Answers are grounded in this paper and cite their sources.
General large language models (LLMs) often have limited reliability when addressing specialized bridge pavement maintenance tasks. This study proposes Bridge-RoadLM, a domain-adapted framework for preliminary bridge pavement maintenance support. We developed an automated data engineering pipeline to construct the Bridge-Road Corpus, comprising 10,000 instruction-following samples derived from authoritative specifications and technical documents. Using parameter-efficient fine-tuning (PEFT/LoRA), representative Qwen, DeepSeek, LLaMA, and Gemma models were adapted to bridge pavement maintenance tasks. A multilevel evaluation framework was established, including automated linguistic metrics, a 500-question expert-validated test bank, and a blinded multiengineer evaluation of open-ended inspection-report-style cases. Results show improved controlled task performance for several fine-tuned models, and the strongest improvement was for Qwen3-8B. Its accuracy in pavement defect diagnosis increased from 45.0% to 81.0%, with 95% Wilson confidence intervals of 35.6%–54.8% and 72.2%–87.5%, respectively. Its accuracy in bridge monitoring and maintenance increased from 25.0% to 75.0%, with 95% Wilson confidence intervals of 17.5%–34.3% and 65.7%–82.5%, respectively. In the blinded open-ended evaluation, six engineers with 10–17 years of professional experience scored anonymized model outputs using five predefined criteria: correctness, completeness, traceability, hallucination control, and safety. Qwen3-8B achieved the highest average score among the nine evaluated models. The results also show that fine-tuning effects varied across model families, indicating that domain-specific PEFT does not uniformly improve all LLMs. Overall, this study provides a specification-based instruction corpus and a controlled evaluation workflow for adapting locally deployable LLMs to bridge pavement maintenance, supporting preliminary specification referencing, defect interpretation, and draft-level maintenance-plan preparation within a human-in-the-loop workflow.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: