Learning is not only the acquisition of knowledge, but also the ability to express that knowledge when needed. Previous work using non-reinforced probe trials showed that animals exhibit complete acquisition of stimulus-action associations far before they express them under reinforcement. Why do animals exhibit this gap in performance between acquisition and expression despite already acquiring the stimulus-action associations? Early in learning, animals may (1) exhibit motor biases that they slowly learn to suppress, (2) continue to explore different choice alternatives, or (3) base decisions on recent trial history, including choice and reward, rather than current stimuli. To test between these and other potential drivers, we trained mice on a balanced, two-alternative forced choice auditory task. During probe trials, animals exhibited surprisingly high accuracy early in learning even when performance in reinforced trials was near chance levels. In addition, animals exhibited less directional biases in choices during probe comparing to reinforced trials early in learning. We investigated the nature of this directional bias further using a generalized linear model to separate the influence of action bias and trial history. The full model fit individual learning curves well; action bias, but not trial history, emerged as the most important contributor. Removing both factors from the model ‘rescued’ behavioral performance and bridged the gap between reinforced and probe trials. The action bias itself was itself dynamic – animals would execute blocks of right or left choices with rapid transitions between the two, reflecting a potential sampling strategy rather than a motor bias. Choices during biased blocks were also faster and less deliberative. Together, our results suggest that behavioral expression reflects a potential exploratory process unique to individual animals that is uncoupled from the acquisition of the core task knowledge.