💬 NLP

Brain-LLM Alignment Tracks Training Data, Not Typology

This study demonstrates that cross-linguistic brain-LLM alignment is primarily driven by a model's training-language dominance rather than an inherent English advantage, with residual variations reflecting genuine typological distances that disproportionately affect syntactic brain regions.

Original authors: Dongxin Guo, Jikun Wu, Siu Ming Yiu

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dongxin Guo, Jikun Wu, Siu Ming Yiu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Question: Is English Special?

Imagine you have a super-smart robot (a Large Language Model, or LLM) that has read millions of books. Scientists have discovered that when this robot "thinks" about a sentence, its internal math looks surprisingly similar to how a human brain lights up when hearing that same sentence.

For a long time, this "brain-robot match" was only tested with English. Because the robot was trained mostly on English, and the human brains were listening to English, the match was perfect.

This led to a big question: Is the human brain somehow "wired" to match English specifically? Or is the match just a coincidence because the robot learned English?

The Experiment: Swapping the Recipe

To answer this, the researchers set up a massive experiment using a story called The Little Prince. They recorded the brains of 112 people listening to this story in three different languages: English, Chinese, and French.

They then tested seven different AI models against these brain scans.

The "English Robot": A model trained mostly on English (like LLaMA-2).
The "Chinese Robot": A model trained mostly on Chinese (Baichuan2).
The "Multilingual Robot": A model trained on many languages equally (like XLM-R).

The Magic Trick:
The researchers compared the "English Robot" and the "Chinese Robot." These two robots are twins in every way—they have the same size, the same brain structure, and the same number of layers. The only difference is what they ate (their training data). One ate mostly English; the other ate mostly Chinese.

The Big Discovery: It's About What You Ate, Not Your DNA

The results flipped the script entirely:

The English Robot matched best with English brains and worst with Chinese brains.
The Chinese Robot did the exact opposite: it matched best with Chinese brains and worst with English brains.

The Analogy:
Think of the human brain like a lock.

The "English Robot" is a key cut specifically for an English lock. It fits perfectly.
The "Chinese Robot" is a key cut specifically for a Chinese lock.
The researchers proved that the "perfect fit" isn't because the English lock is special. It's simply because the key was cut to match that specific lock. If you cut a key for a different lock, it fits that one instead.

Conclusion: The "English advantage" in previous studies wasn't because English is the "universal language of the brain." It was just because the AI models were trained on English data. The brain doesn't care about the language; it cares about the pattern the AI learned.

Other Surprising Findings

1. The "Translation Glitch" (Tokenization)

When the AI reads Chinese, it often has to break one Chinese character into multiple tiny pieces (tokens) to understand it, whereas English words are often one piece.

The Analogy: Imagine reading a book where every English word is one brick, but every Chinese word is a pile of 2.4 bricks glued together.
The Result: Because of this "brick pile," the AI has to look deeper into its own layers to understand the meaning of a Chinese word. The researchers found that about 60% of the reason the AI "thinks" differently for Chinese is just this mechanical glitch in how it breaks words apart, not a deep difference in how the brain processes meaning.

2. Grammar vs. Meaning (The Brain's Neighborhoods)

The human brain has different neighborhoods for different jobs.

The "Meaning Neighborhood" (PTL): This area handles vocabulary and concepts.
The "Grammar Neighborhood" (IFG): This area handles sentence structure and rules.

The researchers found that the "Meaning Neighborhood" matched the AI almost perfectly, no matter what language was spoken. The AI and the brain agree on what words mean.

However, the "Grammar Neighborhood" was very picky. It only matched the AI well if the AI had learned the specific grammar rules of that language.

The Analogy: If you and I both know what "apple" means, we are on the same page. But if I learned to drive on the left side of the road and you learned on the right, we will argue about how to turn the steering wheel. The brain's "grammar" part is sensitive to these rules, while the "meaning" part is universal.

3. The "Universal" Multilingual Robot

The researchers also tested a robot trained on many languages at once. This robot was the most flexible. It didn't have a strong preference for English or Chinese; it matched the brains of all three languages almost equally well. This suggests that if you teach an AI to be truly multilingual, it can understand human brains in any language.

Summary

Myth Busted: The brain isn't "English-centric." The AI just happened to be trained on English.
The Real Driver: The AI's performance depends on what language it was trained on. If you train it on Chinese, it understands Chinese brains best.
The Nuance: While the meaning of words is universal across languages, the grammar rules are specific. The brain's grammar centers are sensitive to these differences, but the meaning centers are not.
The Technical Glitch: Some of the differences we see between languages in AI are just because the AI struggles to count Chinese characters correctly (tokenization), not because the brain works differently.

In short: The AI is a mirror. If you show it English, it reflects English. If you show it Chinese, it reflects Chinese. The human brain is the same in both cases; only the reflection changes.

