aiqbench analysis

This article investigates whether large language models possess an inherent visual understanding by analyzing their performance on tasks involving Chinese characters, which are logographic rather than alphabetic. The study reveals that these models can learn to "see" character structures through text-only training data due to a specific inductive bias toward visual patterns embedded in the writing system. Consequently, this research highlights how non-visual modalities like language processing can implicitly capture spatial and structural information relevant to computer vision tasks.