Large Language Models as Bayesian Priors for Interpretable Scientific Discovery

Zhang et al. (2024, LLMExplainer) show how LLMs can serve as a Bayesian inference module for graph explanations, but the broader potential remains untapped. Here, the idea is to formalize how LLMs—pretrained on vast scientific literature—can provide nuanced, structured priors for Bayesian analysis in fields like biology (Han, 2025), physics, or medicine, where data may be scarce but domain knowledge is rich. The model would use LLM-generated priors to guide hypothesis generation and posterior updating, ensuring that the results are both data-driven and consistent with expert knowledge. The innovation is in making LLMs not just black-box predictors, but interpretable, updatable Bayesian partners in the scientific discovery process—potentially accelerating breakthroughs in understudied or low-data scientific areas.

References:

1. Inter-Discipline Between Mathematical Modeling, Statistical Inference and Biology. Yuliang Han (2025). Theoretical and Natural Science.
2. LLMExplainer: Large Language Model based Bayesian Inference for Graph Explanation Generation. Jiaxing Zhang, Jiayi Liu, Dongsheng Luo, Jennifer Neville, Hua Wei (2024). arXiv.org.