Tag
This paper develops a benchmark to evaluate the multilingual ability of vision-language models to use spatial deictic expressions (e.g., 'this', 'that') across four languages, revealing that models differ from humans in selecting appropriate demonstratives based on distance.