Introduction
As recent Nobel Prizes recognize AI's contributions to science, large language models (LLMs) are transforming scientific research by boosting productivity and reshaping scientific methods. Today, LLMs are already involved in experimental design, data analysis, and workflows, with particularly notable impacts in chemistry and biology.
Recent advances in artificial intelligence have transformed many aspects of society, the global economy, and academic research practices. Generative AI and LLMs bring unprecedented opportunities to transform scientific practice, drive scientific development, and accelerate technological innovation. The Nobel Prizes in Physics and Chemistry have been awarded to several AI leaders, recognizing their contributions to AI and cutting-edge models such as LLMs. This signals that LLMs will transform or advance scientific research by enhancing productivity and supporting all stages of the scientific method. The application of AI in science is burgeoning, spanning numerous scientific fields and influencing different parts of the scientific process.
Despite the great potential of LLMs in hypothesis generation and data synthesis, AI and LLMs still face challenges in fundamental science and scientific discovery. Therefore, our perspective is that, so far, AI's impact on fundamental science—defined here as the discovery of new principles or new scientific laws—has been limited. This article reviews how LLMs are currently used as technical tools to augment the scientific process, and how they might be used in the future as they become more powerful and evolve into robust scientific assistants. By combining data-driven techniques with symbolic systems, such systems can merge into hybrid engines, potentially leading new research directions. We aim to describe the gap between LLMs as technical tools and as "creative engines" capable of fostering high-quality scientific discoveries and posing new questions and hypotheses to human scientists. We first review current applications of LLMs in science, aiming to identify limitations that need to be addressed to move toward creative engines.
Current Applications
LLMs have been widely used in scientific literature analysis, experimental protocol generation, data parsing, and preliminary hypothesis formation. For example, in drug discovery, LLMs can help predict molecular properties; in materials science, they can recommend synthesis pathways. However, most of these applications remain at the assistive level and are not yet capable of independently generating original scientific insights.
Challenges and Limitations
Core challenges facing LLMs include a lack of causal understanding of the physical world, a tendency to "hallucinate," reliance on the quality and biases of training data, and insufficient cross-domain reasoning abilities. In fundamental science, discovering entirely new principles requires abstract reasoning capabilities that go beyond pattern matching—an area where current LLMs are deficient.
Future Directions
To achieve the leap from tools to creative engines, it is necessary to develop hybrid systems that combine LLMs' linguistic capabilities with symbolic reasoning, knowledge graphs, and physical simulations. In addition, establishing clear evaluation metrics is crucial to ensure that LLM outputs align with human scientific goals. Interdisciplinary collaboration and ongoing human-machine alignment will be key.