Researchers including Amirmohammad Izadi and colleagues have identified a mechanism they call 'Grounding IDs' to explain why simple visual cues improve large vision-language models (LVLMs). In a paper accepted at ICML 2026, they show that external cues such as symbols or grid lines cause the model to form latent identifiers that bind related visual and textual elements.

These Grounding IDs appear when cues partition the input into structured groups. The authors find that such partition-inducing cues produce stronger alignment between corresponding visual and text representations, and that attention is enhanced between related components. This leads to better cross-modal grounding and fewer hallucinations across a range of multimodal tasks.

The work offers a concrete explanation for a previously vague effect: rather than merely focusing the model's attention, external cues create an internal representational structure. The authors suggest that understanding Grounding IDs could guide the design of prompts or interfaces that make LVLMs more reliable.