Strategy execution in cognitive skill learning: an item-level test of candidate models

This article investigates the transition to memory-based performance that commonly occurs with practice on tasks that initially require use of a multistep algorithm. In an alphabet arithmetic task, item response times exhibited pronounced step-function decreases after moderate practice that were uniquely predicted by T. C. Rickard's (1997) component power laws model. The results challenge parallel strategy execution models as developed to date and they demonstrate that the shift to retrieval is an item-specific, as opposed to task-general, learning phenomenon. The results also call into question the entire class of smooth speed-up functions as global empirical learning laws. It is shown that overlaying of averaged item fits on averaged data can provide a sensitive test for model sufficiency. Strategy probes agreed with strategy inferences that were based on step-function speed-up patterns, supporting the validity of the probing technique.