Every day, we make decisions between multidimensional options. Learning each option’s value by simultaneously attending to all its dimensions is challenging. However, humans navigate this complexity by selectively focusing on relevant information, making real-world, multidimensional reinforcement learning (RL) tractable. Neuroimaging has identified correlates of attention and value learning in the lateral prefrontal (LPFC) and orbitofrontal (OFC) cortex, respectively, yet these methods fall short in elucidating the physiological mechanisms by which these regions coordinate during multidimensional RL. We leveraged human intracranial electrophysiology, which offers the necessary spatiotemporal resolution to test whether LPFC directs attention to relevant information to guide the OFC’s efficient value representation via local neuronal value encoding and interregional low-frequency communication. We recorded local field potentials from LPFC and OFC of 21 neurosurgical participants performing a multidimensional RL task. A selective attention RL model explained performance better than a uniform attention RL model, confirming participants’ deployed selective attention during the task. Higher model-based selective attention correlated with faster reaction times and improved performance. Choice’s expected value (EV) was encoded in OFC and LPFC high gamma activity, with key distinctions in how each region processed this information. LPFC's value signals were modulated by attention, highlighting its role in directing attention to relevant information. The timing of peak EV encoding differed significantly between the regions, suggesting they represent EV with distinct temporal profiles. LPFC to OFC information flow, assessed by directed cross-regional theta-band connectivity, was modulated by attention. Under greater selective attention, pre-choice LPFC-OFC theta connectivity increased with increasing EV, reflecting LPFC’s role in guiding attention. By leveraging behavioral modeling and iEEG, results reveal how attention shapes neuronal activity during value learning by modulating both the magnitude and timing of local value encoding and interregional communication. Findings provide novel insight into the neurocomputational mechanisms underlying adaptive behavior in complex, real-world environments.