Title: Prioritizing Significance: Grounding Contextual Knowledge via Dynamic Indices for Sequential Retrieval

Abstract: Retrieval-Augmented Generation (RAG) has emerged as a leading strategy to reduce hallucinations in Large Language Models (LLMs) by bringing in external information. Nevertheless, current RAG architectures frequently encounter difficulties in synthesizing and reasoning about critical proof points that are fragmented across cluttered retrieved texts, a challenge that is especially pronounced in multi-hop contexts. To address this, we introduce KAIR, a framework designed for Iterative Retrieval that grounds knowledge within the retrieved material to steer LLMs toward essential details. As the iterative retrieval process unfolds, KAIR continuously refines the knowledge index to secure prominent evidence extracted from the documents. This dynamic index functions as a navigational anchor, allowing the LLM to evaluate whether the available knowledge is sufficient and to construct subsequent queries for retrieval. Ultimately, KAIR produces responses by combining the gathered documents with the finalized anchoring index. Our evaluations across four multi-hop question answering benchmarks reveal that KAIR consistently surpasses robust RAG baseline methods. Additional analysis confirms that KAIR successfully isolates vital knowledge and reduces contextual interference during iterative retrieval, thereby enhancing the LLM’s capacity to link and reason over scattered evidence found in the retrieved documents. The associated code and data can be accessed at https://github.com/NEUIR/KAIR.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC