Overcoming the costs of selective attention: Resolving rule-uncertainty supports acquisition of generalizable memory traces.

Research on category learning shows that people gradually allocate attention to deterministic features. Categorization models formalize this shift in attention as a process of "stretching" and "shrinking" dimensions of an internal representation, such that attended features bear a greater influence on choice than those ignored. Consistent with this view, gaze allocated to deterministic features increases alongside categorization accuracy. Going beyond the established relationship between gaze and performance, this work investigates whether gaze determines which features are encoded in the first place, as would be predicted by theories in which attention gates access to memory. Undergraduate participants (N=120) completed a category learning task while gaze was recorded, followed by tests of feature memory and category generalization. A baseline group (1) was cued to deterministic features to encourage selective gaze. Two additional groups tested manipulations intended to broaden gaze during learning: (2) was not cued to deterministic features; (3) was cued, but was also forewarned of the upcoming memory test. Mixed effects model analyses showed that gaze allocation during learning predicted both immediate categorization performance and the breadth of features encoded into memory. Effects were most robust when comparing groups (1) and (2). These findings suggest that selective attention constrains the representational content acquired during learning, but this cost is mitigated by early uncertainty about which features are most task-relevant.