Prefrontal cortex (PFC) plays a key role in flexible, context-dependent behavior, but the exact mechanisms underlying this function remain unclear. This study investigates PFC activity in macaque monkeys performing a task requiring flexible selection and integration of visual stimuli. The results show that the complex responses of individual PFC neurons can be understood within a population-level dynamical framework. A trained recurrent neural network (RNN) successfully models these dynamics, suggesting a novel mechanism for context-dependent selection and integration of task-relevant inputs. This mechanism indicates that selection and integration are aspects of a single dynamical process within the same PFC circuits, providing a general framework for understanding context-dependent computations. The study demonstrates that PFC neurons do not gate irrelevant sensory inputs before integration, but instead select relevant inputs late in the process. The population-level dynamics show that relevant inputs are integrated along the choice axis, while momentary evidence from motion and color inputs forms arcs away from this axis. These findings challenge existing models of early sensory selection and suggest that context-dependent computations rely on a recurrent network mechanism involving a line attractor and a selection vector. This mechanism allows the same input to influence behavior differently depending on context, as the selection vector is context-dependent and orthogonal to irrelevant inputs. The study also shows that PFC activity can be modeled using a recurrent neural network, which reproduces key physiological observations and provides insights into the mechanism of context-dependent selection and integration. The model's dynamics, including a line attractor and selection vector, explain how relevant inputs are integrated while irrelevant inputs are filtered out. These findings highlight the importance of population-level dynamics in PFC for flexible, context-dependent behavior and suggest that PFC functions through the coordinated activity of large neuronal populations. The results challenge the notion that sensory modulation is necessary for input selection and emphasize the role of recurrent dynamics in context-dependent computations.Prefrontal cortex (PFC) plays a key role in flexible, context-dependent behavior, but the exact mechanisms underlying this function remain unclear. This study investigates PFC activity in macaque monkeys performing a task requiring flexible selection and integration of visual stimuli. The results show that the complex responses of individual PFC neurons can be understood within a population-level dynamical framework. A trained recurrent neural network (RNN) successfully models these dynamics, suggesting a novel mechanism for context-dependent selection and integration of task-relevant inputs. This mechanism indicates that selection and integration are aspects of a single dynamical process within the same PFC circuits, providing a general framework for understanding context-dependent computations. The study demonstrates that PFC neurons do not gate irrelevant sensory inputs before integration, but instead select relevant inputs late in the process. The population-level dynamics show that relevant inputs are integrated along the choice axis, while momentary evidence from motion and color inputs forms arcs away from this axis. These findings challenge existing models of early sensory selection and suggest that context-dependent computations rely on a recurrent network mechanism involving a line attractor and a selection vector. This mechanism allows the same input to influence behavior differently depending on context, as the selection vector is context-dependent and orthogonal to irrelevant inputs. The study also shows that PFC activity can be modeled using a recurrent neural network, which reproduces key physiological observations and provides insights into the mechanism of context-dependent selection and integration. The model's dynamics, including a line attractor and selection vector, explain how relevant inputs are integrated while irrelevant inputs are filtered out. These findings highlight the importance of population-level dynamics in PFC for flexible, context-dependent behavior and suggest that PFC functions through the coordinated activity of large neuronal populations. The results challenge the notion that sensory modulation is necessary for input selection and emphasize the role of recurrent dynamics in context-dependent computations.