Cognitive constraints and thalamocortical architecture explain systematic biases and neural signatures in human hierarchical decision-making

Human cognition operates under capacity limitations, requiring computational adaptations for solving complex tasks. Although such adaptations have been described algorithmically, their neural substrates and mechanistic implementations remain poorly understood. Here, we study a hierarchical decision-making task in which subjects must simultaneously infer latent rule changes and resolve sensory uncertainty. By comparing human behavior with Bayesian observer models, we identify three systematic deviations: humans make decisions hierarchically, persist in the previous context after a context switch, and preferentially explore a new context within the same sensory modality. We argue that these deviations are jointly explained by a cognitive-constraint perspective in which hierarchical decisions are less cognitively demanding and switching context or modality incurs additional cost. Incorporating these constraints substantially improves behavioral fits and reveals correlated subject-level context and modality stickiness parameters, suggesting a shared underlying cost mechanism. We then asked how neural representations support hierarchical decisions. Because hierarchical decisions require selecting a latent context before choosing an action, context representations provide a natural substrate for reducing the effective decision space. Neuroimaging data revealed that context information is represented in mediodorsal thalamus, consistent with a role for thalamus in task-state compression. Motivated by the context signal in thalamus, we trained a thalamocortical recurrent neural network in which a thalamic module and faster corticothalamic plasticity provided an inductive bias for task compression, while switch-dependent working-memory noise instantiated switching costs. These constrained networks better matched human behaviorial bias than standard task-optimized networks and reproduced key neural signatures observed in the data, including context encoding in mediodorsal thalamus and cue/rule representations in prefrontal cortex. Together, these results suggest that structured biases in hierarchical decision-making arise from cognitive constraints and identify thalamocortical dynamics as a candidate mechanism for efficient task-state inference.