Are Multimodal Large Language Models Pragmatically Competent Listeners in Simple Reference Resolution Tasks?

TL;DR

This paper evaluates whether multimodal large language models can pragmatically interpret simple visual references, revealing that they still struggle with basic pragmatic tasks despite their advanced capabilities.

Contribution

It provides an empirical assessment of the pragmatic competence of MLLMs in simple reference resolution tasks involving visual stimuli.

Findings

MLLMs face challenges in context-dependent color interpretation.

Basic pragmatic skills remain difficult for current MLLMs.

The study highlights gaps in pragmatic understanding of multimodal models.

Abstract

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.