<p>Humans often learn better when problems are broken down into parts, but this phenomenon has eluded explanation at the computational level. Here we study how differing training curricula help or hinder learning in a classic probabilistic cue combination task. Training curricula that ‘divide and conquer’ by presenting one cue at a time facilitate later performance on test trials involving multiple cues. This effect is captured by a hybrid learning framework that arbitrates between two different learning strategies: a marginal updating process, which assigns credit to each cue independent of every other, and a joint updating process, which distributes credit across cues on the basis of their joint presence. We use this theory to generate new ‘skewed distribution’ multi-cue curricula that should and should not successfully promote human learning. It makes accurate predictions, demonstrating that we can use computational insights of learning to accelerate human probabilistic learning.</p>