Source-linked AI summary
Deep Research: A Survey of Autonomous Research Agents
Wenlin Zhang, Xiaopeng Li, Yingyi Zhang, Pengyue Jia, Yichao Wang, Huifeng Guo, Yong Liu, Xiangyu Zhao
TL;DR
LLMs remain constrained by internal knowledge boundaries, while deep research agents address this challenge through autonomous planning, retrieval, and synthesis. This survey systematically organizes the deep research pipeline into four stages and reviews methods, optimization techniques, benchmarks, and open challenges.
Problem
LLMs remain constrained by their internal knowledge boundaries, and deep research must capture specific information needs while ensuring broad topic coverage.
Method
The survey provides a capability-centric, systematic overview of deep research agents and analyzes planning, question developing, web exploration, and report generation.
Results
The survey categorizes representative methods for each pipeline stage and summarizes advances in optimization techniques and benchmarks for deep research.
Takeaways & Limitations
Deep research agents can transform retrieved evidence into coherent, factual, and goal-aligned generation, while some agents increasingly self-adjust planning routines in dynamic web environments.
Takeaways & Limitations
Deep research systems remain in their infancy and face multiple limitations, while most current systems rely solely on traditional search engines as their primary external tool.
Abstract
from arXiv · showhide
The rapid advancement of large language models (LLMs) has driven the development of agentic systems capable of autonomously performing complex tasks. Despite their impressive capabilities, LLMs remain constrained by their internal knowledge boundaries. To overcome these limitations, the paradigm of deep research has been proposed, wherein agents actively engage in planning, retrieval, and synthesis to generate comprehensive and faithful analytical reports grounded in web-based evidence. In this survey, we provide a systematic overview of the deep research pipeline, which comprises four core stages: planning, question developing, web exploration, and report generation. For each stage, we analyze the key technical challenges and categorize representative methods developed to address them. Furthermore, we summarize recent advances in optimization techniques and benchmarks tailored for deep research. Finally, we discuss open challenges and promising research directions, aiming to chart a roadmap toward building more capable and trustworthy deep research agents.
1 Introduction
Deep research extends agentic search by combining planning, adaptive questioning, web exploration, and structured report generation. The survey organizes this paradigm around modular capabilities and the challenges of producing coherent, factual, goal-aligned outputs from external evidence.
- Motivation: Deep research addresses the limits of internal LLM knowledge by combining web exploration with structured analysis and iterative generation.This transforms retrieved evidence into coherent, factual, and goal-aligned outputs.
- Pipeline: The pipeline comprises planning, question developing, web exploration, and report generation, with each stage imposing distinct agentic demands.The stages progress from decomposing research goals to integrating evidence into structured reports.
- Web Exploration: Web exploration uses iterative, agent-driven interaction with APIs or browsers to retrieve relevant information and deepen coverage of sparse or scattered evidence.It includes tool selection, query issuance, result parsing, and filtering of noisy or redundant content.
- Planning: Planning converts broad or ambiguous questions into flexible, interpretable, goal-driven plans that guide subsequent retrieval and synthesis.The agent determines what to search for, in what order, and how intermediate information supports downstream synthesis.
- Question Developing: Question developing generates diverse, complementary queries conditioned on sub-goals while balancing specificity with coverage.Adaptive, reinforcement-learning, supervised, and heuristic approaches are discussed as ways to improve query effectiveness, diversity, and relevance.
- Survey Scope: The survey adopts a capability-centric, modular taxonomy that examines planning, retrieval, reasoning, and report generation independently and jointly.It contrasts this fine-grained focus on capability formation and integration with surveys centered on system architecture or enumerating tasks and tools.
2 Planning
Planning is an explicit stage that transforms research questions and prior knowledge into structured subgoals or tool-invocation steps, guiding downstream execution. Recent methods make planning adaptive and learnable, while persistent challenges include brittle plans, weak evaluation, and limited transfer across tasks.
- Planning as structured execution: Explicit planning improves task decomposition, guides retrieval and reasoning systematically, and reduces failures from trial-and-error strategies.It replaces purely reactive step-by-step prompting with advance selection of actions or subtasks.
- Planning as structured execution: Planning transforms an initial research question and agent context into an ordered sequence of subgoals or tool-invocation steps.This structured plan connects high-level goals with low-level execution.
- Structured world knowledge: Planning methods increasingly use world models, simulation, knowledge graphs, iterative refinement, and efficiency-aware strategies to support long-horizon web tasks.These approaches balance environment modeling, trajectory feasibility, coherence, and resource usage.
- Learnable planning: Planning is becoming a learnable capability through architecture search, reusable components, interaction feedback, meta-learning, and large-scale task training.These methods shift planning from hand-crafted or static execution toward continual self-improvement and task-driven adaptation.
