Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning

TL;DR

AutoRefine is a reinforcement learning framework that improves retrieval-augmented reasoning by enabling iterative knowledge refinement during the reasoning process, leading to better performance on complex question-answering tasks.

Contribution

It introduces a novel search-and-refine-during-think paradigm with explicit knowledge filtering steps and tailored rewards, advancing retrieval-augmented reasoning methods.

Findings

Outperforms existing methods on QA benchmarks

Enhances reasoning accuracy in multi-hop scenarios

Issues higher-quality, frequent searches

Abstract

Large language models have demonstrated impressive reasoning capabilities but are inherently limited by their knowledge reservoir. Retrieval-augmented reasoning mitigates this limitation by allowing LLMs to query external resources, but existing methods often retrieve irrelevant or noisy information, hindering accurate reasoning. In this paper, we propose AutoRefine, a reinforcement learning post-training framework that adopts a new "search-and-refine-during-think" paradigm. AutoRefine introduces explicit knowledge refinement steps between successive search calls, enabling the model to iteratively filter, distill, and organize evidence before generating an answer. Furthermore, we incorporate tailored retrieval-specific rewards alongside answer correctness rewards using group relative policy optimization. Experiments on single-hop and multi-hop QA benchmarks demonstrate that AutoRefine significantly outperforms existing approaches, particularly in complex, multi-hop reasoning scenarios. Detailed analysis shows that AutoRefine issues frequent, higher-quality searches and synthesizes evidence effectively.