Benjamin, L.; Morillon, B.; Wyart, V.

The ability to decipher the structure of sensory input is crucial for building an accurate internal model of one\'s environment. Humans can not only extract and integrate the sensory properties of stimulus sequences but also discover hidden rules governing their generation. However, sensory integration and the discovery of latent generative rules in sequences have been studied separately, and whether and how these two cognitive abilities interact remains unclear. We jointly investigated sensory integration and rule discovery in a visual prediction task where a hidden rule governed predictable changes in the sensory properties of presented stimuli - a rule that only half of tested participants discovered. Participants\' performance in sensory integration during the control blocks (without a hidden rule) accurately predicted their ability to discover the rule. Conversely, discovering the rule modified participants\' sensory integration strategy. Using computational modeling of human behavior and simulations of recurrent neural networks, we show that minimizing leak during sensory integration is the sole parameter with a selective causal impact on subsequent rule discovery. Furthermore, participants who discovered the hidden rule appeared to apply it mentally to each presented stimulus, prior to integrating it. Together, our results reveal a deep bidirectional interplay between sensory integration and latent rule discover, to form a comprehensive learning system.