Quantification and object perception in Multimodal Large Language Models deviate from human linguistic cognition

The study published on arXiv investigates the challenges faced by Multimodal Large Language Models (MLLMs) in understanding quantification, a complex linguistic phenomenon. It reveals that MLLMs exhibit clear differences from human cognition in representing quantification, particularly regarding the ordering of quantifiers and biases in numerical perception. By examining these discrepancies, the research aims to enhance our understanding of MLLMs as semantic and pragmatic agents. This exploration is vital for advancing AI language models, as it highlights the need for improvements in their architecture to better align with human linguistic capabilities.