Cross-lingual robustness of LLM-brain alignment and its computational roots

Problem Statement
While Large Language Models (LLMs) have demonstrated robust ability to predict neural activity during language comprehension, several critical gaps remain in understanding the nature of this alignment. First, it is unclear whether LLM-brain alignment extends beyond the classical, left-lateralized perisylvian language network to encompass subcortical structures and distributed cortical functional networks. Second, the spatial overlap of these alignment patterns across typologically distinct languages (e.g., Mandarin, English, French) has not been systematically characterized. Third, the computational roots of this alignment are debated: specifically, whether transformer depth truly mirrors the hierarchical organization of the brain, and whether predictive processing (surprisal) or information compression (intrinsic dimensionality) drives the observed neural predictivity. Finally, recent evidence suggests that contextual embeddings may not consistently outperform static embeddings, challenging the assumption that context-sensitivity is the primary driver of neural alignment.

Methodology
The study employed a multilingual, whole-brain encoding framework using fMRI data from 112 participants listening to naturalistic stories ("The Little Prince") in three languages: Mandarin, English, and French.

• Data: fMRI data were acquired during passive listening to audiobooks. Preprocessing involved standard pipelines (fMRIPrep), confound regression, and mapping to a 400-parcel cortical parcellation (Schaefer) and 16 subcortical regions (Melbourne Subcortex Atlas).
• Models: Seven models were utilized: one multilingual BERT (mBERT), three monolingual BERTs (one per language), and three static FastText models (one per language).
• Encoding: Ridge regression was used to map model embeddings (both static and contextual) to fMRI time series. Contextual embeddings were derived from all 12 transformer layers of mBERT, with context windows defined by entire sentences.
• Analysis:
  • Spatial Alignment: Significant regions of interest (ROIs) were identified via one-sample t-tests with FDR correction. Overlap across languages was visualized on cortical surfaces.
  • Layer-wise Analysis: Pairwise permutation tests assessed differences in brain scores across transformer layers. Preferred layers for each ROI were identified.
  • Computational Metrics: To test mechanistic hypotheses, the authors computed layer-wise surprisal (predictive uncertainty) and intrinsic dimensionality (ID) (representational compression) for mBERT. These metrics were compared against the observed layer-wise brain score profiles.
  • Model Comparison: Contextual (mBERT) vs. static (FastText) performance was compared using permutation tests and linear mixed-effects models.

Key Results

1. Distributed and Subcortical Alignment: Transformer-based models predicted neural activity across a wide distributed landscape, including canonical language areas (IFG, STG, MTG), the Default Mode Network (DMN), limbic networks, and subcortical structures (caudate, putamen, amygdala). This confirms that LLM alignment extends beyond the cortical language network.
2. Cross-Linguistic Stability: Spatial alignment patterns showed substantial overlap across Mandarin, English, and French. Shared significant regions included the posterior STG, MTG, IFG, IPL, PCC, mPFC, ACC, and the amygdala. While some language-specific variations existed, a robust, language-agnostic representational space was identified.
3. Stability Across Layers: Contrary to the hypothesis of a strict layer-to-cortical hierarchy, the study found no systematic progression of alignment from early sensory regions to higher-order areas across transformer layers. Spatial alignment patterns remained largely stable across model depth. In some regions, early or intermediate layers performed comparably to or better than deeper layers.
4. Static vs. Contextual Embeddings: Contrary to previous literature suggesting contextual embeddings outperform static ones, this study found that static FastText embeddings often yielded higher encoding performance than contextual mBERT embeddings across all three languages.
5. Computational Metrics Mismatch: Neither surprisal nor intrinsic dimensionality (ID) mirrored the observed layer-wise brain score profiles. While brain scores exhibited a U-shaped profile across layers (peaking in intermediate layers), surprisal decreased monotonically, and ID peaked at the 10th layer. This indicates that predictive uncertainty and representational compression are not the primary drivers of the observed alignment.
6. Functional Hierarchy Sensitivity: While layer-wise progression was absent, the magnitude of brain scores across functional networks reflected the brain's large-scale organizational gradients. Alignment was strongest in transmodal networks (limbic, ventral attention, DMN) and weaker in primary sensory/motor networks, consistent with the principal cortical gradient.

Significance and Claims
The paper claims that brain-LLM alignment is spatially robust and cross-linguistically stable, generalizing across typologically distinct languages and extending into subcortical regions. However, the authors modestly conclude that this alignment does not necessarily reflect a shared hierarchical computational organization between transformers and the human brain.

Specifically, the study challenges the view that transformer depth mirrors cortical hierarchy or that alignment is driven by predictive processing (surprisal) or information compression (ID). Instead, the authors propose that neural predictivity primarily arises from distributed lexical-semantic correspondences that generalize across languages. The alignment may reflect shared statistical structures of natural language rather than shared computational objectives (such as minimizing prediction error or compressing information). The findings suggest that while LLMs capture aspects of the brain's functional hierarchy (e.g., transmodal vs. sensory processing), they do so through mechanisms that are distinct from the specific layer-wise computational dynamics previously hypothesized.