MedEviGraph: Evidence Organization and Paired Responses in Multimodal Medical Reasoning

Biomedical reasoning workflows need evidence organization that supports inference and answer revision that corrects errors while preserving sound judgments. We introduce MedEviGraph, a benchmark of 9,559 chest-radiography, CT, and histology cases whose typed graphs connect images with annotation-derived observations. Shared evidence objects support comparisons of input structure and paired answers after retaining or removing cited cards. Across nine configurations on 1,451 test cases, explicit relations improve accuracy in seven configurations with observations fixed, adding up to 125 correct decisions. Analysis of 30,035 response triplets further distinguishes systems with identical initial accuracy: their retention correction rates differ by 13.73 percentage points, alongside different rates of newly introduced errors. Fuller observations improve all nine configurations, with the largest gains in histology. MedEviGraph connects evidence organization and subsequent answer quality to support the evaluation of reasoning components for biomedical assistants.