Data-centric evaluation of reference-guided multimodal models for reliable railway fastener inspection

Reliable railway fastener inspection is essential for maintaining infrastructure safety and supporting timely maintenance decisions. Large multimodal models (LMMs) offer a promising alternative to conventional supervised approaches; however, their reliability for fine-grained railway inspection remains poorly understood. Here, we present a data-centric conceptual framework for the systematic cross-paradigm evaluation of a supervised InceptionV3 baseline and large multimodal models under no-reference and reference-guided inspection settings. We further investigate the influence of data partitioning, structural consistency, and Chain-of-Thought (CoT) reasoning on multimodal inspection performance. The supervised baseline achieved stable and reliable results under conventional training conditions, whereas LMMs showed limited capability in no-reference settings. Introducing structurally consistent same-type reference images substantially improved the performance of several multimodal models, particularly under limited-data conditions. However, the effects of CoT reasoning were strongly model-dependent, yielding benefits for some models while degrading performance in others. Collectively, the findings suggest that, under the evaluated experimental conditions, multimodal inspection reliability is influenced more strongly by data-centric factors, including reference alignment, structural consistency, and data availability, than by model complexity alone. These results support the adoption of data-centric and deployment-oriented strategies for developing robust and reliable multimodal railway inspection systems.