GraphLoom: Reliability-Calibrated Graph Evidence Routing for Multimodal KG-RAG

A new study available on arXiv (2608.15056) introduces GraphLoom, a framework aimed at improving the reliability and precision of outputs in multimodal retrieval-augmented generation (RAG). It creates a detailed multimodal knowledge graph at the instance level by using scene descriptions, relational triples, and external commonsense knowledge. Instead of using all retrieved information in the generation process, GraphLoom focuses on a reliability-aware subgraph retrieval method. This allows it to prioritize valuable evidence through a structured graph memory and joint attention within a fixed language model. The authors point out that this approach reduces irrelevant information and enhances multi-hop reasoning, setting it apart from traditional RAG systems that often rely on lengthy or overly complex evidence.