Systematic and dynamic task coding in the frontal cortex

Behavioral flexibility depends on the capacity to form and utilize task representations. Prior work has shown that baseline activity in the prefrontal cortex (PFC) can differ between two tasks. Comparing more than two tasks makes it possible to ask whether different neuronal populations share a common form of task encoding, that is, a conserved representational geometry across tasks. We trained two rhesus monkeys to perform 6 cognitive tasks and recorded spiking activity simultaneously from hundreds of neurons in the frontal eye field (FEF) or the dorsolateral PFC (dlPFC), one area per session. During the fixation period, before any task-specific stimulus appeared, the population activity state already reflected the upcoming task, and task identity could be decoded on single trials. Strikingly, the geometry of the 6 task states was conserved across populations recorded in different sessions, across the two areas, and across the two monkeys. Global signals that differ between tasks, such as expected reward, task difficulty, and attention, measured by reward amount, hit rate, and reaction time, did not fully account for this shared geometry. These findings suggest that the frontal cortex systematically adjusts its population state to encode and prepare for the upcoming task.