Larger language models better align with neural representations of natural language.

Larger transformer-based language models more accurately predict neural activity, with performance peaking in earlier layers as model size increases, across brain regions.

• Why it matters: Understanding how artificial models mirror human language processing can improve both AI and neuroscience, but the relationship between model size and neural alignment remains unclear.
• What they did: Researchers analyzed several families of LLMs, controlling for architecture and training data, and used electrocorticography to measure neural responses in epilepsy patients listening to natural speech.
• The result: Findings show larger models better predict neural signals, with an organized hierarchy across brain regions, enabling improved models of natural language comprehension and brain function.