Source-linked AI summary
A Survey of LLM-based Deep Search Agents: Paradigm, Optimization, Evaluation, and Challenges
Yunjia Xi, Jianghao Lin, Yongzhao Xiao, Zheli Zhou, Rong Shan, Te Gao, Jiachen Zhu, Weiwen Liu, Yong Yu, Weinan Zhang
TL;DR
Search agents address the difficulty of using static or single-turn retrieval for complex, dynamic information needs. This survey provides the first systematic analysis of their search structures, optimization methods, applications, evaluations, and challenges. It identifies a shift toward more dynamic search and highlights gaps in current evaluation and coverage of commercial deployments.
Problem
Static single-turn or rule-based iterative search struggles with complex and dynamic context, motivating autonomous search agents.
Method
The survey systematically analyzes search agents across search structure, optimization, application, evaluation, and challenges.
Results
Search structures are shifting toward dynamic querying, with hybrid structures often preferred without tuning and sequential flexibility increasingly internalized by fine-tuned models.
Takeaways & Limitations
Search Agents proactively leverage context and diverse sources, extending information retrieval into a broader, intelligent process for deep information mining.
Takeaways & Limitations
The survey focuses primarily on academic papers and therefore covers commercial applications less extensively because technical details are often undisclosed.
Abstract
from arXiv · showhide
The advent of Large Language Models (LLMs) has significantly revolutionized web search. The emergence of LLM-based Search Agents marks a pivotal shift towards deeper, dynamic, autonomous information seeking. These agents can comprehend user intentions and environmental context and execute multi-turn retrieval with dynamic planning, extending search capabilities far beyond the web. Leading examples like OpenAI's Deep Research highlight their potential for deep information mining and real-world applications. This survey provides the first systematic analysis of search agents. We comprehensively analyze and categorize existing works from the perspectives of architecture, optimization, application, and evaluation, ultimately identifying critical open challenges and outlining promising future research directions in this rapidly evolving field. Our repository is available on https://github.com/YunjiaXi/Awesome-Search-Agent-Papers.
1 Introduction
Search has evolved from manually consolidating web results to LLM-enhanced retrieval and autonomous Search Agents. The survey systematically examines this emerging paradigm across its holistic pipeline, including search, optimization, application, evaluation, and challenges.
- Traditional Web Search required users to manually select and consolidate relevant information from result lists.
- LLM-enhanced Search introduced query rewriting, but its single-turn or rule-based iterative integration struggles with complex and dynamic context.
- Search Agents autonomously control search, using user intentions and environmental context for adaptive reasoning and dynamic retrieval across diverse information sources.
- The survey presents a systematic analysis organized around how to search, optimize, apply, and evaluate Search Agents.
- Unlike surveys centered on specific sub-domains or perspectives, this work analyzes the holistic pipeline, including search structure, optimization, application, evaluation, and challenges.
2 Task Formulation
A Search Agent receives a user intention and context, then iteratively plans and acts to gather information. It updates its plan through reflection on observations and prior trajectory until sufficient information is acquired.
- A search agent iteratively plans and acts to gather information and fulfill a user’s intention given intention q and context C.
- The agent begins with π0 = Plan(q, C), then updates its plan through reflection on observations and prior trajectory.
- After planning, the agent acts by searching or browsing, producing retrieved information as a new observation for subsequent steps.
- The process continues until sufficient information is acquired, forming an information-seeking trajectory.
3 How to Search
Search Agents scale beyond single queries through parallel, sequential, and hybrid structures. The field is shifting toward dynamic, context-generated search, with hybrid structures favored without tuning and sequential flexibility increasingly internalized by fine-tuned models.
- Search-agent querying has evolved from single queries toward dynamic, context-dependent queries guided by search structures and feedback.
- Parallel Structure: Parallel search reformulates one query into multiple queries processed simultaneously, including decomposition-based and diversification-based strategies.
- Parallel Structure: Decomposition-based parallel search breaks complex intents into sub-queries, executes them in parallel, and synthesizes the results.
- Sequential Structure: Sequential search dynamically decides whether and what to search next using prior results and reflection, unlike advance-fixed parallel querying.
