Relevance prediction between user queries and items is a fundamental part of e-commerce search systems. Traditionally, high-quality relevance labels come from human judgment, which is reliable but expensive and slow, creating a bottleneck for large-scale deployment.

The paper explores an alternative: using a large-scale teacher model to guide the training of smaller, multimodal student models. This knowledge-distillation approach aims to capture the teacher's understanding of relevance while producing students that are efficient enough for practical use.

By shifting away from heavy reliance on human annotation, the method could make it faster and cheaper to update search relevance models, while still leveraging multimodal signals such as text and images. The trade-off between teacher quality and student efficiency is the central focus of the work.