Tag
ConceptFormer learns continuous latent concept representations for visual document retrieval, bridging visual evidence and semantic relevance without text intermediates, achieving significant improvements over baselines.