Perceptual learning reformats working memory codes along a reverse visual hierarchy

Mastering fine perceptual skills requires maintaining precise visual information while simultaneously processing new inputs. In the visual cortex, sensory and mnemonic representations coexist through region-wide multiplexing, yet how perceptual experience reshapes these neural codes remains unknown. Here, combining five days of intensive perceptual training with fMRI and multivariate pattern analyses, we tracked how perceptual learning reformats mnemonic neural representations in a motion-direction discrimination task. Learning improved motion-direction discrimination sensitivity and reshaped the distribution of sensory- and mnemonic-format codes across the visual hierarchy. Sensory-format maintenance was enhanced in V1, and the magnitude of this enhancement closely tracked the behavioral learning effect across individuals, whereas mnemonic-format coding was reduced in V3A. Within V1, learning further increased temporal coupling between sensory and mnemonic codes while leaving their near-orthogonal coding geometry largely unchanged. Our results suggest that perceptual learning reformats working memory codes along a reverse visual hierarchy, strengthening high-fidelity sensory-like representations while preserving the separable organization of multiplexed codes. These findings identify reverse-hierarchy reformatting as a hallmark of experience-driven working memory plasticity.