Learning shapes neural codes for sensory-motor integration in the tail of the striatum.

Separating meaningful sensory stimuli from irrelevant ones requires learning sensorimotor associations, but how sensory-linked striatal circuits acquire and maintain these associations is unclear. We longitudinally imaged direct- and indirect-pathway (D1 and A2a) spiny projection neurons (SPNs) in the tail of the striatum (TS) as mice learned to push or pull a joystick in response to auditory cues in either a stimulus-response association (go/omit) task or a two-alternative forced choice (2AFC) task. Learning in both tasks increased the fraction and strength of task-modulated TS SPNs across the sound, action, and reward epochs, yet individual neuron selectivity often switched over days between behavioral epochs. In spite of individual neuron variability, population activity of direct and indirect pathways became aligned with characteristic behavioral features during learning: D1-SPNs dominated the 'action' category, A2a-SPNs were biased toward the 'mixed' category (multiple epochs), and both SPN types showed 'sound' category specificity that depended on the sound-action association. Trial-wise modeling revealed a reweighting of behavioral predictors within the action window, with reward gaining and movement losing predictive weight. Learning the two-choice task led to a higher prevalence of association-preferring neurons and better behavioral decoding within the sound window than in the action/reward window, reflecting a task-dependent prioritization of sensory information. Association-preferring neurons also showed a stable local distance-similarity relationship, with nearby neurons more similar than distant neurons across learning. Together, our results support a population mechanism in TS during learning in which neurons from both direct and indirect pathways are recruited and take on distinct behavioral roles that vary with performance and task complexity.