Statistical learning prioritizes abstract over item-specific representations.

Statistical learning optimizes limited working memory by abstracting probabilistic associations among specific items. However, the cognitive mechanisms responsible for the working memory representation of abstract and item-specific information remain unclear. This study developed a learning-memory representation paradigm and tested three participant groups across three conditions: control (Experiment 1), item-specific encoding (Experiment 2), and abstract encoding (Experiment 3). All groups were first shown picture-artificial-character pairs that contained abstract semantic categories at high (100%), moderate (66.7%), and low (33.3%) probability levels and item-specific information (16.7%). Participants then completed an online visual search task that simultaneously assessed statistical learning and memory representation by examining how abstract or item-specific distractors influenced their speed for searching artificial characters. In the control condition, participants spent more time searching abstract than item-specific distractors across all probability levels, indicating abstract prioritization. In the item-specific condition, abstract prioritization was absent. In the abstract condition, enhanced prioritization of abstract information was observed for moderate and low, but not high, probability items. These findings suggest that statistical learning is central to the abstraction process, with input probabilities and encoding strategies jointly shaping the formation of abstract and item-specific representations. This process depends on a flexible working memory system that dynamically adjusts prioritization, particularly when inputs are uncertain.