LOGOS: Precision Retrieval via Logical Document Graphs for Retrieval-Augmented Generation

ICLR 2026 Conference Submission19048 Authors

Keywords: Graph Neural Network, RAG

TL;DR: Enhancing Contextual Understanding in Long Documents via Cross-Page Heterogeneous GNN

Abstract: Retrieval-Augmented Generation (RAG) systems struggle with long documents because conventional retrieval methods provide noisy, page-level context that degrades generation quality. These methods are fundamentally limited by treating documents as a linear sequence of pages, which breaks the crucial logical dependencies—like tables, paragraphs, or references—that span across page boundaries. To overcome this limitation, we propose Precision Retrieval via Logical Document Graphs for Retrieval-Augmented Generation (LOGOS), a new RAG method that achieves precision retrieval by modeling a document's intrinsic logical structure. LOGOS transforms a document into a graph where semantic regions are nodes and logical connections are edges, effectively bridging page breaks. A Graph Neural Network then generates fine-grained, context-aware representations for each node, enabling a more concise and semantically relevant context for the generator. Extensive experiments on the ViDoRe and MMDOCIR benchmarks show that LOGOS sets a new state-of-the-art, significantly outperforming strong baselines by up to 2\% in average Recall@1.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 19048