Fig. 2. Functional architecture of the prefrontal cortex contributing to adaptive behavior.

Medial and lateral view of the human prefrontal cortex (PFC) and its main anatomical regions shown in relation with their contribution to adaptive behavior. The laOFC (lateral orbitofrontal cortex), vmPFC (ventromedial PFC), dmPFC (posterior and anterior dorsomedial PFC) along with premotor cortex (and possibly clPFC) are present presumably in all mammals from rodents to primates and humans. The lateral PFC including clPFC and especially midlPFC (mid-lateral PFC) emerges in primates, whereas fpPFC (frontopolar cortex) is specific to humans. In the proposed framework, the laOFC encodes stimulus reward values (S-Rew), posterior dmPFC action reward values (A-Rew). The vmPFC encodes predictive models involving learning (Stimulus–)Action–Outcomes associations. The lateral premotor cortex encodes low-level selective models (Stimulus–Action associations), whereas the clPFC encodes higher-level selective models (Cues-(S–A) associations). Task sets (TS) form large-scale neural frames linking such internal models encoded in these various PFC regions in order to potentially invoke them together to guide behavior. TS reliability is the ability of TS internal models to jointly predict external contingencies. midlPFC learns contextual models predicting TS reliability according to external cues. The actor TS is the TS driving ongoing behavior and which reliability is monitored in the vmPFC. Counterfactual TS are the TSs which reliability are monitored in the fpPFC without contributing to ongoing behavior. White arrows indicate major information flows related to actor task set reliability (medial PFC) and counterfactual task-set reliabilities (lateral PFC). Black arrows indicate major information flows related to reward values of action outcome (ventral PFC) and reliability-based inhibition or selection of actor TS in the dorsal PFC. See text for more explanations.