New study finds LLMs' grammar skills drive brain-activity predictions stronger in left hemisphere

Researchers studying the relationship between large language models and the human brain have uncovered a striking pattern: as LLMs develop formal grammatical competence, their internal representations increasingly align with brain activity in the left hemisphere—the region typically responsible for language processing in humans.

Brain scans and model activations converge#

When people read text, functional magnetic resonance imaging can detect patterns of brain activity. Scientists have long observed that these neural signatures correlate with activations inside trained LLMs processing the same material. A new study from arXiv pushes this observation further, tracing how this relationship evolves during a model's training.

The researchers tracked the OLMo-2 7B language model at different stages of training, comparing its internal representations against fMRI data from English-speaking volunteers. They found that predictive power—the ability to infer brain activity from model activations—improved asymmetrically: gains were significantly steeper in the left hemisphere than the right.

Grammar ability, not general reasoning, drives the pattern#

The critical finding is what accounts for this hemispheric bias. The team tested whether various competencies correlated with the emerging asymmetry:

• Formal linguistic ability showed a tight match. The models' capacity to assign higher probability to grammatically correct sentences (versus intentionally malformed minimal pairs) and to generate well-formed text tracked closely with left-hemisphere asymmetry growth.
• General reasoning and world knowledge, by contrast, did not. Performance on arithmetic tasks, Dyck language problems, or reasoning-heavy benchmarks failed to predict the neural asymmetry.

Generalization across models and languages#

The pattern held when the team replicated their analysis using the Pythia family of models and when testing with fMRI data from French and Chinese speakers, suggesting the phenomenon reflects a robust principle rather than an artifact of a single model or language.

The findings suggest that LLMs acquire representational structures that increasingly mirror how the human brain—particularly its language-specialized hemisphere—organizes linguistic information, specifically through learning formal grammatical rules.