The Part of an AI Model That Lines up With a Listening Brain Is Tiny

Put a person in a room with a narrative podcast playing and a grid of electrodes resting on the surface of their cortex, and you can watch the language areas of the brain respond word by word. Feed the same transcript to GPT-2 and you get a second record of the same half hour: a long list of numbers for every word. Fitting one to the other has become a standard move in neuroscience, and it works well enough that large language models are now routinely called computational models of human language. What nobody had put a number on is how much of the model is involved.

A preprint posted to arXiv on Oct. 29, 2026, puts one there. Zach Paris and Emily Finn, of the Department of Psychological and Brain Sciences at Dartmouth College, report that a linear prediction built from 1–3 % of a model's dimensions tracks recorded brain activity as well as a linear prediction built from all of them. The work remains unpublished in a peer-reviewed journal, and its findings have yet to be independently reviewed

Dimensions here are just the length of the number list the model writes for each word: 768 for GPT-2, several thousand for the largest models in the study. So 1–3 % is a few dozen directions through that space, and in the podcast recordings the average was 20. Matching the full model means landing within a hair of the best score the authors' own sweep reached, on recording sites the full model already predicted well. That is a statement about one kind of prediction, not about the model as a whole.

How that subspace is found shapes what it can mean. Paris and Finn use a regression capped at a set number of directions, reduced-rank regression, and they fit it against the brain recordings. The compact space is therefore not something sitting visibly inside the model; it is what is left when the recordings get a vote. Their sweep covers six conditions: four models trained to predict the next word (GPT-2, GPT-2 XL, GPT-J 6B, Llama 3 8B), one trained to fill in missing words (RoBERTa Large), and a GPT-2 whose weights were never trained, as a control.

The recordings come from three sets collected by other labs: nine people with electrodes inside the skull listening to a narrative podcast; three people with depth electrodes watching the film Cars 2; and eight people listening to spoken stories in an fMRI scanner, which tracks blood flow. About twenty people in all, which is ordinary for work that needs electrodes inside a head and is still a small base for a general claim.

Two tests then ask whether the subspace is more than a quirk of whoever was in the room. A space found in one group still predicts the recordings of people whose data was never used to build it, and in two of the three datasets it does so at about the same low number of dimensions. The exception is the three-person film dataset, where the transferred version needed more. Between datasets the transfer is only partial, and the size of the shortfall tracks how far apart the stimuli were, which they tie to what people were listening to rather than to their brains.

One further comparison gives the result its edge. If those few directions were simply where the model varies most, then principal component analysis, which finds exactly those axes, should do the same job on the recordings. It does not: it needs hundreds of dimensions where the fitted subspace needs tens. Whatever the electrodes and the scanner are tracking, it is not what the model itself leans on.

There is a peer-reviewed reason to hold the interpretation loosely. Nima Hadidi, Jonathan C. Kao and colleagues published a large robustness check of this literature in Nature Communications on April 27, 2026, across three widely used neural datasets. Their conclusions: shuffled training and test splits have produced influential results that do not hold up, and how activations are pulled out of a model can favor some model classes over others. The third is the one that lands here. Simple confounds, in particular where a word sits in a sentence and how fast words arrive, perform competitively with trained models, and for untrained models they fully account for how well the predictions work.

Paris and Finn cite that paper and sidestep its central trap: their held-out data are contiguous blocks of words and whole unseen stories, not shuffled samples. But one of their six conditions is the GPT-2 that was never trained, the result is reported as holding across all six, and sentence position is among the features they find most readable from the compact subspace itself. The preprint does not report a word-position or word-rate baseline against that subspace. So the counting result holds up on its own terms; the reading of the subspace as the place where a model and a brain share something about language is contested in the reviewed literature.

What holds the number down to a few dozen is left open. The authors suggest the recordings may be too coarse to carry more, or the models may be missing part of what the recordings hold. The code is public and all three datasets were already open, so the next group to take the subspace apart does not need anyone's permission to start.