Research map / Large language models / Language, safety and interpretability
Neural Language Models are not Born Equal to Fit Brain Data, but Training Helps
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the model instead of using untrained weights | GloVe, LSTM, GPT-2, and BERT · The Little Prince fMRI, SRM25 | R-test: GloVe (1.5% SE=0.02%); LSTM (3.1% SE=0.02%); GPT-2 (4.5% SE=0.02%); BERT (4.4% SE=0.02%) | Helped |
| Increase depth from 1 to 4 layers (ablation) | Trained GPT-2 · The Little Prince fMRI, SRM25 | brain score: 2% SE=0.006% | Helped |
| Increase depth from 1 to 4 layers (ablation) | Trained BERT · The Little Prince fMRI, SRM25 | brain score: 1% SE=0.006% | Helped |
| Increase depth from 1 to 4 layers (ablation) | Trained LSTM · The Little Prince fMRI, SRM25 | brain score: 0.4% SE=0.005% | Helped |
| Increase depth from 1 to 4 layers (ablation) | Untrained LSTM · The Little Prince fMRI, SRM25 | brain score: -0.02% SE=0.002% | No clear effect |
| Increase depth from 1 to 4 layers (ablation) | Untrained GPT-2 · The Little Prince fMRI, SRM25 | brain score: -0.6% SE=0.004% | Hurt |
| Increase depth from 1 to 4 layers (ablation) | Untrained BERT · The Little Prince fMRI, SRM25 | brain score: -0.02% SE=0.003% | No clear effect |
| Use an untrained LSTM rather than an untrained GPT-2 | Untrained language models · The Little Prince fMRI, SRM25 | brain score difference: 1.6% SE=0.02% | Helped |
| Use an untrained LSTM rather than an untrained BERT | Untrained language models · The Little Prince fMRI, SRM25 | brain score difference: 0.7% SE=0.004% | Helped |
| Use a context-sensitive LSTM rather than static GloVe | Trained LSTM and GloVe · The Little Prince fMRI, language network | brain score | Helped |
| Use bidirectional BERT rather than incremental GPT-2 | Trained transformer models · The Little Prince fMRI, language network | brain score | Helped |
| Increase LSTM training data from Wikipedia to Gutenberg datasets | LSTM · The Little Prince fMRI, SRM25 | brain score: 0.0828; 0.0973; 0.0957; 0.0934; 0.0920; 0.0962 | Mixed |
Closest papers by meaning
- Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.91)
- Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.91)
- fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.90)
- Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.89)
- BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.88)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.88)
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.