- Hybrid Structure: Hybrid structures combine parallel and sequential exploration through tree- or graph-based designs, enabling multiple paths and, for graphs, backtracking or revised decisions.
- Hybrid Structure: Hybrid structures are often preferred in tuning-free settings, while fine-tuned models tend to internalize flexibility within sequential structures for efficiency and effectiveness.
4 How to Optimize
Search-agent optimization divides into tuning-free and tuning-based approaches. Tuning-free systems use human-designed workflows, multi-agent organization, and test-time computation, whereas tuning-based methods learn search behavior through supervised fine-tuning, reinforcement learning, or their combination.
- Optimization methods are broadly categorized into tuning-free and tuning-based approaches.
- Tuning-Free Approaches: Tuning-free approaches rely on human knowledge and predefined workflows, increasingly using multi-agent architectures and test-time scaling.
- Tuning-Free Approaches: Single-agent systems handle planning, query generation, and synthesis, commonly using iterative refinement or reflection chains to structure complex search.
- Tuning-Free Approaches: Multi-agent systems distribute complex search among planner, search, and generation agents, using fixed execution orders or centralized supervision.
- Tuning-Free Approaches: Test-time scaling allocates additional inference computation through reasoning-centric and search-centric techniques such as Self-Consistency, Best-of-N, and MCTS.
- Tuning-Based Approaches: Tuning-based methods train agents to learn context-dependent actions through imitation and exploration.
- Tuning-Based Approaches: Supervised fine-tuning internalizes reasoning and search capabilities from high-quality trajectories or actions, either in components or end to end.
- Tuning-Based Approaches: Reinforcement learning supports flexible behavior through environmental exploration and increasingly targets end-to-end planning, searching, reflection, and generation.
5 How to Apply
Search Agents apply agentic information seeking both externally across diverse domains and internally to improve reasoning, memory navigation, and tool selection.
- External Applications: Search Agents conduct in-depth information seeking across diverse external domains, extending retrieval beyond traditional web search.Applications include AI assistants, e-commerce, finance, code, medicine, biology, chemistry, teaching, and research.
- External Applications: Deep Research systems search diverse sources, synthesize large amounts of information, and present organized findings.
- Internal Agent Enhancements: Agentic search improves tool selection through multi-turn reasoning and search when agents have access to many tools.
- Internal Agent Enhancements: Agentic search extracts queries from ambiguous user intentions and searches extensive agent memory for relevant content.
- Internal Agent Enhancements: Agents can retrieve relevant internal experiences and combine them with external knowledge to support more robust reasoning.
6 How to Evaluate
Search Agent evaluation spans complex QA, open-ended research, and domain-specific datasets, using task outcomes, process quality, and multifaceted judgments. Open-ended evaluation is especially demanding because answers lack a single correct response.
- Datasets: Evaluation datasets primarily cover complex QA and reasoning problems requiring extensive information seeking.
- Datasets: Closed-ended evaluation includes multi-hop QA, challenging QA, and fact-checking because simple datasets are insufficient for multi-turn search.
- Datasets: Open-ended datasets assess broad, non-factual, expert-level, and sometimes multimodal research questions without a single unambiguous answer.
- Datasets: Domain-specific benchmarks evaluate iterative information seeking in fields including finance, business, medicine, and agriculture.
- Metrics and Judgment: Closed-ended and domain-specific tasks mainly use final-output metrics such as Exact Match (EM), F1, Accuracy, and Pass@k, while some evaluate reasoning and retrieval.
- Metrics and Judgment: Open-ended tasks require metrics covering key-point coverage, informativeness, breadth, depth, coherence, organization, readability, citation accuracy, and win rate.
- Metrics and Judgment: Judging has evolved from rule-based metrics toward LLM- and Agent-based paradigms, while future frameworks should assess citation reliability and credibility discrimination.
7 Challenges and Future Directions
Future Search Agents must address broader information fusion, imperfect retrieval, multimodality, customized reinforcement learning, infrastructure, scheduling, and autonomous self-evolution.
