ISO-RAG Improves Multi-Hop QA with Noise Control

Retrieval-Augmented Generation (RAG) systems are crucial for mitigating hallucinations in large language models (LLMs). However, traditional dense retrieval methods struggle with complex multi-hop question answering (QA) that requires navigating intricate reasoning paths. Graph-based RAG approaches can capture these relationships but often suffer from semantic drift and high latency due to noisy, global graph traversals. To overcome these limitations, researchers propose ISO-RAG (ISOperimetric Retrieval-Augmented Generation). This novel framework employs a geometry-aware approach by projecting the underlying knowledge graph into a hyperbolic Poincaré ball. This projection allows for the precomputation of node-wise isoperimetric profiles, which are then used to prune spurious edges during retrieval. By restricting the search space to a strictly localized subgraph, ISO-RAG effectively controls noise and regulates the Personalized PageRank diffusion process, ensuring exact and low-latency convergence. Experiments on multi-hop QA benchmarks demonstrate that ISO-RAG significantly outperforms state-of-the-art baselines, achieving substantial gains in both retrieval recall and downstream exact match accuracy, while also eliminating the latency bottleneck of global traversals.