In skill-learning tasks, reaction times (RTs) typically decrease with practice. For example, in alphabet arithmetic tasks(e.g. J + 7 = ?), learners respond correctly (e.g. Q) faster on later than on earlier trials. A number of mathematicalmodels have been proposed to account for the functional form of practice-related RT speedup. We aim to evaluate whichof two candidates better fits observed speedup data for individual learners across several tasks. In particular, we comparea process-shift account in which learners initially execute an algorithm in constant time, but as trials accumulate, exhibitpower-law speedup as they directly retrieve a memorized solution to a delayed exponential model in which RTs decreaseexponentially after learners eventually achieve insight into a task-appropriate strategy. Using hierarchical Bayesian modelsof each account (which can flexibly model learning in individual subjects), we show that the process-shift model betterpredicts out-of-sample data than the delayed-exponential model.