Overview
- A coordinated set of arXiv preprints, published Tuesday, July 28, 2026, introduced new GraphRAG and RAG variants such as DeCoRAG, HCG‑RAG, VecTree‑RAG, HyCE‑RAG, MM‑R2, TriShieldRAG, and APS‑RAG that target multimodal and multi‑hop question answering.
- Several papers replace exhaustive flat extraction with structured representations like schema‑constrained causal graphs, hypergraphs, and tree navigation to cut graph size, reduce LLM calls, and make evidence easier for humans to audit.
- Agentic and intent‑grounded designs move reasoning before retrieval or split corpus routing from local traversal so systems request fewer tokens and yield clearer, more verifiable retrieval trajectories.
- Security and provenance defenses appear for the first time as integrated pipelines, with TriShieldRAG reporting a drop in a controlled PoisonedRAG attack from about 91% to about 13% on a 5,000‑document Wikipedia set while noting some validation sweeps remain unrun.
- One paper documents an operational deployment, APS‑RAG at the Advanced Photon Source, which fuses dense, sparse and graph channels and reports measurable gains on a 50‑question APS‑Bench, but most results are preprints that need peer review and broader replication.