Generative machine learning methods have advanced crystal discovery but cannot fully explore material candidates that are both novel and useful. A reinforcement learning-based method now steers candidate generation toward these areas, enabling the design of novel functional materials. The approach is described in a News & Views article published on 03 August 2026 in Nature Machine Intelligence by Zhendong Cao and Lei Wang of the Institute of Physics, Chinese Academy of Sciences, Beijing, China. The authors declare no competing interests.