The emergence of all-or-none retrieval of chunks in verbal serial recall.

People often subdivide a list into smaller pieces, called chunks. Some theories of serial recall assume memories are stored hierarchically, with all-or-none retrieval of chunks, but most mathematical models avoid hierarchical assumptions. Johnson (Journal of Verbal Learning and Verbal Behavior, 8(6), 725-731, 1969) found steep drops in errors following correct recalls (transitional-error probabilities) within putative chunks during multi-trial letter-list learning, and viewed this as evidence for all-or-none retrieval. Here we test whether all-or-none retrieval occurs in lists studied only once. In serial recall of six-word lists (Experiment 1), transitional-error probabilities were inconsistent with all-or-none retrieval, both when participants were instructed to subdivide and when temporal grouping induced subdivision. Curiously, the same analysis of previous temporally grouped nine-letter lists produced compelling evidence for all-or-none retrieval, which may result from recoding rather than the formation of chunks. In Experiment 2, participants were pre-trained on three-word chunks. For nine-word lists constructed from those trained chunks, transitional-error probabilities exhibited more pronounced evidence of all-or-none retrieval. Nearly all effects reversed with post-cued backward recall, suggesting mechanisms that play out over the course of recall rather than encoding of the list. In sum, subdivided lists do not result in hierarchical memories after a single study trial, although they may emerge in lists formed from chunks that are previously learned as such. This suggests a continuous transition from non-hierarchical subdivision of lists to all-or-none retrieval over the course of chunk formation.