Learning to predict others' minds is critical for social cognition, but the underlying computation and neural mechanisms remains largely unknown. According to theories in social cognition, a simple conception is that humans simulate others' mental processes by directly recruiting one's own process to model others' minds. In this review, we first describe our recent finding and discuss its possible implications. Using human fMRI with model-based analysis on frameworks of reinforcement learning and value-based decision making, we found that simulation involves two hierarchical learning signals: a reward prediction error, generated by simulation of direct recruitment to model others' valuation, and an action prediction error, based on simulation and observation of the other's choices to track others' variability. These findings show that humans can learn to predict others' minds from simulation, using a scaffold of mentalizing signals. Then, we discuss prospects that theoretical neuroscience and computational approaches will play significant roles in understanding human behavior and neural mechanisms, leading to the so-called computational psychiatry as well as synthesis over different disciplines to study human.