SPECIALIZED COMPUTATIONS FOR GENERALIZED WORLD MODELLING IN PREFRONTAL CORTEX

What separates human flexibility and reasoning from current artificial systems is the endowment of an internal world model. The prefrontal cortex (PFC) is thought to learn generalizable world models but the specializations that enable this are unknown. We tested whether prefrontal specializations are representational (selecting domain-specific features) or computational (implementing domain-general computations). During fMRI, participants learned probabilistic features of virtual environments that represented spatial, social, and sequential domain knowledge. Although each domain used different features, they shared the same feature-to-latent state mapping, matching their computational demands. This design dissociates domain-specific features from invariant latent computations.

Using Representational Similarity Analysis (RSA) we found no evidence for observable domain feature-specific representations in PFC.
Instead, PFC patterns revealed a triad of specialized yet domain-general computations adaptively learning internal models. Comparing multiple learning mechanisms, a Monte Carlo sampling model best explained participants’ behaviour. The trial-by-trial estimates reflected participants’ internal model updates and was subsequently used for RSA. The ventromedial PFC patterns reflected probabilistic inference, abstracting out hidden probability distributions into task states. Neural patterns in this region were sensitive to (posterior) state changes within a low-dimensional latent space. Anteromedial PFC patterns tracked local directional shifts within each task state and switches between different states, organized along orthogonal axes, suggesting a global task coordinate system. Dorsomedial PFC patterns represented future observations given past history and current state, signalling task transitions. Together, these results reflect a general-purpose inductive bias for world-modelling in medial PFC, possibly indicating a modular specialized architecture humans use for internal models.