What Are Large Language Models Mapping to in the Brain? A Case Against Over-Reliance on Brain Scores

TL;DR

This paper challenges the interpretation of brain scores as direct evidence of brain-like computing in large language models. It introduces a rigorous deconstruction framework using simple linguistic features and the Orthogonal Autocorrelated Sequences Model (OASM) to separate true linguistic processing from autocorrelation-driven variance across Pereira, Fedorenko, and Blank datasets. The findings show that untrained LLMs are largely explained by sentence length and position, and that trained LLM predictivity on Pereira is mostly captured by non-contextual features such as SP, SL, and WORD, with only modest gains from sense and syntax representations; on Fedorenko, word position accounts for most variance, while Blank yields chance-level predictions. Collectively, the work cautions against over-interpreting brain scores and emphasizes dissecting what neural signals LLMs actually map to, with implications for how brain-likeness should be assessed in NLP-to-brain studies.

Abstract

Given the remarkable capabilities of large language models (LLMs), there has been a growing interest in evaluating their similarity to the human brain. One approach towards quantifying this similarity is by measuring how well a model predicts neural signals, also called "brain score". Internal representations from LLMs achieve state-of-the-art brain scores, leading to speculation that they share computational principles with human language processing. This inference is only valid if the subset of neural activity predicted by LLMs reflects core elements of language processing. Here, we question this assumption by analyzing three neural datasets used in an impactful study on LLM-to-brain mappings, with a particular focus on an fMRI dataset where participants read short passages. We first find that when using shuffled train-test splits, as done in previous studies with these datasets, a trivial feature that encodes temporal autocorrelation not only outperforms LLMs but also accounts for the majority of neural variance that LLMs explain. We therefore use contiguous splits moving forward. Second, we explain the surprisingly high brain scores of untrained LLMs by showing they do not account for additional neural variance beyond two simple features: sentence length and sentence position. This undermines evidence used to claim that the transformer architecture biases computations to be more brain-like. Third, we find that brain scores of trained LLMs on this dataset can largely be explained by sentence length, position, and pronoun-dereferenced static word embeddings; a small, additional amount is explained by sense-specific embeddings and contextual representations of sentence structure. We conclude that over-reliance on brain scores can lead to over-interpretations of similarity between LLMs and brains, and emphasize the importance of deconstructing what LLMs are mapping to in neural signals.

Figures (10)

• Figure 1: Within each panel, EXP1 results are on the left and EXP2 results are on the right (a)$R^2$ values across layers for GPT2-XL on shuffled train-test splits (gray) and contiguous (unshuffled) splits (blue). (b) Each dot shows the mean $R^2$ value across voxels within a participant, with bars indicating mean $R^2$ across participants.
• Figure 2: For all panels, EXP1 results are on the left and EXP2 results are on the right. (a)$R^2$ for different combinations of features. Each dot represents $R^2$ values averaged across voxels in a single participant, with bars showing mean across participants. (b) 2D histogram of $R^2$ values ($100$ bins) for the best model without GPT2-XLU (SP+SL), and the best model with GPT2-XLU (GPT2-XLU+SP+SL). The dotted lines show $y=x$, $y=0$, and $x=0$. Values below $y=0$ or left of $x=0$ were clipped when averaging, but are shown here to visualize the full distribution. (c) Same as (a), but after voxel-wise correction for SP+SL and SP+SL+GPT2-XLU; lines connect data-points from the same participant. (d) Glass brain plots, for a representative participant (different for each EXP), showing $R^2$ values of SP+SL* (left) and SP+SL+GPT2-XLU* (right). Neural encoding performance across brain areas is highly similar for both models.
• Figure 3: For all panels, EXP1 results are on the left and EXP2 results are on the right. (a) For each model, we display the best sub-model which must include the last feature space. Dots represent participants and bars are mean across participants. Gray dashed line is the performance of GPT2-XL alone. (b)$2$D histogram comparing the best sub-model which includes GPT2-XL to the best sub-model that does not. (c) Same as (a) but after voxel-wise correction for SP+SL+WORD and SP+SL+WORD+GPT2-XL. (d) Glass brain plots showing $R^2$ values of SP+SL+WORD* (left) and SP+SL+WORD+GPT2-XL* (right) for a representative participant (different for each EXP).
• Figure 4: (a) Across-layer $R^{2}$, averaged across electrodes in Fedorenko, for GPT2-XL with and without shuffled splits. (b, c, d) Dots represent participants and bars are mean across participants.
• Figure 5: For all panels, EXP1 is on the left, and EXP2 is on the right. Layer $0$ is the input, static layer. (a)) Across layer performances in Pereira for GPT2-XLU for each functional network when using the sum-pooling method. b) Same as (a) but for GPT2-XL, also using the sum-pooling method. c) Same as (a) but when using the last token method. Dotted grey line shows performance of best layer of GPT2-XLU in language network when sum-pooling. d) Same as (b) but when using the last token method. Dotted grey line shows performance of best layer of GPT2-XL in language network when sum-pooling.