Orthogonal representational geometry in dACC underpins human hierarchical reasoning.

Flexible cognitive control requires inferring the causes of feedback and adjusting accordingly. When tasks are hierarchically structured with high-level rules governing low-level perceptual judgments, good performance requires hierarchical reasoning. Yet how the human brain represents and integrates multiple task-relevant variables to support the computations required for hierarchical reasoning remains unclear. Here, building on Bayesian modeling and recurrent neural network simulation of hierarchical reasoning, we show that such algorithm and its representations are implemented in the human dorsal anterior cingulate cortex (dACC). Accumulated errors and perceptual difficulty are integrated in dACC to estimate high-level rule switch confidence. These variables are represented along two separable dimensions that can be approximated by basis functions. Orthogonality between them supports a two-dimensional representation well suited for hierarchical reasoning. A closer-to-orthogonal representational geometry also correlates with accurate computation and better performance. Together, these findings provide an integrated computational and neural explanation for hierarchical reasoning.