BioDivergence: A Benchmark and Evaluation Framework for Hidden Contextual Contradictions in Biomedical Abstracts

BioDivergence is a newly introduced evaluation framework designed to address hidden contextual contradictions in biomedical abstracts, featuring a six-class conflict taxonomy and a 13-axis divergence ontology. The framework includes BioDivergence-Silver-v1.0, a benchmark consisting of 11,865 article-disjoint claim pairs across five biomedical domains, revealing significant performance variations with models like Mistral-7B-Instruct-v0.3 achieving 0.5523 accuracy and 0.3894 contextual-F1 on the primary test set. This framework is crucial for practitioners as it enhances the understanding of contextual divergence, thereby improving the reliability of model outputs in biomedical research.