Do Large Language Models Think Like the Brain? Sentence-Level Evidences from Layer-Wise Embeddings and fMRI

TL;DR

This study asks whether the brain-like processing observed in large language models is driven by scaling or architectural alignment. Using a sentence-level, layer-wise comparison of embeddings from 14 LLMs with fMRI data collected as participants listened to a naturalistic story, the authors construct encoding models to map model representations to brain activity. They find that intermediate-layer representations align best with language-related brain regions, and that instruction tuning enhances both semantic comprehension and neural alignment, with left-hemisphere regions showing stronger Language processing correlations and right-hemisphere regions contributing to higher-level cognition. Overall, the work argues for a nuanced brain–LLM correspondence shaped by semantic depth, training objectives, and hemispheric specialization, suggesting paths toward more cognitively plausible language models.

Abstract

Understanding whether large language models (LLMs) and the human brain converge on similar computational principles remains a fundamental and important question in cognitive neuroscience and AI. Do the brain-like patterns observed in LLMs emerge simply from scaling, or do they reflect deeper alignment with the architecture of human language processing? This study focuses on the sentence-level neural mechanisms of language models, systematically investigating how layer-wise representations in LLMs align with the dynamic neural responses during human sentence comprehension. By comparing hierarchical embeddings from 14 publicly available LLMs with fMRI data collected from participants, who were exposed to a naturalistic narrative story, we constructed sentence-level neural prediction models to identify the model layers most significantly correlated with brain region activations. Results show that improvements in model performance drive the evolution of representational architectures toward brain-like hierarchies, particularly achieving stronger functional and anatomical correspondence at higher semantic abstraction levels. These findings advance our understanding of the computational parallels between LLMs and the human brain, highlighting the potential of LLMs as models for human language processing.

Figures (8)

• Figure 1: A brief presentation of the experimental design. We expose both LLMs and real humans to a narrative story ("The Little Prince") and aim to compare the correlation between LLMs' and human brains' language processing.
• Figure 2: Multi-stage pipeline to analyze the alignment between LLM representations and neural responses during naturalistic language comprehension. The methodology includes auditory stimulus presentation, layer-wise embedding extraction from LLMs, voxel-wise regression modeling, and region-of-interest (ROI)-based brain-model alignment analysis. Panel (a) outlines the neuroimaging data acquisition and preprocessing steps, while panel (b) describes the brain-LLM alignment analysis.
• Figure 3: Performance comparison of 14 LLMs
• Figure 4: Correlation between model predictions and brain activity across layers. Shaded areas are 95% confidence intervals.
• Figure 5: Average correlation of LLMs across 12 ROIs.