LEARNING DYNAMICS IN A HIERARCHICAL SEQUENCE TASK

Sequential behaviour is often structured hierarchically, requiring animals to track latent task states over extended action sequences. Recent work in mice performing a looping hidden-state sequence task (ABCD) demonstrated that populations in medial frontal cortex encode, at any given moment, the animal's future trajectory through the task, enabling zero-shot inference and generalisation across task structure (El-Gaby et al., 2024). How such representations extend to hierarchical sequences remains unknown. Here, we trained rats on an eight-step looping sequence task with two embedded levels of hierarchy. Animals performed the repeating sequence ABABCDCD on a poke-wall apparatus, in which each physical action is shared across 2 distinct latent states. Rats differentiated their next action significantly between at least one set of aliased states (e.g., 1st B vs 2nd B and/or 1sr D vs 2nd D) in each task, indicating that they tracked their position within the sequence across at least five states and identified differences in transition probabilities between otherwise identical action states. In the hierarchical task, we observe distinct behavioural strategies before animals reach stable performance, suggesting that multiple intermediate solutions can support partial learning of task structure. We are modelling learning in this task using recurrent neural networks and tracking how internal representations evolve across training. These behavioural and modelling observations motivate recording neural activity in the hierarchical task, where multiple representational solutions may support performance at different stages of learning.