Particle.news
Download on the App Store

Two New RAG Preprints Propose Theme-Aligned Hypergraphs and Hierarchical Agents

The authors report gains over strong baselines, with claims posted as unreviewed preprints.

Overview

  • Cog-RAG introduces a dual-hypergraph design that models global themes across chunks alongside high‑order entity relations to better guide retrieval.
  • Its two-stage, top‑down strategy activates theme-level context first, then diffuses to entity details to align generation from broad topics to specifics.
  • The Cog-RAG paper reports significant improvements over state-of-the-art RAG baselines in experimental evaluations.
  • Separately, GHAR targets healthcare tasks with a dual-agent setup in which a top agent decides whether to retrieve and a lower agent compiles task-relevant evidence once retrieval is triggered.
  • GHAR formalizes agent–retriever–generator interaction as an MDP with tailored rewards and reports superior results across three datasets and three tasks.