Neural dynamics of adaptive learning in sensorimotor cortex

Neural activity is tightly coupled to body dynamics, and predictive internal models are thought to support sensorimotor control by computing prediction errors. Here, we test whether such errors are encoded in primary somatosensory (S1) and motor (M1) cortex during motor adaptation in mice. Using control theory–derived features, we show that functionally distinct neurons map onto specific computational motifs. Layer 2/3 population dynamics encode command-like signals and sensorimotor prediction errors (SPEs), with S1 neurons representing SPEs more prominently than M1 and exhibiting larger learning-related changes in latent dynamics. To identify the coordinate frameworks of these errors, we developed a 50-muscle biomechanical model of the mouse forelimb in a physics simulator. We find that SPEs are represented in both 3D task space and muscle space. Together, our results provide a new model of how S1 dynamics support adaptive motor learning.