To what extent can large language models assess the reliability of published works in biomedical sciences

• Grant number: 25/03746-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: May 01, 2025
• End date: April 30, 2026
• Field of knowledge: Health Sciences - Medicine
• Principal Investigator: Konradin Metze
• Grantee: Matheus da Silva Fago
• Host Institution: Faculdade de Ciências Médicas (FCM). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

The number of retracted scientific papers is increasing considerably across science in general. Obviously, retracted papers cannot serve as a basis for designing new scientific studies or interpreting data. Unfortunately, the retraction of a document may occur long after its publication, and many researchers are often unaware of the retraction. Therefore, it is mandatory for scientists to frequently check the validity of the articles in their personal scientific libraries. However, this process is time-consuming and tedious, and as a result, it is not conducted rigorously. In this context, the use of LLMs (large language models) could provide valuable assistance. However, previous studies have shown that these systems exhibit significantly more errors than expected based on their advertised capabilities. We aim to test the reliability of LLMs in assessing the supposed retraction of scientific papers by creating a dataset of 500 real bibliographic references. Half of them represent publications that were later retracted, while the other 250 are non-retracted articles with titles also present in the retracted group. We will compare the performance of different LLMs based on the percentage of retracted documents that were not recognized as such and the percentage of non-retracted papers that were falsely identified as retracted. These analyses will be repeated after 3 and 6 months to detect potential improvements or deteriorations in LLM performance.