GraphLoom is a new framework for multimodal knowledge-graph retrieval-augmented generation (RAG) systems. It aims to improve the reliability and faithfulness of evidence routing by selectively retrieving high-utility evidence from a graph memory. GraphLoom constructs an instance-level multimodal knowledge graph from various sources, including scene descriptions, relational triples, and external commonsense knowledge. The framework uses hierarchical graph memory slots and joint graph-sequence attention in a frozen language model to route evidence. Experimental results show that GraphLoom outperforms strong baselines on several benchmarks, including ScienceQA, MultiModalQA, and OK-VQA. --- Why it matters: GraphLoom's reliability-calibrated graph evidence routing is relevant to researchers working on multimodal knowledge-graph RAG systems because it addresses the issue of noisy evidence and unsupported generation in these systems. By selectively retrieving high-utility evidence, GraphLoom can improve the quality of answers and evidence faithfulness. Source: https://arxiv.org/abs/2608.15056