LLMs Reveal Detailed Visual Cortex Selectivity

Takuya Matsuyama et al. (arXiv v2, Mar 9, 2026) present LLM-assisted Visual Cortex Captioning (LaVCa), which uses large language models to generate natural-language captions for images that selectively activate individual voxels. They report LaVCa produces captions that more accurately and quantitatively capture voxel selectivity and fine-grained properties across ROIs than prior methods. This reveals intra-voxel concept multiplicity and suggests LLM-based descriptions can clarify human visual representations.