PRISM: Problem Discovery via Structural Motifs in Knowledge Graphs

While recent advancements in artificial intelligence have automated various scientific workflows, existing discovery systems primarily explore variations of user-specified inputs rather than autonomously identifying critical gaps in the broader literature. To overcome this reliance on predefined prompts, we propose a framework that formulates research problem discovery as the systematic detection of structural gaps within scientific knowledge graphs. Our approach decomposes scientific papers into structured components to build a heterogeneous graph, enabling structural-level reasoning across methods, evaluated tasks, and failure conditions. The system queries this graph for recurring inconsistencies and underexplored patterns, formulating these discoveries into structured problem objects. Through an evolutionary refinement process and localized subgraph reasoning, our framework generates novel, feasible research problems that demonstrate high traceability and robust grounding in existing scientific evidence.