AI's Brain-Like Language Skills Come from Just 1-3% of Its Code
This tiny fraction could explain how AI understands language—and lead to better brain treatments.
Scientists found that only 1–3% of an AI language model’s original embedding dimensionality was enough to recover full linear encoding performance for predicting neural responses during listening. Across three naturalistic language-listening datasets spanning intracranial electrophysiology and fMRI, and six model conditions, these low-dimensional brain-predictive subspaces generalized across held-out participants and, to a lesser extent, across datasets.
The directions most useful for predicting neural activity were distinct from those capturing the most overall variance in the LLM representations. Together, the results show that LLM-brain alignment is supported by a small, structured subset of model representations and provide a framework for better understanding exactly why these representations align.
- Only 1-3% of an AI's internal code is needed to predict human brain activity during listening.
- These key directions work across different people and brain-scanning methods, suggesting a shared brain code.
- The most useful directions for predicting brains are not the most obvious ones in the AI, so we need to look deeper.
Why It Matters
This could lead to better AI that understands language like humans and new ways to study brain disorders.