Pre-trained large language models demonstrate inherent capability to generate stable crystal structures for materials discovery without additional fine-tuning.

• MatLLMSearch framework combines LLMs with evolutionary search algorithms
• Achieves 78.38% metastable rate validation
• Generates novel material structures with desired properties
• Eliminates need for specialized training on materials databases

This breakthrough enables faster materials discovery for engineering applications by leveraging existing AI models, potentially accelerating development of new materials with specific industrial properties.

Large Language Models Are Innate Crystal Structure Generators