Research / Inside the model
Language-Neutral or Language-Specific? Disentangling Multilingual LLM Subspaces with Naturalistic fMRI
Uses multilingual encoders as factorized probes of naturalistic language processing, showing that a cross-lingually shared component predicts semantic brain regions better than the language-specific residual in all six confirmatory tests.
- Published
- 23 March 2026
- Author
- Ali Uyar
Abstract
Large language model features increasingly explain naturalistic fMRI responses, but most model–brain alignment studies remain effectively monolingual. Multilingual encoders, by contrast, contain both cross-lingually shared structure and language-specific residual structure, which makes them useful probes for a more mechanistic question: during comprehension, is the brain better predicted by what several languages agree on or by the information that remains specific to the heard language? We tested this question in the Le Petit Prince multilingual naturalistic fMRI corpus by aligning English, French, and Chinese sentence spans, extracting sentence representations from XLM-R-base and the encoder of NLLB-200-distilled-600M, decomposing each target-language representation into a leave-target-out shared component and an orthogonalized target-language-specific residual, and evaluating those components with cross-validated ROI encoding models. In the semantic ROI family, the shared component outperformed the specific residual in all six model-by-language confirmatory tests, and every H1 effect survived Holm correction. The stronger prediction that the semantic-family shared advantage would exceed the auditory-family shared advantage remained mixed, because auditory and superior temporal ROIs also showed robust SHARED–SPECIFIC effects. Correct-content controls remained directionally healthy, including broad SHARED > MISMATCHED_SHARED behavior, whereas text-geometry-to-brain coupling was weak and non-significant in the present checkpoint. Robustness analyses supported the main semantic-family effect under several perturbations, but a voxelwise-within-ROI target produced negative French rows in both model families. The most defensible conclusion is therefore narrower than a clean semantic-versus-auditory dissociation: multilingual LLMs can serve as factorized probes of naturalistic language processing, but the brain-side distribution of shared and specific sensitivity is regionally heterogeneous.
Why it matters
Most model–brain alignment studies are effectively monolingual. A multilingual encoder holds both structure shared across languages and structure specific to one, and separating the two lets each be tested against the brain on its own.
Cite
@misc{uyar2026language,
title = {Language-Neutral or Language-Specific? Disentangling Multilingual LLM Subspaces with Naturalistic fMRI},
author = {Uyar, Ali},
year = {2026},
doi = {10.5281/zenodo.19185476},
publisher = {Zenodo},
url = {https://doi.org/10.5281/zenodo.19185476}
}