Emergence of Value and Action Codes in Bimodular Spiking Actor-Critic Networks

Abstract

Decision-making depends on coordinated computations distributed across dorsal and ventral circuits, often described as actor-critic systems. We examined how this division of labor gives rise to value- and decision-related representations by training a bimodular recurrent network on an economic choice task and transferring it to an expanded gated spiking model that preserved its latent dynamics while stabilizing the underlying neural representations. The two modules assumed complementary roles: the actor encoded action-value differences and spatial contingencies required for selecting between actions, whereas the critic represented object values predictive of reward. Neurons across modules were selective for object, action, total, and difference values, supporting decision- and confidence-related activity. Low-dimensional analyses revealed structured trajectories reflecting temporal evolution, spatial configuration, and value-dependent divergence. These results clarify how distributed circuits can jointly implement valuation and action selection, providing a foundation for linking reinforcement-trained recurrent models with anatomically inspired dorsal-ventral frameworks.