Billion-parameter AI models trained on brain scans fail to predict how well you think, while a simple statistical measure succeeds — and does so without any GPU or pretraining. That's the finding of a new study that uncovers a fundamental blind spot in today's most advanced brain models.
The research
Giovanni Marraffini and colleagues at the French National Institute for Research in Digital Science and Technology (Inria) tested three state-of-the-art brain foundation models (BFMs) — large AI systems pretrained on fMRI data from thousands of people. They evaluated how well each model could predict individual cognitive performance, using standard cognitive tests from large public datasets like the Human Connectome Project.
The result was startling: every BFM predicted cognition worse than a simple linear regression using the functional connectivity (FC) matrix — a matrix of ~80,000 numbers describing pairwise correlations between brain regions. Worse, larger models performed more poorly: BrainLM's 650-million-parameter version underperformed its 111-million-parameter counterpart.
The team traced the problem to what they call a variance allocation problem. BFMs are trained to reconstruct the fMRI signal as accurately as possible, which means they focus on the largest, most dominant variance components — but these are largely noise for cognition. Crucially, the models destroy the third-order co-skewness — a statistical measure that captures asymmetric, non-Gaussian relationships between brain regions — which turns out to be far more predictive of cognition than ordinary correlations.
To recover what BFMs lose, the researchers designed a simple linear pipeline: project the fMRI signal into the subspace that best preserves co-skewness, then compute FC in that subspace. This method — requiring no GPU and no pretraining — outperformed raw FC and every BFM across all datasets and brain parcellations tested. It even matched the ceiling performance of BrainLM's forward pass after targeted fine-tuning.
Why it matters
This finding has profound implications for how we understand and measure intelligence. The brain's most cognitively relevant signals may not be the strongest or most obvious — they are subtle, higher-order patterns that conventional AI models ignore. For anyone interested in their own cognitive abilities, it suggests that simple, well-chosen measures can be more informative than complex black-box algorithms.
Importantly, the study reveals that the bottleneck in current brain AI is not model architecture or size but the pretraining objective itself. By shifting focus to third-order statistics, future models could become both simpler and more accurate.
What you can do
While you can't extract co-skewness from your brain at home, you can appreciate that cognition arises from complex, non-linear interactions. Engaging in activities that challenge your brain — like puzzles, learning new skills, or cognitive training — may strengthen these subtle connections. And when you hear claims about AI predicting intelligence, remember that sometimes less is more: the right simple measure can outperform a billion-parameter model.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
