HIPPOCAMPAL REPRESENTATIONS OF ABSTRACT LATENT STATES IN FLEXIBLE BEHAVIOUR

Flexible behaviour depends on the ability to extract structure from experience and apply it to novel situations. This is thought to rely on latent states – internal representations that capture hidden structure in the world and provide behavioural frameworks for planning and decision-making. To support flexibility, these models must also be compositional, allowing generalisation to new scenarios through the combination and rearrangement of their components. The hippocampus is known to organise knowledge by binding stimuli into representations of discrete states, from spatial locations and contexts to maps of social relationships. While growing evidence suggests that it also encodes abstract latent states defined beyond immediate sensory inputs, the neural mechanisms supporting their representation and compositional use remain poorly understood.
Using behavioural modelling, miniscope imaging, and caspase lesions in a probabilistic reversal learning task in mice, we showed that the ventral hippocampus (vH) is required for the inference and representation of abstract latent states. To understand how such representations support more complex, naturalistic behaviour, we used a grid-maze task in which animals navigate between towers connected by bridges to reach different cued locations. Importantly, rather than solving the maze in a step-by-step manner, mice develop stereotyped routes that abstract over individual locations into behaviourally relevant latent states, and sequence them into compositional paths. Preliminary analyses of Neuropixels recordings of vH activity during maze navigation suggest that vH neurons encode extended maze segments. Ongoing work investigates how these representations relate to behavioural routes, their development during learning, and compositional use across different maze structures.