errors. Following the development of building information modeling (BIM), technologies such as the internet of things (IoT), point clouds, and visual images have been gradually applied to construction progress identification and plan comparison. However, these methods require the introduction of additional acquisition equipment, and point cloud acquisition equipment is costly. In addition, image processing is easily affected by factors such as occlusion, light, and weather. Therefore, the present study proposes a construction progress updating method based on BIM and large language models (LLMs). This approach enables construction personnel to verbally report progress information to the LLM, allowing a three-dimensional (3-D) building model to be accordingly updated. Methods: This research develops a system that automatically extracts relevant information from natural language, which identifies the corresponding component using the planned construction time in the component database and visualizes the progress status of the 3-D building model in Blender. The system does not require detailed information such as precise component IDs but completes the progress update by recognizing fuzzy information (e.g., construction section, floor, and other relevant information). Specifically, this study first parses the industry foundation classes (IFC) format BIM file and construction schedule to extract and correlate the component IDs, location information, and scheduled construction time. It then constructs a database of building components. Subsequently, the LLM is enhanced through prompt engineering so that it can generate accurate information query instructions based on natural language inputs, retrieve component information from the database, assess the progress status, and generate corresponding model update instructions to achieve dynamic updates in Blender. Results: This study tested the accuracy and consistency of the proposed method using a BIM model with 716 components and a dataset of 200 progress reports in various natural language formats. The testing results showed that after prompt fine-tuning, the LLM-based method achieved an average accuracy of 96% in progress assessment and model updating and 62.5% improvement over the non-fine-tuned model. The consistency reached 87% or an increase of 68% over the non-fine-tuned model, demonstrating the effectiveness and feasibility of this method for construction progress updating. Conclusions: This study has successfully combined BIM and LLMs to develop a construction progress updating method, including the construction of component retrieval database and schedule updating process based on LLMs. The case studies show that the method effectively improves the accuracy and consistency of the LLM in generating progress update instructions without providing additional equipment and significant computing costs. The method allows construction personnel to describe progress information in natural language and achieves accurate progress updating of the 3-D model of a building, which meets the demand for visualizing and updating progress information on construction sites. However, this study suffers from certain limitations. Due to the use of prompt fine-tuning for the LLM, consistency remains a challenge. Future work is expected to improve the model's accuracy and consistency by training a local model.