[Paper Review] Left-right asymmetry in predicting brain activity from LLMs' representations emerges with their formal linguistic competence

Laurent Bonnasse-Gahot, Christophe Pallier|arXiv (Cornell University)|Feb 13, 2026

Neurobiology of Language and Bilingualism0 citations

TL;DR

The study shows that left-right brain predictivity asymmetry from LLM representations emerges as models acquire formal linguistic competence, aligning with grammatical ability rather than world knowledge or arithmetic tasks, and generalizes across models and languages.

ABSTRACT

When humans and large language models (LLMs) process the same text, activations in the LLMs correlate with brain activity measured, e.g., with functional magnetic resonance imaging (fMRI). Moreover, it has been shown that, as the training of an LLM progresses, the performance in predicting brain activity from its internal activations improves more in the left hemisphere than in the right one. The aim of the present work is to understand which kind of competence acquired by the LLMs underlies the emergence of this left-right asymmetry. Using the OLMo-2 7B language model at various training checkpoints and fMRI data from English participants, we compare the evolution of the left-right asymmetry in brain scores alongside performance on several benchmarks. We observe that the asymmetry co-emerges with the formal linguistic abilities of the LLM. These abilities are demonstrated in two ways: by the model's capacity to assign a higher probability to an acceptable sentence than to a grammatically unacceptable one within a minimal contrasting pair, or its ability to produce well-formed text. On the opposite, the left-right asymmetry does not correlate with the performance on arithmetic or Dyck language tasks; nor with text-based tasks involving world knowledge and reasoning. We generalize these results to another family of LLMs (Pythia) and another language, namely French. Our observations indicate that the left-right asymmetry in brain predictivity matches the progress in formal linguistic competence (knowledge of linguistic patterns).

Motivation & Objective

• Investigate which competencies in LLMs drive the emergence of left-right hemispheric asymmetry in brain predictivity.
• Examine the alignment between training progress, linguistic benchmarks, and brain score asymmetry.
• Test generalizability of findings across model families (OLMo-2, Pythia) and languages (English, French).
• Differentiate the influence of formal linguistic competence from non-linguistic or functional language tasks on brain asymmetry.

Proposed method

• Compute voxelwise brain scores by regressing fMRI signals on LLM activations using ridge-regularized linear models across layers.
• Define left/right brain scores by averaging correlations from voxels in each hemisphere and focusing on the 25% most reliable voxels.
• Evaluate LLM competence with minimal-pair benchmarks (BLiMP, Zorro) and non-linguistic tasks (Arithmetic, Dyck language).
• Assess linguistic acceptability of generated text as a measure of formal competence.
• Replicate analyses in Pythia models and in French using language-specific benchmarks (fr-grammar, French Hellaswag).
• Fit sigmoids to training-trajectory curves to compare transition points (x0) and slopes (beta) between brain asymmetry and benchmark performance.

Experimental results

Research questions

• RQ1 Does the emergence of left-right brain score asymmetry track the acquisition of formal linguistic competence in LLMs?
• RQ2 Is the asymmetry aligned with linguistic benchmarks (formal competence) rather than non-linguistic or functional tasks?
• RQ3 Do findings generalize across different model families and languages?
• RQ4 What is the relative timing of brain asymmetry development versus functional language capabilities during training?

Key findings

• Left-right brain score asymmetry shows a phase transition during training that aligns with the rise of formal linguistic competence.
• BLiMP and Zorro benchmarks track the same transition in brain asymmetry, while Arithmetic and Dyck benchmarks do not.
• Linguistic acceptability of generated text mirrors the left-right asymmetry trajectory, whereas ARC and Hellaswag do not.
• Results generalize to Pythia models and to French, though formal competence develops more slowly in French due to English-focused training.
• Cerebellar right-left asymmetry mirrors the cortical left-right pattern, also aligning with formal linguistic competence.
• The left-right asymmetry correlates with syntactic rather than solely lexical-semantic content, suggesting syntactic processing contributes to brain-model alignment.