World-to-Words: Grounded Open Vocabulary Acquisition through Fast Mapping in Vision-Language Models

38823

Related Entities

No entities found for this article.