brain-llm-alignment-training-data

name	brain-llm-alignment-training-data
description	Brain-LLM alignment is driven by training-language dominance, not an inherent property of English. Tests with fMRI from 112 participants across English, Chinese, French and 7 LLMs (English-dominant, Chinese-dominant, multilingual). Baichuan2-7B reverses alignment gradient entirely; typological distance independently affects alignment degradation in syntax regions (IFG). Accepted at CoNLL 2026. Activation: brain-LLM alignment, cross-linguistic brain encoding, training data dominance, multilingual fMRI, typological alignment.
arxiv_id	2605.23032
published	2026-05-21
authors	Dongxin Guo, Jikun Wu, Siu Ming Yiu
tags	["brain-llm-alignment","cross-linguistic","fmri","neurolinguistics","training-data-dominance","computational-neuroscience"]

Brain-LLM Alignment Tracks Training Data, Not Typology

This paper shows that the apparent "English advantage" in brain-LLM alignment is an artifact of training data composition. Using fMRI from 112 participants across three languages (English, Chinese, French) and 7 LLMs, it demonstrates that training-language dominance, not English per se, drives alignment patterns.

Source: arXiv: 2605.23032 | Accepted at CoNLL 2026

Core Methodology

Key Innovation

Brain-LLM alignment is well-established in English, but the brain's language network is neuroanatomically universal. This paper asks: does alignment generalize cross-linguistically, and what governs the variation? It provides the first systematic cross-linguistic test of brain-LLM alignment.

Technical Framework

1. fMRI Dataset: Le Petit Prince corpus with 112 participants across English, Chinese, and French
2. LLM Suite: 7 models spanning English-dominant (LLaMA-2-7B, GPT-2 XL), Chinese-dominant (Baichuan2-7B), and multilingual (mT5, BLOOM, XLM-R) architectures
3. Encoding Model: Ridge regression encoding models predicting fMRI responses from LLM layer activations
4. Training-Language Dominance Analysis: Compare alignment gradients between architecture-matched English-dominant (LLaMA-2-7B) and Chinese-dominant (Baichuan2-7B) models
5. Typological Distance Analysis: Quantify how formal typological distance between languages independently affects alignment degradation
6. Brain Region Analysis: Decompose alignment by brain regions — syntax-associated IFG vs. lexico-semantic PTL
7. Tokenization Analysis: Measure how tokenization fertility (tokens per word) affects cross-linguistic optimal encoding layer shifts

Key Results

• Training-language dominance drives alignment: Baichuan2-7B (Chinese-dominant, architecture-matched to LLaMA-2-7B) reverses the alignment gradient entirely — aligns best with Chinese brains and worst with English
• Typological distance independently covaries with alignment degradation across all models
• Syntax regions (IFG) show steeper typological gradients than lexico-semantic regions (PTL)
• Tokenization fertility accounts for ~60% of the cross-linguistic shift in optimal encoding layer
• The "English advantage" is an artifact of training data composition, not an inherent property of the English language

Applications

• Cross-linguistic neuroscience: Study how language processing in the brain depends on model training vs. inherent language structure
• LLM evaluation for brain alignment: Evaluate which models are best for predicting brain responses across different languages
• Neurolinguistic theory: Understand the interplay between training data, typology, and neural language processing
• Multilingual model design: Inform multilingual model development by understanding how training data composition affects brain-relevant representations

Related Skills

• sparse-autoencoder-brain-llm-topography
• brain-llm-key-neurons-grammar
• fcn-llm-brain-network-understanding
• computational-linguistics-brain-perspective