Learning a new task is accompanied by changes at multiple levels in the brain. At the neuronal population level, learning requires coordinated adaptations to achieve optimal sensory processing and appropriate behavioral output. To understand how learning shapes high-dimensional neuronal dynamics across cortical areas, we simultaneously imaged the activity of neuronal populations in mouse somatosensory (S1) cortex and posterior parietal cortex (PPC) while mice were learning a multisensory texture discrimination task. Mice were trained to discriminate two textures, each preceded by a distinct auditory tone. During learning, we measured GCaMP6f calcium signals in S1 and PPC layer 2/3 neurons with a custom-built two-area two-photon microscope. We characterized the changes in four types of population subspaces: variance subspace, encoding subspace, within-area interaction subspace, and cross-area interaction subspace. We found that learning is associated with reconfiguration of the intrinsic subspaces, which aligned to maximize task-related information. Furthermore, behavioral errors are accompanied by unstable and misaligned interaction subspaces. These results show that learning is accompanied by intrinsic changes in population dynamics that optimize information processing and behavioral output.