Adaptive Retrieval for Reasoning-Intensive Retrieval

TL;DR

REPAIR tackles reasoning-intensive retrieval by bridging reasoning-aware planning with selective adaptive retrieval. It introduces a Planning-based Step Reranker (PSR) that outputs an explicit reasoning plan and a dense step-level reward system, and a Step-Adaptive NAR (SAR) that uses these rewards to perform mid-course retrieval corrections. Through dense rewards (including a Baseline reward and a Consistency reward via a Bradley–Terry model), REPAIR directs adaptive retrieval to pivotal steps, achieving improved bridge-document coverage and recall. Empirical results on BRIGHT and multi-hop QA benchmarks show consistent gains over strong reranking and naive NAR baselines, with favorable efficiency characteristics and demonstrated synergy between PSR and SAR.

Abstract

We study leveraging adaptive retrieval to ensure sufficient "bridge" documents are retrieved for reasoning-intensive retrieval. Bridge documents are those that contribute to the reasoning process yet are not directly relevant to the initial query. While existing reasoning-based reranker pipelines attempt to surface these documents in ranking, they suffer from bounded recall. Naive solution with adaptive retrieval into these pipelines often leads to planning error propagation. To address this, we propose REPAIR, a framework that bridges this gap by repurposing reasoning plans as dense feedback signals for adaptive retrieval. Our key distinction is enabling mid-course correction during reranking through selective adaptive retrieval, retrieving documents that support the pivotal plan. Experimental results on reasoning-intensive retrieval and complex QA tasks demonstrate that our method outperforms existing baselines by 5.6%pt.

Figures (4)

• Figure 1: A comparison of reranking and adaptive retrieval paradigms. (a) Reasoning-based reranking first generates a reasoning trace and ranks documents based on the trace, but is limited by bounded recall since it operates on a fixed candidate set. (b) NAR addresses bounded recall by expanding candidates through document neighborhoods, but retrieval is guided solely by top-ranked documents, without awareness of reasoning steps. (c) REPAIR tightly integrates reranking and adaptive retrieval by transforming planning steps into dense, step-level rewards and selectively performing adaptive retrieval for mid-course correction.
• Figure 2: Comparison between retrievers on BRIGHT. $i=1$ corresponds to the first-stage retrieval (BM25). For (Ours) PSR w/ SAR, we used BM25 ranking results up to $i<6$ (See Figure \ref{['fig:consistency']} for related discussion.).
• Figure 3: Compatibility comparison between different rerankers and NAR on BRIGHT benchmark.
• Figure 4: nDCG@10 vs. accumulated iterations used for the Bradley-Terry (BT) model on BRIGHT.