My current research focuses on (1) integrating deep learning with symbolic modules; (2) making reinforcement learning more efficient; (3) applications in real-world natural language understanding and program synthesis tasks. For example, my recent works designed a hybrid neural network model (Neural Symbolic Machines) and proposed novel policy gradient methods for efficient training. It is the first end-to-end neural network model that achieves new stateof-the-art on the most challenging semantic parsing benchmarks. The model learns from weak supervision to answer compositional questions by generating programs to query and process data from knowledge graph or database tables.