ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning

Paper · arXiv 2503.19470 · Published March 25, 2025
Reinforcement LearningDeep Research

We propose ReSearch, a novel framework that trains LLMs to Reason with Search via reinforcement learning without using any supervised data on reasoning steps. Our approach treats search operations as integral components of the reasoning chain, where when and how to perform searches is guided by text-based thinking, and search results subsequently influence further reasoning.

Reinforcement learning (RL) has emerged as a promising avenue for enhancing reasoning capabilities without the need for supervised data regarding reasoning steps [4, 16]. This approach holds potential for training LLMs to exhibit reasoning skills solely based on simple reward signals derived from final outcomes. Recent advancements in RL-based training for LLMs have demonstrated significant improvements in complex reasoning tasks, where models learn to decompose problems into manageable steps through trial and error rather than explicit instruction. Models such as DeepSeek-R1 have shown that rule-based reward functions can effectively guide LLMs to develop sophisticated reasoning patterns autonomously. Despite these successes, current approaches primarily focus on enhancing internal reasoning capabilities, with limited exploration of how to effectively combine this reasoning process with external knowledge retrieval.

In this paper, we propose a novel framework for training LLMs to Reason with Search via reinforcement learning, which we term ReSearch. The reasoning chain in this framework is not only composed of text-based thinking (i.e., enclosed by ) as DeepSeek-R1, but also search query (i.e., enclosed by search> /search>) and retrieval results (i.e., enclosed by ). We treat the search operation as part of the chain-like reasoning process, and the search operation will interact with text-based thinking. Specifically, when and how to perform search will be steered by previous text-based thinking and the search results will infuence subsequent text-based thinking. In the framework, we don’t provide any supervised data on reasoning steps for LLMs to imitate, instead, we leverage reinforcement learning (i.e., GRPO) to incentivize LLMs to perform reasoning with search.