DYNAMIC INTEGRATION OF DISSOCIABLE VALUE CODES FOR REWARD AND INFORMATION IN THE HUMAN PREFRONTAL CORTEX

Balancing the exploitation of known rewards with exploration for new information is a core challenge in decision neuroscience. While the prefrontal cortex is critical for this process, how its subregions dynamically cooperate to mediate this balance remains unclear. Here, we recorded intracranial stereo-electroencephalography (SEEG) from 17 human participants performing a gambling task that orthogonalized expected reward and information gain. Using reinforcement learning-based computational modeling, we confirmed that participants assigned intrinsic value to information. Time-frequency analyses identified functionally dissociable value representations: while ventromedial prefrontal cortex (vmPFC) activity encoded the integrated value of the current option, anterior cingulate cortex (ACC) activity tracked reward and information value in a distinct manner.Network dynamics analysis revealed that undirected Theta synchronization (PLV) supported reward integration, whereas directed information flow (PSI) showed that the vmPFC initiates exploration by driving the ACC. Furthermore, Phase-Amplitude Coupling (PAC) identified specific gating mechanisms: a bottom-up striatal-cortical coupling encoded absolute motivational gain, while a top-down ACC Beta-Gamma coupling facilitated deliberation during high-conflict exploration. Crucially, state-space analysis of neural population dynamics confirmed a computational double dissociation: the vmPFC functioned as a value integrator where representations sharpened over time, while the ACC functioned as an information differentiator, with its neural state velocity precisely tracking the theoretical rate of information gain.Collectvely, these results map a hierarchical architecture where exploration is initiated by value regions, gated by oscillatory control, and realized through a dynamic interplay between the integral of value and the derivative of information.