Are Multimodal Large Language Models Pragmatically Competent Listeners in Simple Reference Resolution Tasks?
This work addresses the problem of evaluating pragmatic competence in AI models for researchers in natural language processing and multimodal AI, though it is incremental as it focuses on a specific, simple task.
The study investigated multimodal large language models' ability to resolve references using simple visual stimuli like color patches and grids, finding that basic pragmatic capabilities such as context-dependent color interpretation remain major challenges for state-of-the-art models.
We investigate the linguistic abilities of multimodal large language models in reference resolution tasks featuring simple yet abstract visual stimuli, such as color patches and color grids. Although the task may not seem challenging for today's language models, being straightforward for human dyads, we consider it to be a highly relevant probe of the pragmatic capabilities of MLLMs. Our results and analyses indeed suggest that basic pragmatic capabilities, such as context-dependent interpretation of color descriptions, still constitute major challenges for state-of-the-art MLLMs.