Technical Summary: Brain–LLM Alignment Tracks Training Data, Not Typology

Problem Statement
While the human brain's language network is neuroanatomically universal across 45 languages, the extent to which Large Language Models (LLMs) align with this network across typologically diverse languages remains unclear. Prior research has established strong brain–LLM alignment almost exclusively using English data, creating a potential blind spot regarding the universality of these representations. A critical confound in existing literature is the inability to distinguish whether alignment patterns are driven by an inherent "English advantage," the typological distance between languages, or the dominance of specific languages in a model's training data. Furthermore, it is unknown whether cross-linguistic variation in alignment is concentrated in syntactic or semantic processing regions, a question with implications for Construction Grammar (CxG) and theories of language universality.

Methodology
The study utilizes fMRI data from 112 participants listening to Le Petit Prince in three languages: English ( $n=49$ ), Mandarin Chinese ( $n=35$ ), and French ( $n=28$ ). The authors employ a controlled experimental design comparing seven LLMs:

English-dominant: GPT-2 Medium and LLaMA-2-7B.
Chinese-dominant: Baichuan2-7B.
Multilingual: mBERT, XLM-R Large, BLOOM-7B, and Qwen2.5-7B.

Crucially, the study pairs LLaMA-2-7B and Baichuan2-7B, which share identical architecture and parameter counts (7B, 32 layers) but differ primarily in training-language composition (~90% English vs. ~55% Chinese).

The analysis pipeline involves:

Encoding Models: Voxelwise ridge regression to predict neural activity from LLM hidden states, normalized against inter-subject noise ceilings.
Typological Metrics: Quantitative typological distance calculated using Grambank (195 binary grammatical features), supplemented by WALS and lang2vec.
Tokenization Analysis: Measurement of "tokenization fertility" (subword tokens per orthographic word) to assess its impact on layer dynamics.
Statistical Framework: Linear mixed-effects models (LMM) controlling for training data proportion, with cluster bootstrap resampling at the language level to address limited degrees of freedom (3 languages).

Key Contributions and Results

Training Language Dominance Drives Alignment Gradients:
The central finding is that the "English advantage" in brain–LLM alignment is an artifact of training data composition, not an inherent property of English or the brain.
- Reversal of Gradient: The Chinese-dominant Baichuan2-7B reverses the alignment gradient observed in LLaMA-2-7B. Baichuan2 aligns best with Chinese brains ( $\tilde{r} = .85$ ) and worst with English ( $\tilde{r} = .59$ ), whereas LLaMA-2 shows the opposite pattern.
- Multilingual Uniformity: Multilingual models (e.g., XLM-R) exhibit high uniformity across languages ( $\tilde{r} \approx .81-.83$ ), suggesting that multilingual training creates representations that transcend specific language dominance.
Formal Typological Distance Covaries with Alignment:
Beyond training dominance, formal typological distance independently predicts alignment degradation.
- In mixed-effects models controlling for training proportion, Grambank distance is associated with a decrease in alignment performance ( $\beta = -0.41$ ).
- The authors treat this as a well-supported descriptive pattern requiring confirmation across $\ge10$ languages due to limited degrees of freedom.
Tokenization Fertility Explains Layer Shifts:
The study identifies a rightward shift in the optimal encoding layer for Chinese (peaking at layer 18 vs. layer 15 for English/French).
- Tokenization fertility analysis reveals Chinese requires significantly more subword tokens per word (2.4 tokens/word for XLM-R) compared to English (1.4 tokens/word).
- Controlling for fertility accounts for approximately 60% of this layer shift, indicating that what appears to be a cross-linguistic processing difference is substantially a tokenization artifact.
Region-Specific Neural Gradients (Syntax vs. Semantics):
The typological gradient is not uniform across the language network.
- The Inferior Frontal Gyrus (IFG), associated with syntactic processing, shows a typological gradient 2.3 times steeper than the Posterior Temporal Lobe (PTL), associated with lexico-semantic processing.
- This suggests that cross-linguistic divergence is concentrated in structural/syntactic processing, while semantic representations remain more universal.

Significance and Claims
The paper claims to resolve the "English advantage" debate by demonstrating that alignment patterns are driven by training-language dominance. The apparent superiority of English in prior studies reflects the predominance of English-dominant models and English fMRI data, not a special neurocognitive relationship between English and the brain.

The results support a nuanced view of universality: biological and artificial language processors converge on shared semantic representations (universal across languages and models) but diverge in structural/syntactic representations, where typological distance and training distribution matter. The findings also highlight that tokenization artifacts can masquerade as cross-linguistic processing differences, urging the field to report tokenization fertility alongside encoding results.

The authors explicitly note that while the IFG/PTL gradient is consistent with Construction Grammar (which posits variation in form-function pairings), it is also derivable from generic syntax–semantics dissociation theories or cognitive control accounts. Therefore, the paper claims the results are consistent with CxG but do not uniquely prove it without construction-level analysis. Similarly, the typological distance effect is presented as a descriptive pattern requiring larger-scale validation rather than a definitive causal proof.