LLM 'Semantic Neurons' Map to Human Brain Regions with 94% Accuracy

A team led by Dongxin Guo, Jikun Wu, and Siu Ming Yiu at HKU bridges mechanistic interpretability and computational neurolinguistics by applying sparse autoencoders (SAEs) to GPT-2 XL and Llama-3.1-8B. Decomposing each layer into 16K–32K features, they created a human-validated taxonomy (κ≥0.74) showing that semantic features alone explain 94% of the peak encoding performance of LLM activations against human brain responses (r=0.285, p<0.001, d=1.31). This quantifies why intermediate LLM layers best predict brain activity—semantic content, not syntax or position, drives the alignment.

Crucially, the SAE-discovered features recapitulate known cortical semantic organization: five predefined semantic subcategories (e.g., social, spatial, emotional) map onto distinct brain regions with high convergence (Spearman ρ=0.72, hypergeometric p=0.007). The features also predict human reading times beyond lexical baselines (ΔlogLik=38.4, p<0.001), with exploratory evidence that the brain encodes unexpected semantic content. Results hold across English, Chinese, and French, suggesting universal neural alignment. Accepted at CoNLL 2026, this work offers a mechanistic explanation for brain-LLM similarity and a new tool for studying language in the cortex.