Compositional computations in the primate frontal cortex

A hallmark of intelligence is the ability to apply familiar operations to new inputs without relearning. How neural circuits carry a learned computation into states never visited during training remains unclear. Here we combined neural recordings, circuit modeling and human experiments to investigate compositional generalization. Monkeys learned to combine a start location with a displacement specified by the direction and duration of an optic-flow stimulus, then encountered novel combinations of these familiar inputs. Frontal population activity separated the computation into spatial and temporal components: start location and flow direction established the neural state, from which activity evolved along shared timing dynamics. Novel combinations elicited accurate responses when trajectories maintained this organization, whereas departures predicted systematic behavioral biases. Circuit models revealed a three-way alignment problem: input representations, recurrent dynamics and readout operations must jointly implement the required computation. Mismatches produced systematic biases that increased with distance from trained conditions. Humans exhibited similar biases when familiar sensory cues and spatial transformations were linked through an unfamiliar mapping. Together, these findings show how the coordination of reusable neural resources supports new computations, and how its disruption produces predictable failures of generalization.