The integration of machine learning and automatic reasoning is a new trend in artificial intelligence. Constraint satisfaction is a classic problem in artificial intelligence. A large number of scheduling, planning, and configuration problems in the real world can be modeled as constraint satisfaction problems, and efficient solving algorithms have always been a research hotspot. In recent years, many new methods of applying machine learning to solve constraint satisfaction problems have emerged. These methods based on “learn to reason” open up new directions for solving constraint satisfaction problems and show great development potential. They are featured by better adaptability, strong scalability, and online optimization. This study divides the current “learn to reason” methods into three categories including message-passing neural network-based, sequence-to-sequence-based, and optimization-based methods. Additionally, the characteristics of various methods and their solution effects on different problem sets are analyzed in detail. In particular, a comparative analysis is conducted on relevant work involved in each type of method from multiple perspectives. Finally, the constraint solving method based on “learn to reason” is summarized and prospected.