Update: Our research on evaluating large language models' capacity to predict neuroscience findings, including BrainGPT, has been published in Nature Human Behaviour. - https://www.nature.com/articles/s41562-024-02046-9

The scientific literature is exponentially increasing in size. One challenge for scientists is keeping abreast of developments. One solution is a human-machine teaming approach in which scientists interact with a vast knowledge base of the neuroscience literature, referred to as BrainGPT. BrainGPT is trained to capture data patterns in the neuroscience literature, taking advantage of recent machine learning advances in large-language models.

BrainGPT functions as a generative model of the scientific literature, allowing researchers to propose study designs as prompts for which BrainGPT would generate likely data patterns reflecting its current synthesis of the scientific literature. Modellers can use BrainGPT to assess their models against the field's general understanding of a domain (e.g., instant meta-analysis). BrainGPT could help identify anomalous findings, whether because they point to a breakthrough or contain an error.

Importantly, BrainGPT does not summarize papers nor retrieve articles. In such cases, large-language models often confabulate, which is potentially harmful. Instead, BrainGPT stitches together existing knowledge too vast for human comprehension to assist humans in expanding scientific frontiers.