CrystalReasoner: LLM Framework for Crystal Structure Generation via Reasoning and RL

CrystalReasoner is a comprehensive framework for large language models (LLMs) designed to create crystal structures from textual instructions. It overcomes challenges found in current generative models: LLMs often lack atomic accuracy, and diffusion-based approaches struggle to incorporate advanced scientific insights, leading to the creation of invalid or unstable structures. By introducing physical priors as conceptual tokens—such as crystallographic symmetry, local coordination environments, and anticipated physical properties—CrystalReasoner effectively connects natural language with three-dimensional structures before generating atomic coordinates. Additionally, it utilizes reinforcement learning with a multi-objective, dense reward function to ensure that the generation process adheres to physical constraints. This framework is elaborated in arXiv paper 2605.14344.