How Language and Speech Models Mirror Brain Activity Through Semantic Abstraction

Recent research from arXiv suggests that the ability of large language and speech models to predict human brain activity stems from how these systems represent meaning, not from their core prediction mechanisms.

Why Middle Layers Are Special#

Studies have long shown that intermediate hidden layers from language and speech models can effectively predict neural responses measured via fMRI and ECoG—yet the precise reason remained unclear. A new study finds that it's not the output layers (which handle final predictions) but rather the middle layers that excel at this task.

The key insight: models extract higher-order linguistic features in their middle layers, creating richer semantic representations. These intermediate representations align surprisingly well with how the brain processes language.

The Role of Abstraction and Complexity#

The researchers point to intrinsic dimension—a measure of feature complexity—as the critical factor. They found a marked peak in intrinsic dimension at middle layers, corresponding to where models construct more sophisticated semantic features. Crucially:

• A layer's intrinsic dimension strongly predicts its ability to explain observed brain signals
• This relationship between complexity and brain predictivity emerges during model pre-training itself
• When models are fine-tuned to better predict the brain, their representations gain both higher intrinsic dimension and richer semantic content

Meaning Over Mechanics#

The findings suggest that semantic richness and brain predictivity mirror each other, challenging the assumption that language models' ability to predict neurons comes from their next-word prediction task. Rather, language modeling appears to be sufficiently complex to require rich meaning abstraction—though it may not be the only way to achieve such abstraction.

This work illuminates how artificial and biological intelligence converge around deep understanding of meaning.