OmniScience++: A Hierarchical Corpus for Evidence-Grounded Scientific Reasoning

Scientific conclusions often depend on observations spread across multiple panels, measurements, and experimental conditions. Understanding these figures requires identifying the relevant evidence and determining how the observations support a conclusion. We introduce OmniScience++, a hierarchical corpus for evidence-grounded scientific reasoning, organizing 1.52 million figures into a figure–panel–subfigure structure with 10.02 million localized atomic-subfigure records. An evidence-first data engine turns this corpus into questions with explicit visual support. The engine selects observations and composes a reasoning program before generating the question and answer. Constraints on entities, conditions, and measurements guide evidence composition, while source regions and intermediate operations remain traceable. This construction supports scientific tasks such as comparing experimental outcomes and using an intermediate finding to select evidence for a subsequent inference. We evaluate these capabilities with a 600-question human-reviewed benchmark covering quantitative analysis, visual structure, and experimental relations. Across eight multimodal models, answer accuracy and evidence localization produce different rankings. Providing reference evidence crops improves answer accuracy by 2.36-17.70 percentage points across four models. By enabling the construction and evaluation of scientific reasoning tasks with explicit visual support, OmniScience++ provides a foundation for future AI for scientific research that connects conclusions to their underlying experimental evidence.