Task demands dynamically structure feature selection, routing, and integration in the human brain

Flexible visual categorization requires selecting task-relevant information from complex stimuli and transforming it through structured computations. Yet how task demands flexibly control these computations in the human brain remains unresolved. Dynamic faces provide a powerful test case, conveying stable identity through 3D shape and transient emotional expression through 4D movement. Using generative 4D face modelling, participant-specific MEG (N = 24), and information-theoretic analyses, we show that the brain implements a structured processing architecture. Task relevance first gates feature representations in occipital cortex, selectively sustaining relevant features while suppressing irrelevant ones. These features are then routed along pathway-specific channels—3D identity features to the ventral pathway and 4D emotion features to the lateral pathway. Finally, these inputs converge in temporal cortex, where identity and emotion features are synergistically integrated for learned identities. These results show how task demands dynamically structure feature selection, gating, routing, and integration, thereby constraining the computations that support flexible categorization.