New AI model generates brain activity for unseen cognitive tasks

A team of researchers introduced a flow matching model with in-context priors that generates whole-cortex fMRI brain dynamics for cognitive tasks never seen during training. Unlike prior generative models restricted to categorical conditioning, this diffusion transformer processes per-timestep language inputs and optional spatial priors, enabling compositional and zero-shot generalization. The model can produce realistic neural time series for hundreds of held-out task conditions, accurately recovering region-specific recruitment patterns from text descriptions alone. When spatial priors are added, they anchor generation in task regions where language degrades, preserving the compositional structure needed for counterfactual task specification.

This work represents the first generative model of whole-cortex fMRI dynamics for unseen cognitive tasks. By allowing researchers to simulate brain activity for novel experiments before running costly empirical studies, it advances counterfactual neuroscience and data-driven experimental design. The authors evaluated predictive performance across the training manifold and found that the model maintains high fidelity even for out-of-distribution conditions. Code and pretrained models are publicly available, opening the door for broader adoption in neuroscience and AI research.