Despite the impressive performance achieved by pre-trained language-and-vision models in downstream tasks, it remains an open question whether this reflects a proper understanding of image-text interaction. In this work, we explore to what extent they handle basic linguistic constructions—active-passive voice, coordination, and relative clauses—that even preschool children can typically master. We present BLA, a novel, automatically constructed benchmark to evaluate multimodal models on these Basic Language Abilities. We show that different types of Transformer-based systems, such as CLIP, ViLBERT, and BLIP2, generally struggle with BLA in a zero-shot setting, in line with previous findings. Our experiments, in particular, show that most of the tested models only marginally benefit when fine-tuned or prompted with construction-specific samples. Yet, the generative BLIP2 shows promising trends, especially in an in-context learning setting. This opens the door to using BLA not only as an evaluation benchmark but also to improve models’ basic language abilities.

Several recent state-of-the-art transfer learning methods model classification tasks as text generation, where labels are represented as strings for the model to generate. We investigate the effect that the choice of strings used to represent labels has on how effectively the model learns the task. For four standard text classification tasks, we design a diverse set of possible string representations for labels, ranging from canonical label definitions to random strings. We experiment with T5 on these tasks, varying the label representations as well as the amount of training data. We find that, in the low data setting, label representation impacts task performance on some tasks, with task-related labels being most effective, but fails to have an impact on others. In the full data setting, our results are largely negative: Different label representations do not affect overall task performance.