Behavioral performance during goal-directed learning is typically measured in the presence of reinforcement. In this context, learning has been described as a slow process with high inter-subject variability. Exploration of the neural mechanisms, therefore, has focused on identifying dynamics concomitant with these slow performance improvements. Recent work, however, has shown that task acquisition is much faster and more stereotyped than previously thought1. Performance was evaluated daily in reinforced and non-reinforced (‘probe’) trials. These probe trials revealed a rapid and stereotyped acquisition of task contingencies early in learning which was only expressed much later in reinforced trials. Here we ask whether and how sensory cortical networks encode and control the acquisition of this latent knowledge. We used longitudinal, two-photon calcium imaging of the same large population of excitatory neurons in layer II/III of the auditory cortex (AC) while mice learned to lick to a tone to obtain a water reward (S+) and withhold from licking to another tone (S−) to avoid a timeout. We used unsupervised low-rank tensor decomposition to uncover low-dimensional network dynamics at different timescales. We identified a subset of neurons that were initially S+-driven but then shifted within the first 400 trials to firing at the time of reward, suggesting a role in reward learning. Another subset of S−-responsive neurons exhibited a rapid enhancement of their stimulus-driven response, suggesting a role in behavioral inhibition. Latent knowledge of the task was thus manifest in cortical networks during reinforced trials despite poor task performance. To test the causal role of these dynamics, we optogenetically silenced the AC on reinforced trials and assayed performance on light-off probe trials. AC silencing led to a striking delay in the acquisition of stimulus-action associations. Overall, our work argues that latent task knowledge emerges rapidly in the AC and is crucial for goal-directed learning.