- Open challenges: Despite these advances, current plans remain brittle under ambiguity, may contain inconsistent or hallucinated steps, and are often evaluated only through end-task accuracy.These limitations make plan quality difficult to diagnose and strategy comparisons difficult to interpret.
- Open challenges: Many systems treat research questions independently, limiting accumulation of generalizable planning knowledge and transfer of strategies across tasks.Open challenges also include ambiguity, long-range consistency, and cross-task strategy transfer.
3 Question Developing
Question developing dynamically converts research subgoals into targeted, contextualized queries guided by the overall plan and accumulated evidence. The survey categorizes methods by optimization strategy and highlights their benefits, limitations, and open challenges.
- Overview: Question developing generates targeted, contextualized queries beyond a single static query to elicit useful evidence for multi-step reasoning and synthesis.The process guides retrieval toward task-aware search rather than keyword matching alone.
- Formalization: Given plan P, subgoal s_i, and accumulated evidence E, a query-generation model produces a set of search queries Q_i.The retrieved evidence set E is updated through the retrieval process described in the web-exploration stage.
- Taxonomy: Question developing methods divide into reward-optimized approaches using trial-and-error reward signals and supervision-driven approaches using demonstrations or designed strategies.The taxonomy organizes methods by training paradigm and modeling strategy.
- Reward-Optimized Methods: Reward-optimized methods learn adaptive query policies from answer, retrieval, evidence-quality, efficiency, tool-use, or multimodal rewards.Examples include staged format-and-answer rewards, simulated search, retrieval metrics such as Recall@K or NDCG@K, and visual-search rewards.
- Reward-Optimized Methods: Reward-optimized methods can learn interpretable query-development policies that improve behavior over time through reinforcement feedback in noisy, dynamic, and high-stakes environments.The survey describes these methods as adapting search decisions to query utility and end-task outcomes.
- Supervision-Driven Methods: Supervision-driven methods offer controllable, semantically faithful, stable, and diagnosable query generation, but depend on the quality and coverage of demonstrations.They avoid noisy or expensive search interaction during training and are useful when data efficiency and development simplicity matter.
- Challenges: Existing methods remain vulnerable to vague subgoals, weak use of query history and task context, and overly narrow queries for open-ended problems.These weaknesses can cause retrieval deviation, repetitive questions, and limited exploratory coverage.
4 Web Exploration
Web exploration retrieves and organizes evidence from heterogeneous online sources through browser-based agents or API-based systems. The survey compares their capabilities and limitations, emphasizing multimodal processing, verification, and hybrid architectures for more precise and trustworthy retrieval.
- Overview: Web exploration is critical because relevant evidence is sparsely distributed across heterogeneous web pages, making accurate location and extraction difficult.Deep-research workflows require precise, context-aware, and trustworthy evidence.
- Retrieval Paradigms: Web retrieval agents dynamically browse, click, extract, and filter content, while API-based systems obtain ranked documents or snippets from search engines.The two categories represent browser interaction and direct search-engine integration.
- Browser-Based Autonomous Web Agents: Browser-based agents make contextual navigation decisions through real or simulated browsers, enabling access to interactive or unindexed content.Text-based and DOM-based implementations support commands, page interaction, and programmatic extraction.
- Multimodal Web Agents: Text-oriented browser agents struggle when critical information appears in visual layouts or embedded visualizations, motivating multimodal agents.Multimodal systems combine textual content with screenshots, charts, interface patterns, and spatial organization.
- Multimodal Web Agents: 59% task success on real-world benchmarks was achieved by WebVoyager using GPT-4V, outperforming text-only baselines through screenshot analysis and HTML processing.Other systems embed screenshots in reasoning, reconstruct HTML/CSS, or provide broad interface evaluation frameworks.
- Multimodal Web Agents: Multimodal agents can answer complex queries with citations and synthesize findings from heterogeneous sources, although robustness and reliability remain challenging.The survey connects these capabilities with more expert-like web navigation and a shift toward higher-level analysis.
- Discussion: Web retrieval agents adapt to complex inquiries but are resource-intensive and lack real-time trustworthiness assessment, whereas APIs are efficient and reliable but may miss niche or dynamic content.Both approaches remain disconnected from upstream question formulation and evidence planning.
- Discussion: The survey points toward hybrid architectures that combine retrieval paradigms with evidence extraction, correctness verification, content-quality assessment, multimodal processing, and real-time verification.Such categorization and integration are presented as foundations for more precise and trustworthy evidence retrieval.
5 Report Generation
Report generation synthesizes fragmented web evidence into coherent, structured, and faithful reports. The survey organizes methods around structure control and factual integrity, while highlighting unresolved coupling and consistency challenges.
- Report generation synthesizes web information into comprehensive reports that are coherent, logically organized, and faithful to underlying evidence.
- The survey categorizes report-generation methods into structure control and factual integrity.Structure control targets coherence and consistency, whereas factual integrity targets faithfulness to retrieved evidence.
