Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

TL;DR

Search-R1 introduces a reinforcement learning framework enabling large language models to autonomously generate search queries during reasoning, significantly improving question-answering performance by integrating real-time retrieval with stable training techniques.

Contribution

It presents a novel RL-based method for LLMs to learn optimal search interactions during reasoning, enhancing retrieval-augmented question answering.

Findings

Improves QA performance by up to 41% over baselines.

Demonstrates effective RL optimization for search-query generation.

Provides empirical insights into retrieval-augmented reasoning dynamics.

Abstract

Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities to use search engines during inference is often suboptimal, as the LLM might not fully possess the capability on how to interact optimally with the search engine. This paper introduces Search-R1, an extension of reinforcement learning (RL) for reasoning frameworks where the LLM learns to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM reasoning trajectories with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 41% (Qwen2.5-7B) and 20% (Qwen2.5-3B) over various RAG baselines under the same setting. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.