SPRINT trains large reasoning models to dynamically spot and run independent subtasks in parallel, reducing sequential tokens by up to 40% on complex tasks like math without sacrificing accuracy compared to sequential reasoning baselines.