PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey of Deep Research: Systems, Methodologies, and Applications

Renjun Xu, Jingwen Peng

arXiv:2506.12594v1cs.AIcs.MA

TL;DR

Manual literature reviews, experimental design, and data analysis are increasingly supplemented by intelligent systems; this survey traces Deep Research systems and identifies technical patterns. Reported figures include 26.6% accuracy on Humanity’s Last Exam (HLE) [212] and averaging 72.57% on the GAIA benchmark.

  • Problem

    Maintaining factual accuracy is a fundamental challenge for Deep Research systems.

    Source: §4.4.1 · p. 33

  • Method

    The survey uses behavioral analysis, publicly available documentation, standardized benchmarks, and qualitative assessment for consistent evaluation.

    Source: §9.2 · p. 79

  • Results

    Reported figures include 26.6% accuracy on Humanity’s Last Exam (HLE) [212] and averaging 72.57% on the GAIA benchmark.

    Source: §5.3 · p. 39

  • Takeaways & Limitations

    The survey traces Deep Research systems’ development and identifies key technical patterns across implementations.

    Source: §9 · p. 76

    Comprehensive technical analysis is constrained by limited implementation details.

    Source: §9.2 · p. 78

Abstract

from arXiv · show

This survey examines the rapidly evolving field of Deep Research systems -- AI-powered applications that automate complex research workflows through the integration of large language models, advanced information retrieval, and autonomous reasoning capabilities. We analyze more than 80 commercial and non-commercial implementations that have emerged since 2023, including OpenAI/Deep Research, Gemini/Deep Research, Perplexity/Deep Research, and numerous open-source alternatives. Through comprehensive examination, we propose a novel hierarchical taxonomy that categorizes systems according to four fundamental technical dimensions: foundation models and reasoning engines, tool utilization and environmental interaction, task planning and execution control, and knowledge synthesis and output generation. We explore the architectural patterns, implementation approaches, and domain-specific adaptations that characterize these systems across academic, scientific, business, and educational applications. Our analysis reveals both the significant capabilities of current implementations and the technical and ethical challenges they present regarding information accuracy, privacy, intellectual property, and accessibility. The survey concludes by identifying promising research directions in advanced reasoning architectures, multimodal integration, domain specialization, human-AI collaboration, and ecosystem standardization that will likely shape the future evolution of this transformative technology. By providing a comprehensive framework for understanding Deep Research systems, this survey contributes to both the theoretical understanding of AI-augmented knowledge work and the practical development of more capable, responsible, and accessible research technologies. The paper resources can be viewed at https://github.com/scienceaix/deepresearch.

1 Introduction

The introduction frames Deep Research as AI systems that automate end-to-end research workflows, outlining the field’s definition, scope, evolution, and practical significance. It presents the survey’s technical taxonomy and comparison of representative approaches, while posing questions about technical innovation, performance, usability, ethics, challenges, and future priorities.

1.1 Definition and Scope of Deep Research

  • Deep Research combines knowledge discovery, end-to-end workflow automation, and human-AI collaboration; the survey focuses on systems exhibiting at least two dimensions and using LLMs as foundational reasoning engines.Source: §1.1 · p. 4
  • The survey excludes purely bibliometric or single-stage systems and specialized tools lacking integrated, end-to-end research capabilities.Source: §1.1 · pp. 4–5

1.2.1 Origin and Early Exploration (2023 - February 2025).

  • Early workflow and agent frameworks predated the Deep Research boom, while Google Gemini introduced an initial implementation in December 2024 with basic multi-step reasoning and knowledge integration.Source: §1.2.1 · p. 5

1.2.2 Technological Breakthrough and Competitive Rivalry (February - March 2025).

  • OpenAI’s February 2025 Deep Research, powered by o3, demonstrated autonomous planning, cross-domain analysis, and report generation, with accuracy rates exceeding prior benchmarks on complex tasks.Perplexity also launched a free-to-use Deep Research product in February 2025.Source: §1.2.2 · p. 5

1.2.3 Ecosystem Expansion and Multi-modal Integration (March 2025 - Present).

  • The ecosystem matured through localized open-source deployments and commercial multimodal and multi-agent capabilities, while Claude/Research introduced agentic search with verifiable citations.Source: §1.2.3 · pp. 5–6

1.3 Significance and Practical Implications

  • Deep Research is presented as supporting academic innovation through automated literature synthesis and broader interdisciplinary discovery, and enterprise decisions through analysis of market trends, competitors, and strategic opportunities.Source: §1.3 · pp. 6–7
  • Open-source implementations are described as lowering barriers to sophisticated research capabilities for people and organizations regardless of technical expertise or resource constraints.Source: §1.3 · p. 7

1.4 Research Questions and Contribution of this Survey

  • The survey asks how architectural choices—including system architecture, implementation approach, and functional capabilities—shape Deep Research systems.Source: §1.4 · p. 7
  • The survey asks what innovations have emerged in fine-tuning, retrieval, and workflow orchestration, and how systems balance performance, usability, and ethics, including n8n and OpenAI/AgentsSDK.Source: §1.4 · p. 7
  • The survey proposes a taxonomy that categorizes systems by technical architecture, from foundation models to knowledge-synthesis capabilities.Source: §1.4 · p. 7
  • It compares representative systems across evaluation metrics to highlight the strengths and limitations of different approaches.Source: §1.4 · p. 7
  • It identifies key challenges and outlines a future-development roadmap focused on emerging architectures and integration opportunities.Source: §1.4 · p. 7

2 The Evolution and Technical Framework of Deep Research

This section traces the technical evolution of Deep Research systems through a taxonomy of core capabilities and representative approaches, covering coordinated tool use, adaptive planning and execution, information evaluation, and structured reporting. It also describes advances in evidence-integrated reports and interactive interfaces for exploration and iterative refinement, with attention to reliability and resource constraints.

2.1 Foundation Models and Reasoning Engines: Evolution and Advances

  • The foundation of Deep Research systems is their AI models and reasoning capabilities, which have evolved from general-purpose language models to research-oriented architectures.Source: §2.1 · p. 7

2.1.1 From General-Purpose LLMs to Specialized Research Models.

  • Current research systems move beyond early general-purpose LLMs with minimal task-specific optimization, and GPT-4-to-o3 demonstrates improvements in abstraction, multi-step reasoning, and knowledge integration for research tasks.AutoGLM-Research uses specialized training regimes to optimize ChatGLM for research-intensive tasks, achieving significant performance gains through targeted enhancements.Source: §2.1.1 · pp. 8–9

2.1.2 Context Understanding and Memory Mechanisms.

  • Modern episodic buffers, hierarchical compression, and attention retrieval address early context limits; Perplexity/DeepResearch analyzes hundreds of sources while maintaining coherent reasoning.Its structured information management uses DeepSeek-R1 capabilities.Source: §2.1.2 · p. 9

2.1.3 Enhancements in Reasoning Capabilities.

  • Reasoning has advanced from zero- or few-shot prompting to chain-, tree-, and graph-based methods, with o3 self-critique and recursive refinement; math models reached IMO silver-medal standards.Debate-driven reasoning is also described as supporting weak-to-strong generalization through structured argumentation [132].Source: §2.1.3 · pp. 9–10

2.2 Tool Utilization and Environmental Interaction: Evolution and Advances

  • Deep Research systems interact with external environments to gather and process information, extending beyond core language-model functions.Source: §2.2 · p. 10

2.2.1 Web Interaction Technology Development.

  • Web interaction has evolved from simple API searches to dynamic content handling, authentication, semantic extraction, and adaptive multi-page navigation.These capabilities expand access to web sources and insights from complex web environments.Source: §2.2.1 · p. 10

