NaviRAG: Towards Active Knowledge Navigation for Retrieval-Augmented Generation

Jihao Dai^1,2, Dingjun Wu¹, Yuxuan Chen¹, Zheni Zeng², Yukun Yan¹, Zhenghao Liu³, Maosong Sun¹,

¹Tsinghua University, ²Nanjing University, ³Northeastern University

📖 Introduction/Overview

NaviRAG is a navigation-based Retrieval-Augmented Generation (RAG) framework designed for complex reasoning question answering. Existing RAG research primarily focuses on cross-document retrieval and multi-hop information integration, approximating reasoning as the localization and aggregation of dispersed facts. However, in complex long-chain reasoning scenarios, queries are constrained by explicit contextual conditions, and the required evidence is distributed across different semantic levels of a text. The relationship between evidence and queries is jointly governed by contextual semantics, thereby imposing higher demands on retrieval mechanisms.

Inspired by Information Foraging Theory, NaviRAG models evidence acquisition as a multi-stage, navigable, and dynamic exploration process. The framework constructs a hierarchical semantic representation grounded in traceable raw text and adopts a staged retrieval strategy of “locate first, then forage.” It first identifies relevant semantic subspaces within the knowledge base and subsequently performs coarse-to-fine, multi-step navigational retrieval to progressively acquire evidence. This design enables efficient adaptation to queries of varying granularity while supporting context-sensitive retrieval.

Extensive experiments on multiple complex reasoning question answering benchmarks demonstrate that NaviRAG achieves significant performance improvements over mainstream RAG methods while maintaining competitive reasoning costs.

🚧 Code Availability

The codebase is currently under active organization. It will be released and documented in this repository as soon as it is ready.

🥰 Citation

@misc{dai2026naviragactiveknowledgenavigation,
      title={NaviRAG: Towards Active Knowledge Navigation for Retrieval-Augmented Generation}, 
      author={Jihao Dai and Dingjun Wu and Yuxuan Chen and Zheni Zeng and Yukun Yan and Zhenghao Liu and Maosong Sun},
      year={2026},
      eprint={2604.12766},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.12766}, 
}