Computational basis of hierarchical and counterfactual information processing

Abstract

Humans solve complex multistage decision problems using hierarchical and counterfactual strategies. Here we designed a task that reliably engages these strategies and conducted hypothesis-driven experiments to identify the computational constraints that give rise to them. We found three key constraints: a bottleneck in parallel processing that promotes hierarchical analysis, a compensatory but capacity-limited counterfactual process, and working memory noise that reduces counterfactual fidelity. To test whether these strategies are computationally rational—that is, optimal given such constraints—we trained recurrent neural networks under systematically varied limitations. Only recurrent neural networks subjected to all three constraints reproduced human-like behaviour. Further analysis revealed that hierarchical, counterfactual and postdictive strategies—typically viewed as distinct—lie along a continuum of rational adaptations. These findings suggest that human decision strategies may emerge from a shared set of computational limitations, offering a unifying framework for understanding the flexibility and efficiency of human cognition.