Abstract

Understanding what individual neurons encode is a central challenge in neuroscience. We develop an automated framework that describes neural selectivity in natural language rather than hand-crafted mathematical models. Using digital twins of macaque V1 and V4, we build a three-stage pipeline that converts images to text descriptions, synthesizes these into semantic hypotheses about what drives or suppresses each neuron, and validates predictions by generating novel images. Synthesized images based on linguistic descriptions drove 96.1% of V4 neurons to extreme activation levels. V1 neurons showed similar results but were less describable through language for suppression patterns. We show that linguistic compression is lossy yet semantically faithful — semantic descriptions preserve the geometric structure of neural relationships. This work demonstrates how AI can accelerate neuroscientific discovery while remaining interpretable and verifiable.