Although Vision-Language Models (VLMs) have achieved remarkable progress, they remain limited in performing fine-grained spatial reasoning at the region level—especially in complex environments such as warehouses—due to their reliance on coarse region representations and the lack of effective multimodal interaction mechanisms between RGB and depth modalities. To address this issue, we propose a dedicated Region Feature Enhancer module and conduct a comparative study of two architectural paradigms: an unstructured approach based on Transformers and a structured approach using Graph Neural Networks (GNNs). Our method is further optimized by integrating an auxiliary Region Classifier and a multi-stage training paradigm. Experiments on the PhysicalAI Spatial Intelligence Warehouse dataset show that while both approaches significantly outperform the baseline, the Transformer-based variant consistently delivers better performance across most tasks. This finding suggests that the flexibility of the self-attention mechanism provides an advantage over explicit geometric structures in spatial reasoning contexts, offering important insights for future VLM architecture design.