Authors: ** Zhengliang Shi¹, Yiqun Chen², Haitao Li³, Weiwei Sun⁴, Shiyu Ni⁵, Yougang Lyu⁶, Run‑Ze Fan⁷, Bowen Jin⁸, Yixuan Weng⁹, Minjun Zhu⁹, Qiujie Xie⁹, Xinyu Guo¹⁰, Qu Yang¹¹, Jiayi Wu¹¹, Jujia Zhao¹², Xiaqiang Tang¹¹, Xinbei Ma¹¹, Cunxiang Wang³, Jiaxin Mao², Qingyao Ai³, Jen‑Tse Huang¹³, Wenxuan Wang², Yue Zhang⁹, Yiming Yang⁴, Zhaopeng Tu¹¹, Zhaochun Ren¹² ¹ Shandong University, ² Renmin University of China, ³ Tsinghua University, ⁴ Carnegie Mellon University, ⁵ UCAS, ⁶ University of Amsterdam, ⁷ University of Massachusetts Amherst, ⁸ University of Illinois Urbana‑Champaign, ⁹ Westlake University, ¹⁰ University of Arizona, ¹¹ Tencent, ¹² Leiden University, ¹³ Johns Hopkins University — **
📝 Abstract
Large language models (LLMs) have rapidly evolved from text generators into powerful problem solvers. Yet, many open tasks demand critical thinking, multi-source, and verifiable outputs, which are beyond single-shot prompting or standard retrieval-augmented generation. Recently, numerous studies have explored Deep Research (DR), which aims to combine the reasoning capabilities of LLMs with external tools, such as search engines, thereby empowering LLMs to act as research agents capable of completing complex, open-ended tasks. This survey presents a comprehensive and systematic overview of deep research systems, including a clear roadmap, foundational components, practical implementation techniques, important challenges, and future directions. Specifically, our main contributions are as follows: (i) we formalize a three-stage roadmap and distinguish deep research from related paradigms; (ii) we introduce four key components: query planning, information acquisition, memory management, and answer generation, each paired with fine-grained sub-taxonomies; (iii) we summarize optimization techniques, including prompting, supervised fine-tuning, and agentic reinforcement learning; and (iv) we consolidate evaluation criteria and open challenges, aiming to guide and facilitate future development. As the field of deep research continues to evolve rapidly, we are committed to continuously updating this survey to reflect the latest progress in this area.
💡 Deep Analysis
📄 Full Content
Deep Research: A Systematic Survey
Zhengliang Shi1 Yiqun Chen2 Haitao Li3 Weiwei Sun4 Shiyu Ni5 Yougang Lyu6
Run-Ze Fan7 Bowen Jin8 Yixuan Weng9 Minjun Zhu9 Qiujie Xie9 Xinyu Guo10 Qu Yang11
Jiayi Wu11 Jujia Zhao12 Xiaqiang Tang11 Xinbei Ma11 Cunxiang Wang3 Jiaxin Mao2
Qingyao Ai3 Jen-Tse Huang13 Wenxuan Wang2 Yue Zhang9 Yiming Yang4
Zhaopeng Tu11,
Zhaochun Ren12,
1Shandong University
2Renmin University of China
3Tsinghua University
4Carnegie Mellon University
5UCAS
6University of Amsterdam
7University of Massachusetts Amherst
8University of Illinois Urbana-Champaign
9Westlake University
10University of Arizona
11Tencent
12Leiden University
13Johns Hopkins University
Abstract: Large language models (LLMs) have rapidly evolved from text generators into powerful
problem solvers. Yet, many open tasks demand critical thinking, multi-source, and verifiable outputs,
which are beyond single-shot prompting or standard retrieval-augmented generation. Recently, nu-
merous studies have explored Deep Research (DR), which aims to combine the reasoning capabilities
of LLMs with external tools, such as search engines, thereby empowering LLMs to act as research
agents capable of completing complex, open-ended tasks. This survey presents a comprehensive and
systematic overview of deep research systems, including a clear roadmap, foundational components,
practical implementation techniques, important challenges, and future directions. Specifically, our
main contributions are as follows: (i) we formalize a three-stage roadmap and distinguish deep research
from related paradigms; (ii) we introduce four key components: query planning, information acquisi-
tion, memory management, and answer generation, each paired with fine-grained sub-taxonomies;
(iii) we summarize optimization techniques, including prompting, supervised fine-tuning, and agentic
reinforcement learning; and (iv) we consolidate evaluation criteria and open challenges, aiming to
guide and facilitate future development. As the field of deep research continues to evolve rapidly, we are
committed to continuously updating this survey to reflect the latest progress in this area.
Corresponding Author
Keywords: Deep Research, Large Language Models, Information Retrieval
Date: November 13, 2025
Code Repository: https://github.com/mangopy/Deep-Research-Survey
Contact: zhengliang.shii@gmail.com chenyiqun990321@ruc.edu.cn z.ren@liacs.leidenuniv.nl
arXiv:2512.02038v1 [cs.CL] 24 Nov 2025
Contents
1
Introduction
5
2
Preliminary Concept of Deep Research
6
2.1
What is Deep Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.2
Understanding Deep Research from Three Phases . . . . . . . . . . . . . . . . . . . .
6
2.3
Comparing Deep Research with RAG . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3
Key Components in Deep Research System
9
3.1
Query Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.1.1
Parallel Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.1.2
Sequential Planning
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
3.1.3
Tree-based Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.2
Information Acquisition
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3.2.1
Retrieval Tools
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3.2.2
Retrieval Timing
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.2.3
Information Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
3.3
Memory Management
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.3.1
Memory Consolidation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
3.3.2
Memory Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3.3.3
Memory Updating
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3.3.4
Memory Forgetting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.4
Answer Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.4.1
Integrating Upstream Information . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.4.2
Synthesizing Evidence and Maintaining Coherence . . . . . . . . . . . . . . .
25
3.4.3
Structuring Reasoning and Narrative . . . . . . . . . . . . . . . . . . . . . . .
26
3.4.4
Presentation Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
4
Practical Techniques for Optimizing Deep Research Systems
27
4.1
Workflow Prompt Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
4.1.1
Deep Research System of Anthropic . . . . . . . . . . . . . . . . . . . . . . . .
28
4.2
Supervised Fine-Tuning
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
2
4.2.1
Strong-to-weak Distillation
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4.2.2
Ite
