BridgeRAG: A Framework for Reasoning over Partitioned Knowledge Graphs

ICLR 2026 Conference Submission25305 Authors

Keywords: RAG, Knowledge Graphs, Multi-hop Question Answering, Multi-Document Reasoning, LLM Agents, Planned Navigation

Abstract: Existing Knowledge Graph-based RAG (Retrieval-Augmented Generation) systems face a fundamental dilemma in multi-document scenarios. They either treat each document as an isolated knowledge graph, which preserves contextual purity but prevents cross-document reasoning, or merge them into a single, massive graph, leading to entity saturation and contextual noise pollution. To resolve this core conflict, we introduce the BridgeRAG framework, designed to elegantly achieve both "rtitioned isolation"and "ross-partition linking"or multiple documents. BridgeRAG is a collaborative framework that integrates static linking and dynamic reasoning. Experiments on multi-hop question answering benchmarks like HotpotQA show that BridgeRAG significantly outperforms state-of-the-art RAG models, especially on complex questions that require deep cross-partition navigation.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 25305