Brain Language Processing Mirrors GPT's Hierarchical Structu

A recent study published in Nature Communications suggests that the human brain's language processing mechanism shares a structural similarity with large language models (LLMs) like GPT. The research indicates that the brain may not primarily rely on traditional grammatical parsing but rather on a predictive, layer-by-layer inference process, akin to how LLMs function.

The findings challenge the long-held belief that human language comprehension is based on rigorous rules, grammar, and structural analysis. Instead, the study posits that the brain's "temporal imprint" aligns with the hierarchical structure observed in GPT models.

Experimental Design and Findings

Researchers conducted an experiment where subjects listened to a 30-minute story while their brain activity was recorded using millisecond-level electrocorticography (ECoG). Simultaneously, the same story text was fed into LLMs, specifically GPT-2 and Llama-2, to extract their internal processing representations at each layer.

The core of the research involved aligning the 48-layer structure of GPT with the human brain's time series. Nine epilepsy patients undergoing pre-surgical monitoring, with high-density ECoG electrodes implanted on their cerebral cortex, participated. This allowed for the recording of real electrical brain activity with millisecond precision. High-gamma EEG signals were collected around the time each word appeared, covering key language pathway areas such as the mSTG (auditory processing), aSTG, IFG (language integration), and TP (higher semantic areas).

The study revealed a temporal correspondence between GPT's hierarchical structure and brain activity. When the LLMs processed a word, their internal representations from the first to the last layer were extracted. These representations were then dimensionally reduced and used in a linear model to predict the electrical activity in the human brain.

The results showed that shallow layers of the LLM corresponded to earlier brain activity, while deeper layers correlated with later activity, creating a "time-depth" correspondence. This pattern was particularly strong in higher-order semantic areas like the TP, aSTG, and IFG, with correlation coefficients (r) of .93, .92, and .85, respectively. In contrast, the mSTG, closer to the auditory cortex, showed almost no hierarchical structure (r≈0), indicating that this area primarily processed sound rather than complex semantics.

Divergence from Traditional Linguistic Models

The research also compared the predictive power of GPT's internal representations against traditional symbolic linguistic models, which are based on phonemes, morphemes, syntax, and semantics. These traditional models, while able to predict some brain activity, did not exhibit the clear "shallow to deep" or "early to late" sequential distribution seen with GPT. They lacked the hierarchical and temporal progression, suggesting a less dynamic representation of language processing at the millisecond level.

This indicates that while symbolic models describe "what language is," GPT's processing more closely reflects "how the brain processes language." The study suggests that the brain's language mechanism is not a simple stacking of symbolic rules but a continuous, deep predictive process. The brain, like Transformer models, performs nonlinear transformations, continuously updating, compressing, and integrating meaning over time.

Implications for Language Understanding

The findings suggest that language may not be a system of fixed rules but a dynamic predictive mechanism. The brain, similar to how GPT is trained, constantly calculates "what might happen next" at every millisecond. This perspective implies that the perceived understanding of large language models might stem from their unexpected alignment with the brain's natural processing rhythm, rather than their adherence to human-defined linguistic rules.

This convergence of computational laws between AI models and the human brain suggests that the essence of language might be continuous, dynamic prediction. This mechanism is used by the brain to understand the world and integrate information, and by models to generate language and simulate intelligence. The study posits that established frameworks in linguistics and cognitive science may require re-evaluation in light of these findings.