Seeking effective school and family supports for adolescents' curiosity in learning: A cross-country comparison among China, the USA and Finland.

The sensory environments we live in are variable, yet highly organized. Understanding the generative organization that governs their structure is crucial for reducing uncertainty. To do so, humans rely on multiple inference abilities, including sensory integration (–), temporal prediction (–), categorization based on arbitrary mappings of sensory features (–), learning statistical dependencies between stimuli (–), and the learning of hidden rules describing the patterns of relations between successive stimuli (–). In a majority of studies, these different learning abilities are assumed, either implicitly or explicitly, to rely on different cognitive and neural mechanisms. But some studies, on the opposite, suggest a single unified cognitive apparatus for those different learning processes (). This debate lacks empirical data, and these abilities themselves have rarely been studied conjointly, resulting in a fragmented description of human learning abilities. Here, we studied three human inference capacities conjointly and empirically tested if they rely on similar or distinct and independent or interrelated cognitive processes: sensory integration, temporal prediction, and rule discovery. So far, inference capacities have mostly been studied in isolation, targeting different research questions and using different methods. In perceptual decisions research, studies have examined how sensory information is integrated across sequences of stimuli—or across samples from the same stimulus—to reach a decision (, , , ). We refer to this ability as . This process is known to be prone to a number of suboptimalities constraining the accuracy of perceptual decisions across individuals (–). Yet, it remains largely unknown whether these sensory integration suboptimalities are confined to sensory integration itself, or whether they act as a broader cognitive bottleneck impairing higher-order inferences, such as rule discovery. To investigate this possible effect of sensory integration limitations on higher-order inferences, we designed a rule-based visual prediction task that requires both sensory integration and the discovery of a hidden rule for accurate predictions. Rule discovery is known to depend on the rule’s algorithmic complexity and to vary across individuals (, ). However, except for a few specific studies (), sequences used in most rule discovery tasks are typically stripped of any variability, eliminating the need to integrate sensory information across stimuli. In our rule-based prediction task, we used rhythmic isochronous sequences of 10 oriented stimuli () sampled from probability distributions whose mean orientations switch (i.e., rotate by 90°) after five stimuli—a latent rule that was not described to participants. After exposing participants to full sequences, we asked them to predict the 10th stimulus from incomplete sequences of 3, 5, 7, or 9 stimuli ( blocks, ). In other blocks ( blocks), we used sequences where all stimuli were drawn from a single distribution—without a latent rule—to investigate if suboptimalities in sensory integration measured in static blocks could impact discovery of the latent rule in switch blocks. By introducing interstimulus temporal jitter, we further investigated whether a third inference ability—temporal prediction—affected either of these two forms of inference, or their interrelation.