- Structure control: Structure control combines planning-based generation, constraint-guided decoding, and structure-aware alignment to organize long-form outputs.Hierarchical outlines and section-level synthesis help maintain global coherence, while critique-augmented supervision aligns local completions with document structure.
- Factual integrity: Factual integrity methods use faithful modeling, conflict resolution, and factuality evaluation to improve alignment with retrieved evidence.Approaches include verification layers, context-aware decoding, conflict modeling, attribution metrics, and benchmarks for long-form factual evaluation.
- Open challenges: Current report-generation methods remain limited by weak coordination with upstream stages and insufficient consistency modeling across multi-document, multi-hop contexts.Many approaches target isolated subskills, rely on fixed outlines, and lack flexibility for task-specific evidence distributions and reasoning flows.
6 Optimization
Deep Research workflows use either unified single-agent systems or modular multi-agent systems, with optimization methods tailored to their coordination patterns. Training approaches include contrastive learning, reinforcement learning, curriculum training, and hybrid strategies.
- Workflow architectures: Single-agent systems manage decomposition, querying, retrieval, and synthesis within one model, often using end-to-end reinforcement learning.
- Workflow architectures: Multi-agent systems assign specialized agents to planning, querying, retrieval, and writing, enabling modular optimization, parallelism, and flexibility.
- Optimization approaches: Contrastive learning trains agents to distinguish effective from ineffective tool-use trajectories, especially for retrieval, search, and summarization.
- Optimization approaches: Reinforcement learning optimizes long-term planning and decisions using rewards tied to retrieval accuracy and final artifact quality.
- Optimization approaches: Curriculum training stages learning from fundamental skills to full workflow orchestration and improves robustness in open-ended scenarios.
- Optimization approaches: Optimization choices often differ by architecture: single-agent systems favor end-to-end reward alignment, whereas multi-agent systems benefit from modular training and agent-level feedback.
7 Benchmark and Evaluation
Deep Research benchmarks range from search-oriented tasks to full-pipeline research evaluations. Full-pipeline benchmarks assess coordination across planning, question developing, web exploration, and report generation, while earlier structured-task benchmarks often under-evaluate final content generation.
- Benchmarks are broadly categorized as search-oriented or research-oriented according to task scope.
- Search-oriented benchmarks: Search-oriented benchmarks primarily evaluate question developing and web exploration through interactive browsing, query reformulation, and multilingual navigation.
- Evaluation gaps: Earlier benchmarks often emphasize planning components while rarely requiring final content generation, leaving report generation under-evaluated.
- Research-oriented benchmarks: Research-oriented benchmarks evaluate long-range reasoning, synthesis, and structured output across the full Deep Research pipeline.DeepResearch Bench and DeepResearchGym cover planning, subqueries, real-time web exploration, and research-style report generation.
- Research-oriented benchmarks: DeepResearch Bench and DeepResearchGym report metrics including clarity, knowledge precision/recall, and agreement scores.
8 Limitations and Future Directions
Deep Research systems remain in an early stage, with limitations in factuality, modality, workflow generalization, personalization, and tool diversity. Future directions emphasize stronger grounding, multimodal reasoning, adaptive orchestration, scalable training, and richer tool integration.
- Multi-tool integration: Most systems rely mainly on traditional search engines, restricting access to diverse, task-specific knowledge sources.Future agents should orchestrate APIs, structured databases, code repositories, documents, tables, and charts dynamically.
- Factuality: Systems can introduce inaccuracies, outdated claims, or unsupported assertions when synthesizing inconsistent sources across multi-step reasoning.The survey recommends source attribution, factuality-aware rewards, and post-hoc verification.
- Multimodal reasoning: Current pipelines are predominantly textual and remain poorly suited to tasks requiring images, PDFs, scanned documents, or structured data.Multimodal processing and reasoning are identified as largely unexplored but essential directions.
- Workflow design and optimization: Existing workflows lack mechanisms that learn strategies generalizing across tasks and evolving with new objectives.Aligning models with dynamic workflows is also challenged by high computational costs and sparse reward signals.
- Personalization: Existing methods lack persistent user modeling and dynamic adaptation, often treating personalization as secondary.The survey calls for scalable, privacy-aware user models with continual learning while addressing overfitting.
9 Conclusion
Deep research has emerged as a promising direction in agent research and is reshaping the search paradigm. This survey maps its four-stage pipeline, associated methods, optimization advances, benchmarks, and open directions toward more capable and trustworthy agents.
- Deep research is presented as one of the most promising directions in agent research and as a transformation of the search paradigm.
- The deep research pipeline comprises planning, question developing, web exploration, and report generation.
- For each pipeline stage, the survey analyzes technical challenges and categorizes representative methods addressing them.
- The survey summarizes recent optimization techniques and benchmarks tailored for deep research.
- The survey discusses open challenges and promising research directions to chart a roadmap toward more capable and trustworthy deep research agents.