The developmental trajectories of generalization and discrimination in reinforcement learning reflect multiple underlying cognitive processes

A hallmark of human intelligence is the ability to appropriately generalize past experiences to new scenarios. Organizing concepts into categories allows the brain to generalize learned stimulus-reward contingencies to new stimuli within the same category. However, when category membership does not fully predict associated rewards, the brain needs to balance generalizing values within categories and discriminating value differences among individual items. How do generalization and discrimination in reward learning develop throughout lifespan as semantic knowledge is gradually acquired? Using a probabilistic reward learning task with a hierarchical value structure organized by semantic categories, we found children can generalize values of stimuli based on semantic categories at least as early as three to four years old. Furthermore, age-related performance improvement is explained by a combination of increased decision certainty, decreased initial exploration, change in memory retention, and a reduced categorical learning rate.