Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Web Search towards Agentic Deep Research: Incentivizing Search with Reasoning Agents

Weizhi Zhang, Yangning Li, Yuanchen Bei, Junyu Luo, Guancheng Wan, Liangwei Yang, Chenxuan Xie, Yuyao Yang, Wei-Chieh Huang, Chunyu Miao, Henry Peng Zou, Xiao Luo, Yusheng Zhao, Yankai Chen, Chunkit Chan, Peilin Zhou, Xinyang Zhang, Chenwei Zhang, Jingbo Shang, Ming Zhang, Yangqiu Song, Irwin King, Philip S. Yu

TL;DR

The paper argues that traditional web search is inadequate for complex, multi-step information needs and proposes Agentic Deep Research, where LLMs with reasoning and agentic capabilities autonomously plan, search, and synthesize. It introduces a test-time scaling law linking inference-time computation to gains in internal reasoning and external search, and supports the claim with benchmarks and open-source momentum. The work outlines prompting, supervised fine-tuning, and reinforcement learning as pathways to incentivize search, and discusses open problems such as trust, domain-specific deep research, and multi-modality. Collectively, it positions agentic deep research as the dominant, human-centric paradigm for future information seeking and knowledge synthesis, while highlighting the need for human oversight and responsible deployment.

Abstract

Information retrieval is a cornerstone of modern knowledge acquisition, enabling billions of queries each day across diverse domains. However, traditional keyword-based search engines are increasingly inadequate for handling complex, multi-step information needs. Our position is that Large Language Models (LLMs), endowed with reasoning and agentic capabilities, are ushering in a new paradigm termed Agentic Deep Research. These systems transcend conventional information search techniques by tightly integrating autonomous reasoning, iterative retrieval, and information synthesis into a dynamic feedback loop. We trace the evolution from static web search to interactive, agent-based systems that plan, explore, and learn. We also introduce a test-time scaling law to formalize the impact of computational depth on reasoning and search. Supported by benchmark results and the rise of open-source implementations, we demonstrate that Agentic Deep Research not only significantly outperforms existing approaches, but is also poised to become the dominant paradigm for future information seeking. All the related resources, including industry products, research papers, benchmark datasets, and open-source implementations, are collected for the community in https://github.com/DavidZWZ/Awesome-Deep-Research.

From Web Search towards Agentic Deep Research: Incentivizing Search with Reasoning Agents

TL;DR

The paper argues that traditional web search is inadequate for complex, multi-step information needs and proposes Agentic Deep Research, where LLMs with reasoning and agentic capabilities autonomously plan, search, and synthesize. It introduces a test-time scaling law linking inference-time computation to gains in internal reasoning and external search, and supports the claim with benchmarks and open-source momentum. The work outlines prompting, supervised fine-tuning, and reinforcement learning as pathways to incentivize search, and discusses open problems such as trust, domain-specific deep research, and multi-modality. Collectively, it positions agentic deep research as the dominant, human-centric paradigm for future information seeking and knowledge synthesis, while highlighting the need for human oversight and responsible deployment.

Abstract

Information retrieval is a cornerstone of modern knowledge acquisition, enabling billions of queries each day across diverse domains. However, traditional keyword-based search engines are increasingly inadequate for handling complex, multi-step information needs. Our position is that Large Language Models (LLMs), endowed with reasoning and agentic capabilities, are ushering in a new paradigm termed Agentic Deep Research. These systems transcend conventional information search techniques by tightly integrating autonomous reasoning, iterative retrieval, and information synthesis into a dynamic feedback loop. We trace the evolution from static web search to interactive, agent-based systems that plan, explore, and learn. We also introduce a test-time scaling law to formalize the impact of computational depth on reasoning and search. Supported by benchmark results and the rise of open-source implementations, we demonstrate that Agentic Deep Research not only significantly outperforms existing approaches, but is also poised to become the dominant paradigm for future information seeking. All the related resources, including industry products, research papers, benchmark datasets, and open-source implementations, are collected for the community in https://github.com/DavidZWZ/Awesome-Deep-Research.

Paper Structure

This paper contains 32 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Traditional Information Search Paradigms
  3. Web Search
  4. LLMs as Chatbots
  5. LLMs with RAG
  6. LLMs with RAG
  7. Towards Agentic Deep Research
  8. Incentivizing Search with Reasoning Agents
  9. Prompting and In-Context Learning: Bridging Search and Reasoning.
  10. Supervised Fine-Tuning: Hard-Coding Search Patterns.
  11. Reinforcement Learning: Optimizing Reasoning-Driven Search in the Wild.
  12. Benchmarks and Open-Source Implementations:
  13. Deep Research Benchmarks
  14. Open-Source Implementations
  15. Test-Time Scaling Law for Deep Research
  16. ...and 17 more sections

Figures (3)

  • Figure 1: The evolution of information search paradigms.
  • Figure 2: Benchmarks of 5 standard LLMs, 4 reasoning LLMs, and 1 agentic deep research model (OpenAI Deep Research) on BrowseComp, BrowseComp-ZH, and Humanity's Last Exam.
  • Figure 3: (a) Open-source star trend for agentic deep research. (star counts recorded up to 22 May 2025). (b) Test-time scaling (TTS) law for agentic deep research, where data before performance normalization for reasoning TTS is from S1 on AIME24 muennighoff2025s1 and for search TTS is from IterDRAG on MuSiQue yueinference. The TTS is measured in inference tokens.