Neural Correlates of Language Models Are Specific to Human Language (2510.03156v1)

Abstract: Previous work has shown correlations between the hidden states of LLMs and fMRI brain responses, on language tasks. These correlations have been taken as evidence of the representational similarity of these models and brain states. This study tests whether these previous results are robust to several possible concerns. Specifically this study shows: (i) that the previous results are still found after dimensionality reduction, and thus are not attributable to the curse of dimensionality; (ii) that previous results are confirmed when using new measures of similarity; (iii) that correlations between brain representations and those from models are specific to models trained on human language; and (iv) that the results are dependent on the presence of positional encoding in the models. These results confirm and strengthen the results of previous research and contribute to the debate on the biological plausibility and interpretability of state-of-the-art LLMs.

• The paper demonstrates that only transformer models trained on human language exhibit significant neural alignment with fMRI responses.
• It employs ridge regression and PCA to reveal that positional encoding markedly enhances brain score correlations.
• Brain-model alignment peaks in deeper layers, with causal models outperforming bidirectional ones, underlining the temporal dynamics of language processing.

Neural Correlates of LLMs Are Specific to Human Language

Introduction

This paper systematically investigates the representational alignment between transformer-based LMs and human brain activity during language comprehension, as measured by fMRI. While prior work has reported correlations between LM hidden states and neural responses, this paper addresses critical open questions: Are these correlations specific to models trained on human language? Do they persist under dimensionality reduction, and are they robust to alternative similarity metrics? What is the role of architectural components, particularly positional encodings, in driving these effects? The analysis leverages a diverse set of 19 transformer models, including both bidirectional (masked language modeling) and causal (next-word prediction) architectures, and introduces rigorous controls using models trained on non-linguistic data.

Methodology

The core experimental design involves mapping model activations to fMRI responses elicited by sentence reading. The fMRI dataset comprises 243 stimuli per subject, with BOLD signals extracted from language-selective ROIs. Model activations are obtained layer-wise, with and without positional encodings, and further decomposed into head-wise and attention-weight representations. The analysis pipeline includes:

• Ridge regression for voxel-wise prediction of fMRI activity from model activations, with cross-validation and normalization by a noise ceiling.
• Dimensionality reduction via PCA to 50 principal components, preserving 75–80% of variance, to test for curse-of-dimensionality artifacts.
• Representational similarity metrics: Pearson correlation (brain scores), centered kernel alignment (CKA), and Gromov-Wasserstein (GW) distance, providing both correlational and geometric/topological perspectives.
• Control models: Transformers trained on protein sequences, matched in architecture and objective but lacking linguistic input, to test for language specificity.

Results

Dimensionality and Positional Encoding Effects

The paper finds that brain-model correlations are robust to aggressive dimensionality reduction. PCA compression does not diminish alignment, indicating that the observed effects are not artifacts of high-dimensional spaces.

Figure 1: Average brain scores (brain-model representation correlations) per condition (PCA and original dimensions; with and without positional encodings).

Crucially, the presence of positional encodings is a major determinant of brain alignment. Removing positional information leads to a substantial drop in brain scores (up to 0.4 in $r$), with the effect being more pronounced in causal models. Bidirectional models, due to their access to both past and future context, can partially compensate for the loss of explicit position, but still show a significant performance gap.

Layer-wise and Head-wise Alignment

Layer-wise analysis reveals that brain alignment generally increases with network depth, peaking in the final layers for both bidirectional and causal models. Causal models (e.g., OPT-2.7B) achieve the highest correlations ($r=0.65$), outperforming bidirectional models (maximum $r=0.46$).

Figure 2: All models' layer-wise brain scores (left: bidirectional, right: causal). Blue: [+pos], orange: [-pos], dashed: PCA data, purple: overall mean.

Head-wise analysis of scaled dot-product attention shows that certain heads contribute disproportionately to brain alignment, with the effect again contingent on positional encoding. In causal models, the absence of positional information results in uniformly low head-wise scores, while bidirectional models display more distributed patterns.

Figure 3: Head-wise (scaled dot-product attention) brain scores for bidirectional (top) and causal (bottom) models. Brighter colors indicate better correlations.

Attention weights alone are only marginally predictive of brain activity, and their contribution is less systematic than that of the full attention mechanism.

Figure 4: Head-wise (attention weights only) brain scores for bidirectional (top) and causal (bottom) models.

Geometric and Topological Similarity

CKA analysis demonstrates that representational similarity between model layers and brain activity is highly sensitive to positional encoding. Without positional information, CKA scores are low and flat; with positional encoding, CKA increases monotonically with depth, peaking at the final layers (0.61–0.87).

Figure 5: Participant-model CKA results for -pos and +pos conditions. Each line: participant; bottom row: mean.

GW distance analysis provides a complementary geometric perspective. GW distances are minimized (indicating greater similarity) in deeper layers when positional encodings are present, and increase sharply when positional information is ablated.

Figure 6: Participant-model Gromov-Wasserstein distance results for -pos and +pos conditions.

Language Specificity

A critical finding is that only models trained on human language exhibit significant brain alignment. Protein-trained transformers, despite architectural and objective similarity, yield brain scores $\leq 0.03$, compared to $0.23$ or higher for the weakest text-trained models. This result rules out generic architectural or task-driven explanations and demonstrates that exposure to linguistic data is necessary for capturing neural correlates of language processing.

Regression Analysis

An OLS regression combining the main experimental metrics (brain score, CKA, GW) yields an adjusted $R^2=0.71$, indicating that these measures capture largely independent but complementary aspects of brain-model alignment.

Figure 7: Results for the OLS regression of the experimental results. Adjusted $R^2=0.71$.

Implications and Future Directions

The findings have several implications for both cognitive neuroscience and the development of biologically plausible LLMs:

• Biological Plausibility: The strong dependence on positional encoding and the specificity to language-trained models suggest that transformer LMs capture aspects of the neural code for language, but only under certain architectural and training constraints.
• Model Selection for Cognitive Alignment: Causal (autoregressive) models, which process input in a left-to-right fashion, more closely mirror the temporal dynamics of human language processing and achieve higher brain alignment.
• Metric Robustness: The convergence of results across correlation, CKA, and GW distance strengthens the claim that the observed effects are not metric artifacts.
• Limits of Generalization: The lack of alignment in protein-trained models underscores the necessity of linguistic exposure for neural correspondence, challenging claims that generic sequence modeling suffices for brain-like representations.

Future work should expand the diversity of both models and neural datasets, incorporate additional modalities (e.g., auditory language, multimodal tasks), and further dissect the contributions of architectural components (e.g., attention heads, feedforward sublayers) to neural alignment. The integration of more temporally resolved neural data (e.g., MEG, ECoG) may also clarify the mapping between model dynamics and rapid syntactic-semantic operations in the brain.

Conclusion

This paper provides a comprehensive and methodologically rigorous assessment of the neural correlates of transformer LLMs. The results demonstrate that representational alignment with human brain activity is specific to models trained on human language, robust to dimensionality reduction, and critically dependent on positional encoding. Causal models, which more closely mimic the temporal structure of human language processing, achieve the highest alignment. These findings inform both the design of cognitively plausible LLMs and the interpretation of brain-model similarity metrics in computational neuroscience.