Neurons Map the Cellular Grammar of Human Language

Recordings from 579 neurons across the human frontotemporal cortex during natural speech show individual cells encode fine-grained linguistic features: some respond to parts of speech, others to sentence constituents or dependency depth, and some track combinatorial feature sets. The neuron population also encodes sentence context, with language encoding being left-lateralized and varying by region. Large language models’ contextual embeddings better predict neural activity than syntax/semantics alone, revealing distributed cellular building blocks and a micro-to-macro map of how language is represented in the brain.