| name | cross-lingual-llm-brain-alignment |
| description | Multi-lingual whole-brain encoding framework examining brain-LLM alignment across three typologically distinct languages (Mandarin, English, French). Shows that transformer-based models predict activity in widely distributed cortical functional networks (limbic, ventral attention, default mode, subcortical) across languages, revealing computational roots of cross-linguistic neural alignment with LLM representations. Activation: cross-lingual brain alignment, multilingual fMRI encoding, LLM-brain alignment, computational neurolinguistics, cross-linguistic neural representations. |
| arxiv_id | 2605.21049 |
| published | 2026-05-20 |
| authors | Ni Yang, Rui He, Philipp Homan, Iris Sommer, Davide Staub, Wolfram Hinzen |
| tags | ["brain-llm-alignment","cross-lingual","neurolinguistics","fmri-encoding","multilingual"] |
Cross-lingual robustness of LLM-brain alignment and its computational roots
Examines brain-LLM alignment across Mandarin, English, and French using a whole-brain encoding framework, revealing distributed cortical and subcortical overlap with shared computational principles.
Source: arXiv: 2605.21049
Core Methodology
Key Innovation
First systematic investigation of brain-LLM alignment across three typologically distinct languages (Mandarin Chinese, English, French) using whole-brain fMRI encoding, extending beyond cortical regions to subcortical structures.
Technical Framework
- Multilingual Naturalistic Stimuli: Participants listened to naturalistic stories in Mandarin, English, and French during fMRI scanning
- Transformer-Based Encoding Models: Extract representations from multiple layers of LLMs (e.g., GPT-2, Llama) trained on each language
- Whole-Brain Voxelwise Modeling: Predict BOLD activity for each voxel across the entire brain using ridge regression
- Cross-Linguistic Alignment Analysis: Compare encoding performance across languages to identify shared vs. language-specific neural patterns
- Subcortical Investigation: Extend beyond cortex to examine limbic, ventral attention, default mode network, and subcortical regions
- Computational Root Analysis: Decompose which linguistic features (syntax, semantics, phonology) drive alignment patterns
Key Results
- Cross-linguistic consistency: LLM-brain alignment generalizes across typologically distinct languages
- Distributed cortical networks: Alignment spans limbic, ventral attention, default mode, and subcortical regions (not just classical language cortex)
- Shared computational principles: Common features across languages drive neural alignment
- Subcortical contributions: Subcortical regions show significant alignment not previously reported
- Layer-specific patterns: Transformer depth correlates with different functional networks across languages
Applications
- Multilingual neuroscience: Study how the human brain processes different languages at the neural level
- Brain-LLM alignment validation: Test whether alignment generalizes beyond single-language settings
- Cross-linguistic NLP and clinical applications for bilingual/multilingual populations
- Subcortical language processing: Investigate subcortical contributions to language comprehension
Activation Keywords
- cross-lingual brain alignment
- multilingual fMRI encoding
- LLM-brain alignment
- computational neurolinguistics
- whole-brain encoding model
- subcortical language processing
- cross-linguistic neural representations
- transformer language models brain
Related Skills
- sparse-autoencoder-brain-llm-topography
- brain-llm-key-neurons-grammar
- brain-llm-alignment-training-data
- mllm-brain-alignment-task-probing
- lpact-brain-lm-alignment-evaluation
