STEM education has expanded dramatically over the past decade, yet significant learning challenges remain. In light of these, my paper examines how students in STEM can be guided to learn effectively through the design of the course grading scheme. To do so, I first gather rich administrative and survey data covering nearly 3,700 undergraduates at a large public university taking an online introductory programming course that has a cumulative structure. The data allow me to monitor students' study time precisely and to characterize whether they are forward-looking. I then develop and estimate a multi-stage behavioural model of student effort supply. The marginal benefits and costs of effort at each stage of the cumulative learning process are credibly identified using field experiments covered in other related work. The estimated model allows me to explore the efficacy of changing assignment grading weights to improve student learning. I find that the simulated weights that maximize learning are decreasing across assignments, serving to increase effort by myopic students early in the course when they acquire foundational skills. Additional simulations suggests that in a course with a strong cumulative structure, incentives should be front-loaded when majority of students are myopic, middle-loaded when a modest share of students are myopic, and end-loaded when the vast majority of students are forward-looking.