DIVER: A Multi-Stage Approach for Reasoning-intensive Information Retrieval

TL;DR

DIVER addresses the challenge of reasoning-intensive information retrieval by a four-stage pipeline that preprocesses documents, performs iterative query expansion with explicit reasoning, uses a synthetic-data trained dense retriever, and applies a hybrid pointwise-listwise reranking strategy. Key innovations include DIVER-DChunk for long-document handling, two-round DIVER-QExpand with a reasoning-based expansion loop, a Qwen3-Embedding-4B backbone trained with InfoNCE on diverse synthetic data, and a combined reranking scheme that leverages both local and global relevance signals. On the BRIGHT benchmark, DIVER achieves state-of-the-art $nDCG@10$ scores (e.g., $46.8$ overall and $31.9$ on original queries in early reports), underscoring the effectiveness of reasoning-aware retrieval and iterative query refinement for complex tasks. The work demonstrates that integrating explicit reasoning into query expansion and retriever training yields meaningful gains over strong baselines, with practical implications for real-world knowledge-intensive search tasks.

Abstract

Retrieval-augmented generation has achieved strong performance on knowledge-intensive tasks where query-document relevance can be identified through direct lexical or semantic matches. However, many real-world queries involve abstract reasoning, analogical thinking, or multi-step inference, which existing retrievers often struggle to capture. To address this challenge, we present DIVER, a retrieval pipeline designed for reasoning-intensive information retrieval. It consists of four components. The document preprocessing stage enhances readability and preserves content by cleaning noisy texts and segmenting long documents. The query expansion stage leverages large language models to iteratively refine user queries with explicit reasoning and evidence from retrieved documents. The retrieval stage employs a model fine-tuned on synthetic data spanning medical and mathematical domains, along with hard negatives, enabling effective handling of reasoning-intensive queries. Finally, the reranking stage combines pointwise and listwise strategies to produce both fine-grained and globally consistent rankings. On the BRIGHT benchmark, DIVER achieves state-of-the-art nDCG@10 scores of 46.8 overall and 31.9 on original queries, consistently outperforming competitive reasoning-aware models. These results demonstrate the effectiveness of reasoning-aware retrieval strategies in complex real-world tasks.

Figures (3)

• Figure 1: DIVER achieves state-of-the-art performance on BRIGHT benchmark.
• Figure 2: Overview of DIVER pipeline. The DIVER pipeline begins with document cleaning and semantic-based rechunking to improve textual coherence. User queries are then iteratively expanded using the DIVER-QExpand module to enhance their expressiveness. Document relevance is scored by both the DIVER Retriever and BM25 Retriever. The top-100 candidates are reranked using DIVER-Rerank, which assigns LLM-based helpfulness scores. Final rankings are obtained by interpolating pointwise and listwise reranking scores to improve overall precision. To ensure fair comparison with other baselines, DIVER-DChunk is excluded from the main experiments and only evaluated separately in \ref{['sec:ablation_diver_dchunk']}.
• Figure 3: The training process of DIVER-Retriever.