Summary: Researchers found that the human brain processes spoken language through a temporal sequence of computations that closely mirrors the layered transformations inside large language models. Electrocorticography (ECoG) recordings taken while participants listened to a 30-minute podcast show early neural responses matching early AI model layers and later brain activity corresponding to deeper model layers, particularly in regions like Broca’s area.
These results challenge traditional, rule-based views of language processing and support a dynamic, context-driven model in which meaning emerges gradually through layered integration. The research team also released a comprehensive dataset aligning neural signals with linguistic features to support further work in language neuroscience.
Key Facts
- Layered alignment: Early neural responses correspond to early model layers, while later responses align with deeper layers.
- Context over symbolic rules: Contextual embeddings derived from AI models predicted neural activity better than classical linguistic units.
- New research resource: The authors released a large dataset pairing ECoG recordings with linguistic annotations to accelerate investigation into how the brain processes language.
Source: Hebrew University of Jerusalem
In a study published in Nature Communications, a team led by Dr. Ariel Goldstein (Hebrew University) in collaboration with Dr. Mariano Schain (Google Research), Prof. Uri Hasson, and Eric Ham (Princeton University) demonstrated a close relationship between the temporal unfolding of language processing in the human brain and the hierarchical layers of modern large language models (LLMs).
Using intracranial electrocorticography data recorded from participants who listened to a 30-minute spoken narrative, the researchers compared time-resolved neural signals to contextual embeddings extracted from models such as GPT-2 XL and Llama 2. They used linear encoding models to predict neural responses from model activations and found a robust correspondence between model depth and the timing of brain activity across language regions.
What the study discovered
When we comprehend speech, each word triggers a cascade of computations across the brain. This study shows those computations unfold over time in a manner that mirrors the progressive transformations inside LLMs. Early layers of AI models—sensitive to basic lexical and phonetic patterns—aligned with early, fast neural responses. Deeper model layers—handling richer contextual integration, semantics, and discourse-level information—aligned with later neural activity, especially in high-level language areas such as Broca’s area.
The pattern was consistent across participants and across model families: deeper model layers peaked later in the brain’s response profile. As Dr. Goldstein remarked, “What surprised us most was how closely the brain’s temporal unfolding of meaning matches the sequence of transformations inside large language models. Even though these systems are built very differently, both seem to converge on a similar step-by-step buildup toward understanding.”
Why this matters
For decades, dominant theories of language processing emphasized symbolic rules and fixed hierarchical grammars. These findings support an alternative picture: comprehension as a gradual, context-sensitive construction of meaning. The study shows that AI-derived contextual embeddings capture features of neural dynamics that classical linguistic units—like isolated phonemes or morphemes—do not predict as well in real time. This suggests the brain relies heavily on context and probabilistic integration when forming linguistic representations.
Beyond theoretical implications, the work positions modern LLMs as useful computational tools for probing human cognition. While artificial systems and biological brains differ in implementation, their convergent computational patterns can offer new hypotheses about the mechanisms underlying natural language understanding.
Open dataset as a community benchmark
To promote transparency and further discovery, the authors released a curated dataset that aligns ECoG recordings with time-stamped linguistic features and model embeddings. This resource enables researchers to test competing theories of language processing, develop better brain–model comparisons, and refine computational models that more closely reflect human neural dynamics during naturalistic language comprehension.
Key Questions Answered:
A: Spoken language is transformed by the brain through a sequence of computations whose temporal progression aligns with progressively deeper layers of large language models.
A: It challenges strictly rule-based theories and supports a model in which meaning is built up dynamically from context through successive integration steps, similar to how LLMs compute contextual embeddings.
A: A publicly available dataset pairing electrocorticography recordings with detailed linguistic annotations and model embeddings to serve as a benchmark for future studies.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by the editorial staff.
About this language and AI research news
Author: Yarden Mills
Source: Hebrew University of Jerusalem
Contact: Yarden Mills, Hebrew University of Jerusalem
Image: Image credit: Neuroscience News
Original Research: Open access. “Temporal structure of natural language processing in the human brain corresponds to layered hierarchy of large language models” by Uri Hasson et al., published in Nature Communications.
Abstract
Temporal structure of natural language processing in the human brain corresponds to layered hierarchy of large language models
Large language models provide a numerical, layered representation of words and context that differs from traditional symbolic linguistic frameworks. In this study, we show that the hierarchical layers of these models align with the temporal dynamics of human language comprehension. Using electrocorticography data from participants listening to a 30-minute narrative, we find that deeper model layers correspond to later brain activity, particularly in Broca’s area and other language-related regions. Contextual embeddings from GPT-2 XL and Llama-2, used within linear encoding models, reliably predict time-resolved neural responses, revealing a strong correlation between model depth and the brain’s temporal receptive fields. Comparing these predictions to symbolic approaches highlights the advantages of deep, context-sensitive representations for capturing neural dynamics. We provide the aligned neural and linguistic dataset as a public benchmark to facilitate testing of competing theories of language processing.