​EMERGENCE OF COGNITIVE VARIABLES IS SUBJECT TO ARCHITECTURAL CONSTRAINTS IN RECURRENT DECISION CIRCUITS

Reinforcement learning specifies goals for neural circuits (i.e., maximize reward), but it does not determine how internal representations should be organized. In biological systems, however, value, choice, and confidence appear in distinct circuit stages, suggesting that architecture constrains which cognitive variables can exist at all. We investigate this tension between learning objectives and circuit structure using a bimodular actor–critic recurrent network trained on an economic choice task and transferred into a spiking twin while preserving latent geometry (Fig. a). The architecture comprises a control-oriented “actor” that selects actions and an evaluative “critic” that predicts reward from the actor’s state. Although both modules are optimized for the same task, population analyses reveal a systematic division of representational roles: action selection is supported by low-dimensional value-difference coordinates in the actor, whereas object-centered value representations emerge selectively in the critic (Fig. b). Decision confidence is not explicitly trained and is not localized to specific neurons; instead, it arises as a population-level invariant of the same geometry that supports choice (Fig. c). By expanding network size while preserving dynamics, we show that valuation variables require population redundancy for stability (Fig. d-e), whereas action selection remains robust in compact circuits. Finally, we demonstrate that a minimal actor–critic architecture cannot generate categorical commitment states, despite successful learning, revealing a principled architectural limit. These results show that circuit hierarchy and population structure fundamentally constrain which cognitive variables can emerge in recurrent neural systems. Therefore, learning goals alone are insufficient to determine internal codes.