What is it about?
Scientific progress relies on novel ideas, but the sheer volume of research makes them difficult to identify, especially across disciplines. We introduce SCIMUSE, an AI system combining a knowledge graph from 58 million papers and large language models to generate personalized research ideas. Yet how compelling are these ideas in practice? To find out, as a core contribution, we conducted a large-scale expert evaluation in which over 100 research group leaders assessed thousands of suggestions across diverse domains. We observe correlations between graph features and expert-rated interest and show that machine learning can use these features to predict interest above chance. This ability suggests a future route toward prioritizing candidate ideas and potentially supporting scientific creativity and cross-disciplinary innovation
Featured Image
Photo by Skye Studios on Unsplash
Why is it important?
Our study shows that AI can generate research ideas that scientists find interesting, but identifying such ideas remains a significant challenge. In a large-scale evaluation involving researchers from many disciplines, we found that only about one quarter of the generated ideas received high interest ratings. We also found that ideas predicted to have greater scientific impact were not necessarily more interesting to researchers. Importantly, our results show that AI models can learn to predict which ideas scientists might find interesting. These findings provide new ways to study scientific interest and could help develop AI systems that generate interesting research ideas.
Perspectives
What I found particularly interesting about this project was the opportunity to learn how scientists from very different fields respond to new research ideas. Organizing such a large evaluation was challenging, and some of the findings were quite unexpected. The experience also made me think more deeply about how scientists develop their research interests and decide which ideas are worth pursuing. I hope this work encourages more research into scientific creativity and how AI can support it. There are still many open questions, and I look forward to exploring some of them in future work.
Xuemei Gu
Friedrich-Schiller-Universitat Jena
Read the Original
This page is a summary of: Generating interesting scientific ideas using knowledge graphs and LLMs: Evaluations with 100 research group leaders, Proceedings of the National Academy of Sciences, November 2026, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2611349123.
You can read the full text:
Contributors
The following have contributed to this page
