Your generation model is multimodal-in, text-out only,...

You can derive localization signals as a byproduct of the model’s existing architecture. By extracting cross-attention weights or forcing the model to emit coordinates, you gain a grounding signal. However, because these signals are not inherently perfect, they must be passed through a verification gate to confirm that the identified region actually supports the generated claim.