Sihan Yang¹ (), Yueying Dong¹, Anastasia Kiyonaga¹; ¹University of California, San Diego

Successful working memory (WM) must be flexible and constructive, allowing visual inputs to be manipulated and organized to optimize behavior. In long-term memory, this structuring may be supported by cognitive maps, which represent non-spatial information in an internal spatial framework to exploit navigation circuits. Such a mapping strategy may also be an efficient means of retaining and indexing novel information for WM. But it is unclear how quickly, and under what constraints, such maps are formed. Here, we show that people construct cognitive maps on the fly, organizing novel visual inputs that vary along multiple feature dimensions into an indexing scheme for WM. Participants viewed sequences of stimuli whose features varied along two dimensions – for instance, animal profiles varying in ear length or tail bushiness. The experiment comprised blocks that emphasized different types of judgments. ‘Match’ blocks required recognition judgments about whether the probe display depicted the exact features of stimuli from the preceding sequence. ‘Relative’ blocks asked participants to make relational judgments about how a feature changed from one stimulus to its successor in the sequence. We specified several hypothetical models of strategies for mentally representing and accessing the stimulus space. These included strategies for traversing a 2D cognitive map (where feature combinations are indexed as map coordinates) and a simple ‘sequence’ model (where the stimuli from a given trial are represented in their presentation order). We found that, in ‘match’ blocks, response times (RTs) corresponded best to the simple sequence strategy. However, in ‘relative’ blocks, RTs corresponded to the path length of traversing the hypothetical cognitive map. Likewise, in a parallel eye-tracking study, gaze dynamics during ‘relative’ blocks also tracked the hypothetical map paths. Thus, cognitive maps may be rapidly constructed for visual WM, especially when the task encourages relational encoding among novel pieces of information.