- Broaden and Fuse Information Sources: Search Agents must integrate private and proprietary data with heterogeneous external sources while reconciling conflicting facts and perspectives.
- Imperfect Retrieval: Imperfect retrieval requires agents to detect noise, bias, and falsehoods through stronger verification, fact-checking, and critical assessment.
- From Text to Multi-Modality: Moving beyond text requires multimodal search infrastructure and reasoning across text, images, audio, and video.
- Customized Reinforcement Learning: Search-specific reinforcement learning must handle long-horizon planning, imperfect feedback, sparse rewards, and open-ended queries.
- Robust Infrastructure: Robust infrastructure should improve computationally intensive RL sampling, high-recall retrieval, and support for complex concurrent search tasks.
- Search Agent Self-Evolution: Self-evolution requires agents to identify limitations, generate improvement hypotheses, and test them autonomously through environmental interaction.
8 Conclusion
The survey presents Search Agents as a transformation of information retrieval toward proactive, context-aware use of diverse sources, while systematically analyzing their mechanisms, optimization, applications, evaluation, and challenges.
- Search Agents proactively leverage context and diverse sources, transforming search into an intelligent process beyond traditional web retrieval.
- The survey provides a systematic analysis of Search Agent mechanisms, optimization, applications, evaluation, and challenges.
Limitations
The survey primarily covers academic research papers, so it less extensively addresses commercial search-agent applications. Limited disclosure by major companies leaves open whether academic directions and performance align with commercial deployments.
- The survey focuses mainly on academic papers and therefore covers commercial application intricacies less extensively.
A Classification of Closed-ended QA
The survey classifies closed-ended QA evaluation around increasingly demanding information-seeking settings, including multi-hop reasoning, challenging web-based questions, and fact-checking. These tasks require iterative retrieval and comparison across sources to varying degrees.
- Multi-hop QA requires agents to connect information from multiple sources through multi-step reasoning and iterative retrieval.HotpotQA, 2WikiMultiHopQA, Bamboogle, and MuSiQue are frequently used datasets.
- Challenging QA benchmarks target real-world web questions that cannot be solved reliably from models’ parametric memory alone.
- Fact-checking evaluates iterative browsing and comparison across disparate sources, extending from text verification to long-form, unstructured, multimodal, and multi-hop scenarios.
B Detailed Challenges and Future Directions
Future search agents must broaden their information sources and modalities, improve verification and reinforcement learning, strengthen infrastructure, and support more autonomous coordination and self-evolution. These challenges span retrieval quality, training, scheduling, and continual adaptation.
- Search agents must integrate private and proprietary datasets while combining heterogeneous formats and resolving cross-source conflicts.
- Imperfect retrieval requires stronger information verification and fact-checking because external sources may contain noise, bias, or falsehoods.
- Multimodal search requires infrastructure for diverse queries and reasoning across text, images, audio, and video.
- Search-specific reinforcement learning must address long-horizon planning, imperfect feedback, sparse rewards, and open-ended queries without single definitive answers.
- Robust systems need efficient RL sampling, high-recall approximate retrieval, and priority-aware scheduling for faster access and resource management.
- Self-evolving agents would identify limitations, generate improvement hypotheses, and test them through environmental interaction without constant human intervention.
C Detailed Tables
The detailed tables organize search-agent methods, optimization strategies, applications, and datasets by structures, training choices, domains, and evaluation criteria. Together they compare tuning-free and tuning-based approaches across internal capabilities and external tasks.
- Methods and structures: Table 1 compares non-tuning methods by search structure, sub-structure, multi-agent architecture, test-time scaling, evaluation methods, and metrics.
- Optimization methods: Table 2 compares tuning-based methods by search structure, evaluation approach, training algorithms, supervision signals, and reward functions.
- Applications: Applications span internal memory, reasoning, and tool use alongside external mathematics, coding, finance, healthcare, and scientific research.
- Datasets and evaluation: Dataset comparisons cover categories, scales, modalities, construction methods, evaluation environments, and evaluation strategies.
- Comparative entries: Representative entries include sequential, hybrid, and parallel structures with rule-based, human, and LLM evaluation using metrics such as accuracy, F1, recall, and exact match.