State-of-the-Art Pre-Trained LLMs for Construction Cost Estimation Tasks: A Conceptual Estimation Scenario Using a Modular Chain of Thought (CoT) Approach

The traditional cost estimation process in construction involves extracting information from diverse data sources in various data forms and often relying on human intuition and judgment. This makes the estimation process time-intensive, subjective, and susceptible to human errors. The advancement of large language models (LLMs) offers a promising avenue to address these inefficiencies; however, their effectiveness in cost estimation tasks remains unexplored. There are previous studies that have explored LLM applications in different construction problems. However, no study has evaluated how existing pre-trained LLMs perform in cost estimation or provided a structured framework for enhancing their accuracy and reliability through prompt engineering to automate estimation workflows. In our previous study, we identified the key estimation burdens and categorized them into three distinct areas of estimation: (1) conceptual estimation, (2) evaluating subcontractor estimates, and (3) change management, version control, and data recycling. This study specifically evaluates the performance of LLMs, including GPT-4o, LLaMA 3.2, Gemini 2.0, and Claude 3.5 Sonnet, on cost estimation tasks under a conceptual estimation scenario to compare zeroshot prompting with a modular chain-of-thought (CoT) framework. The results of this study indicate that zero-shot prompting produced incomplete and inconsistent responses, with an average confidence score of 1.906 (64%). The CoT framework significantly improved accuracy, achieving a confidence score of 2.52 (84%) and significant improvements across quantitative measurements such as BLEU, ROUGE-L, METEOR, content overlap, and semantic similarity metrics. This study makes three key contributions: (i) develops a cost estimation scenario for LLMs, (ii) evaluates the baseline performance of LLMs in construction cost estimation, and (iii) develops a modular CoT framework that improves estimation performance, which offers a scalable, adaptable approach for industry application. The findings serve as a guideline for integrating AI into construction estimation workflows, bridging the gap between pre-trained LLMs and construction industry applications.