Large-scale visual-language models (LVLMs) demonstrate powerful performance on multimodal benchmarks, but suffer from limitations in structural inference and accurate grounding. This study investigates the phenomenon that adding simple visual structures (e.g., segmentation, annotation, etc.) improves accuracy and proposes the concept of "grounding IDs," latent identifiers induced by external cues. Grounding IDs link entities to specified segments across modalities. Representation analysis reveals that these identifiers exhibit robust intra-segment alignment in embedding space, bridging the modality gap between images and text. Causal intervention confirms that these identifiers mediate the association between objects and symbolic cues. Grounding IDs enhance cross-modal grounding and reduce hallucinations by enhancing attention between relevant components. Our findings reveal that grounding IDs are a key symbolic mechanism by which external cues enhance multimodal binding, offering both interpretability and substantial robustness improvements.
