Hippocampal dendritic integration as hidden state inference

Many of the most important variables guiding behaviour, such as reward contingency, spatial location or contextual state, cannot be measured directly. Animals must instead infer them from noisy, incomplete information, a process termed ‘hidden state inference’. This requires combining internally generated estimates with incoming evidence to produce accurate beliefs about the world that guide behaviour. While computational frameworks can describe this process in formal terms, its implementation in the brain remains unclear. Here, we propose that the compartmentalised dendritic architecture of CA1 pyramidal neurons provides a biophysically grounded mechanism for performing the core operations of hidden state inference. In this framework, apical dendrites encode internally generated state estimates, tuft dendrites encode evidence from sensory and contextual inputs, and their nonlinear interaction produces an updated state estimate. Structured heterogeneity in dendritic properties across neurons would naturally generate a population code capable of representing graded belief strength, uncertainty, and multiple competing hypotheses. We illustrate how this scheme can account for CA1 function in both reversal learning and spatial navigation, and suggest that similar dendritic principles may operate more widely across brain regions. This perspective links cellular physiology to high level cognitive computation, and yields concrete, testable predictions for how dendritic integration supports adaptive behaviour in dynamic environments.