Multimodal knowledge graphs (MMKGs) organize information across text, images, and other modalities to support multimodal Graph Retrieval-Augmented Generation. The usefulness of such systems depends heavily on the quality of the constructed graph, but most existing construction methods lack explicit quality guidance. Reinforcement learning approaches that optimize only for final-answer correctness provide indirect, sometimes weak, feedback about whether the graph itself is well built.

To close that gap, the authors propose MMKG-R1, a reinforcement learning framework that trains a multimodal large language model as a "multimodal knowledge constructor." Instead of relying solely on downstream outcomes, MMKG-R1 introduces multiple quality-aware rewards that directly assess both the information extracted from individual sources and the assembled graph structure. Those rewards guide the model's training.

Experiments across several datasets show that MMKG-R1 constructs multimodal knowledge graphs that improve downstream question-answering accuracy. The paper is the only source here, and its reported results are consistent: better graph quality, as defined by the proposed rewards, correlates with better performance on downstream tasks.