A neural circuit framework for economic choice: from building blocks of valuation to compositionality in multitasking

Value-guided decisions are a cornerstone of cognition, yet the underlying circuit-level mechanisms remain elusive. How does the brain compute subjective value by integrating multiple features, such as quantity and probability? Where are learned preferences physically stored? And how do these computations generalize to novel situations? We address these questions by developing a biologically plausible, excitatory-inhibitory recurrent neural network trained on a diverse battery of economic choice tasks. Our analyses reveal a two-stage computational framework. First, value computation occurs upstream of the recurrent circuit, where learned input weights both encode the relative value between goods and approximate the non-linear multiplication of offer features. This feedforward valuation mechanism directly enables the network to generalize to unseen choice options. Second, value comparison is implemented within the recurrent circuit via a Competitive Recurrent Inhibition (CRI) mechanism, in which specific connectivity motifs between excitatory and inhibitory neurons mediate a robust winner-take-all decision. By training a single network on multiple tasks, we show the emergence of compositional representations that combine a shared computational schema with specialized neural modules. Our model reproduces key neurophysiological findings from the primate orbitofrontal cortex, unifying value computation, comparison, and generalization into a coherent framework with testable predictions for the neural basis of economic choice.