Research map / Architectures and efficiency / RNNs and normalization
fMRI predictors based on language models of increasing complexity recover brain left lateralization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Encodes Le Petit Prince fMRI with 28 pretrained LLMs (124M–14.2B) and finds brain-score scaling with log parameters, with a growing left-over-right hemispheric asymmetry as models get larger.
Problem
Symmetric bilateral activation patterns from LLM-based fMRI encoding models conflict with the well-known left lateralization of language; it is unclear whether model scale affects hemispheric asymmetry in brain predictivity.
Approach
Systematically varies the size of pretrained autoregressive LLMs across 8 families and fits ridge-regression encoding models to group-averaged fMRI, showing brain correlation scales with log parameters and that the left-minus-right correlation difference also follows a scaling law.
Key results
- Brain correlation increases linearly with log(number of parameters), r=0.95, p=7.3e-15 on the 25% most reliable voxels.
- Left-right difference in brain correlation follows a scaling law with model size (r=0.89, p=1.8e-10).
- Beyond 3B parameters, the number-of-parameters correlation is no longer significant (r=0.30), while perplexity (r=-0.86) and Hellaswag (r=0.83) remain significant among the ten largest models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase LLM parameter count | Pretrained LLM encoding models · Le Petit Prince English, 25% most reliable voxels | brain correlation: r=0.95, p=7.3e-15 | Helped |
| increase LLM parameter count | Pretrained LLM encoding models · Le Petit Prince English, whole brain volume | brain correlation: r=0.95, p=3.2e-14 | Helped |
| use larger, higher-performing LLMs | LLM encoding models · Le Petit Prince English, 25% most reliable voxels | left-right brain-correlation difference: r=0.89, p=1.8e^{-10} | Helped |
| use models with lower Wikitext-2 perplexity | LLM encoding models · Le Petit Prince English, 25% most reliable voxels | brain correlation: r=-0.91, p=1.1e-11 | Helped |
| use models with higher HellaSwag performance | LLM encoding models · Le Petit Prince English, 25% most reliable voxels | brain correlation: r=0.95, p=4.2e-15 | Helped |
| increase parameter count among the 10 largest models (>3B) | LLM encoding models · Le Petit Prince English, 10 largest models | brain correlation: r=0.30, p=0.41 | No clear effect |
| use models with lower Wikitext-2 perplexity among the 10 largest models | LLM encoding models · Le Petit Prince English, 10 largest models | brain correlation: r=-0.86, p=1.5e-3 | Helped |
| use models with higher HellaSwag performance among the 10 largest models | LLM encoding models · Le Petit Prince English, 10 largest models | brain correlation: r=0.83, p=3.0e-3 | Helped |
| increase model size | LLM encoding models · Le Petit Prince English, language-network ROIs | left-right asymmetry | Mixed |
| increase model size | LLM encoding models · Le Petit Prince English, BA44 ROI | left-right asymmetry: 0.15; 4.6e-01 | No clear effect |
| increase model size | LLM encoding models · Le Petit Prince English, TP ROI | left-right asymmetry: 0.36; 6.2e-02 | No clear effect |
| train the model rather than use random initialization (ablation) | GPT-2 and Qwen1.5 · Le Petit Prince English, 25% most reliable voxels | left-right brain-correlation difference | Helped |
Surprising findings
- Brain-score gains from model size are not monotone for the largest models: the parameter-count relationship breaks down above 3B parameters.
- Left-right asymmetry emerges with model size and training, which the authors note has not been reported in the literature; untrained models show no growing asymmetry.
Limitations
- Main analyses use group-averaged data from 49 English speakers, so individual-level effects are only checked on 5 participants.
- Model-size effects are confounded with architecture, training data and family differences across the 28 models.
- Right-hemisphere and signal-to-noise differences are only partly controlled via inter-subject correlation normalization.
Experimental setup
Datasets: Le Petit Prince fMRI dataset Wikitext-2 Hellaswag
Benchmarks: Le Petit Prince fMRI dataset
Metrics: Pearson correlation between predicted and actual BOLD inter-subject correlation perplexity (Wikitext-2) Hellaswag accuracy
Baselines: random vectors random embeddings GloVe
Builds on: Schrimpf et al. brain score Antonello et al. scaling of brain encoding Glover HRF
Closest papers by meaning
- Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.88)
- Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.88)
- TopoLM: brain-like spatio-functional organization in a topographic language model (ICLR 2025 · similarity 0.88)
- Scaling and context steer LLMs along the same computational path as the human brain (NeurIPS 2025 · similarity 0.88)
- Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.87)
- BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.