Lei Lin, Ronghao Wang, Chunbao Zhou, Jue Wang, Yangang Wang
DN-Hypo-Pipeline is an AI workflow that uses scientific explanations as prior knowledge to enable LLMs to systematically generate hypotheses.
Scientific hypothesis generation is the first step in research, but existing methods merely summarize literature or find patterns without deep scientific reasoning. There is a lack of systematic methods for researchers to deeply understand phenomena and derive new hypotheses.
DN-Hypo-Pipeline takes the conclusion (explanandum) of a research paper as input, identifies underlying laws, theories, and principles supporting that conclusion, and reconstructs a new, yet-to-be-verified explanation for the observed phenomenon. It automates this process using LLMs and validates quality through statistical inference and human expert evaluation.
Evaluated on three highly cited papers in data science modeling, the pipeline generated superior hypotheses compared to direct generation methods. Two highest-scoring hypotheses were used to develop novel algorithms that outperformed the baseline models in the original papers. Additionally, the workflow provides a theoretical framework encompassing theory-guided data science modeling methods and is extendable to other scientific domains.