Policy regularization as a unifying theory of the striatal division of labor in learning

Policy regularization explains the DLS–DMS division of labor in learning by modeling DMS as KL-regularized flexibility around a DLS history-dependent default policy.

• Why it matters: The work addresses how agents balance cached, history-driven actions with the need to adapt when reward contingencies shift, a trade-off linked to striatal subregions.
• What they did: It formalizes a KL-regularized reward objective where a flexible DMS controller pays a cost for deviating from a default policy conditioned on action history, with DLS caching frequently used transitions.
• The result: DLS consolidates transitions shaped by DMS value learning, yielding robust but inflexible action selection and a single model that reproduces DLS–DMS dissociations across multiple task paradigms.

AI summary of the abstract · may contain errors · not medical advice · read the paper.