Learning about marriage quality has been proposed as a key mechanism for explaining how the probability of divorce evolves with marriage duration, and why people often cohabit before getting married. I develop four theoretical models of divorce, three of which include learning. I use data from the Survey of Income and Program Participation to test reduced form implications of these models. The data is inconsistent with models including a substantial amount of learning. On the other hand, the data is consistent with a model without any learning, but where marriage quality changes over time.