Abstract: With a focus on the issue of whether vision-language models (VLMs) can provide substantial gains beyond traditional deep vision for welding defect detection, relevant literature was systematically analyzed within the scope of three types of detection objects: surface defects, internal defects, and weld formation anomalies. The results indicate that this research direction is still in the early exploratory stage, is complementary to existing defect detection algorithms, and has great development potential. The advantages of existing VLMs are manifested in high-level tasks such as semantic interpretation, few-shot recognition guidance, and structured report generation; the disadvantages are manifested as follows: insufficient feature extraction ability for non-natural image data such as X-rays and ultrasound; detection accuracy in fine-grained pixel localization lagging behind object detection algorithms based on CNNs; large inference latency in real-time edge deployment scenarios. Therefore, it is considered that a more reasonable engineering path at present is a hybrid complementary architecture: Traditional vision methods are responsible for precise localization and real-time front-end detection, while VLM is responsible for upper-layer semantic interpretation and structured report output, forming a hierarchical collaborative relationship rather than a replacement relationship between the two.