SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis

TL;DR

<3-5 sentence high-level summary> SimpleDeepSearcher tackles the data bottleneck and computational cost of deep search by replacing RL with strategically engineered supervised fine-tuning trained on real-web, multi-turn reasoning trajectories. It combines realistic data synthesis in live web environments, a diversity-aware query sampling strategy, and a four-dimensional response curation pipeline to produce 871 high-quality training examples. The approach yields substantial gains over RL-based baselines across five benchmarks and demonstrates strong generalization across model backbones and domains. The work highlights data-centric pathways to efficient deep search and provides practical guidance for building real-world, search-enabled LLMs.

Abstract

Retrieval-augmented generation (RAG) systems have advanced large language models (LLMs) in complex deep search scenarios requiring multi-step reasoning and iterative information retrieval. However, existing approaches face critical limitations that lack high-quality training trajectories or suffer from the distributional mismatches in simulated environments and prohibitive computational costs for real-world deployment. This paper introduces SimpleDeepSearcher, a lightweight yet effective framework that bridges this gap through strategic data engineering rather than complex training paradigms. Our approach synthesizes high-quality training data by simulating realistic user interactions in live web search environments, coupled with a multi-criteria curation strategy that optimizes the diversity and quality of input and output side. Experiments on five benchmarks across diverse domains demonstrate that SFT on only 871 curated samples yields significant improvements over RL-based baselines. Our work establishes SFT as a viable pathway by systematically addressing the data-scarce bottleneck, offering practical insights for efficient deep search systems. Our code is available at https://github.com/RUCAIBox/SimpleDeepSearcher.

Figures (8)

• Figure 1: Overall framework of our proposed SimpleDeepSearcher approach. $r$ denotes the reasoning content, $q$ represents the search query, and $d$ refers to the retrieved document after summarization. $t_s$ and $t_e$ are special tokens indicating the beginning and end of the search query, and $a$ denotes the final answer.
• Figure 2: Average reasoning length across different benchmarks w/ and w/o reasoning data for training.
• Figure 3: Domain distribution of the data before filtering.
• Figure 4: Changes in Sequence Length and Reward During REINFORCE++ Training.
• Figure :