CogNav: Cognitive Navigation for Complex Multi-Hop Retrieval via Chain-of-Thought Reasoning

Although retrieval-augmented generation (RAG) can effectively mitigate outdated knowledge and hallucinations in large language models, its retrieval performance on multi-hop reasoning tasks still faces substantial challenges. Most existing approaches construct evidence chains via sequential retrieval; however, in the absence of reliable reasoning guidance, the retrieval process is often inefficient and prone to error accumulation. Recent advances in the complex reasoning capabilities exhibited by large language models create new opportunities to address this issue. To this end, we propose CogNav, a cognition-navigation-based retrieval-augmented framework. The key idea is to actively plan and execute multi-step retrieval trajectories through structured chain-of-thought reasoning. CogNav formulates multi-hop retrieval as a state-aware sequential decision-making problem. With carefully designed prompt templates, the model is guided to progressively decompose complex questions into single-hop sub-queries, while dynamically determining when to terminate retrieval. Experiments across multiple multi-hop question answering benchmarks demonstrate that CogNav substantially outperforms conventional single-shot retrieval and existing progressive retrieval methods, yielding consistent gains in both retrieval sufficiency and answer accuracy. Overall, this work introduces a new paradigm for reasoning-guided retrieval systems and provides an effective pathway for evidence acquisition in complex, knowledge-intensive tasks.