2.2.2 Content Processing Technology Advancements.

  • Modern systems process structured tables, visualizations, PDFs, and interactive applications; o3-based systems extract semantic structure and integrate insights across modalities.The survey also describes specialized extraction modules and content-normalization pipelines for varied document formats.Source: §2.2.2 · p. 10

2.2.3 Specialized Tool Integration Progress.

  • Tool integration has advanced from basic search and APIs to orchestrated domain-specific workflows, including ToolLLM’s [222] access to over 16,000 APIs and on-demand code generation.Advanced systems select tools dynamically and compose research workflows from available capabilities.Source: §2.2.3 · p. 11

2.3.1 Research Task Planning Development.

  • Planning increasingly refines hierarchical tasks using intermediate results and resource constraints, while AI-assisted code review can lengthen pull-request closure times [59].The section describes a shift from linear execution to adaptive strategies that account for time, computation, and information availability.Source: §2.3.1 · pp. 11–12

2.3.2 Autonomous Execution and Monitoring Advances.

  • Execution control now uses concurrent paths, monitoring, explicit success criteria, failure detection, and autonomous recovery to improve reliability on complex tasks.Agent-RL/ReSearch learns execution strategies from experience and adapts to failures and intermediate results.Source: §2.3.1 · p. 12

2.3.3 Multi-Agent Collaboration Framework Development.

  • Multi-agent systems replace monolithic agents with specialized roles, coordination, information sharing, and conflict resolution, enabling complex collaborative workflows [49].Multi-agent debate can improve evaluation consistency [48], while generative-AI voting shows resilience to model biases [162].Source: §2.3.1 · p. 12

2.4.1 Information Evaluation Technology Development.

  • Information evaluation has progressed from reputation heuristics to assessing source and content characteristics, uncertainty, contradictions, and evidence, improving output reliability.Generative-AI retrieval can also help source and verify information [306].Source: §2.4 · p. 13

2.4.2 Report Generation Technology Advances.

  • Reports have evolved from simple summaries to coherent, evidence-integrated structures with outputs adapted to audience expertise, information needs, and presentation context.The section links this evolution to improved usability and impact of Deep Research outputs.Source: §2.4.2 · p. 13
  • OpenDeepResearcher generates structured research reports with explicit attribution, structured arguments, and integrated supporting evidence.The survey presents these capabilities as enhancing the utility and trustworthiness of Deep Research outputs.Source: §2.4.2 · p. 13
  • MegaWika [22] provides a large-scale multilingual resource of millions of articles and referenced sources for collaborative AI.Source: §2.4.2 · p. 13

2.4.3 Interactive Presentation Technology Development.

  • Modern interfaces enable drill-down, source verification, alternative-viewpoint examination, and iterative refinement through user feedback and queries.This contrasts with early systems’ fixed textual outputs and minimal user interaction.Source: §2.4.2 · p. 14
  • Auto-Deep-Research lets users explore findings, examine supporting evidence, and refine analysis through iterative interaction.The survey associates these interactive presentation capabilities with more accessible outputs and effective knowledge transfer and utilization.Source: §2.4.2 · p. 14

3 Comparative Analysis and Evaluation of Deep Research Systems

This section compares Deep Research systems across reasoning, tool use, planning, synthesis, personal knowledge-management adaptability, and evaluation performance, considering their fit for academic, enterprise, and personal knowledge work. It describes tradeoffs between specialized and broad capabilities and between research quality and output needs, alongside differences in benchmark performance, output coherence and factual reliability, customization, and strengths in specific use cases.

3.1.1 Foundation Model and Reasoning Efficiency Comparison.

  • OpenAI and Google use proprietary models for coherent processing of larger information volumes; o3 is strong in complex reasoning, while Gemini 2.5 Pro excels at integrating diverse sources.Perplexity/DeepResearch achieves competitive performance with open-source DeepSeek-R1 through optimized implementation, while Camel-AI/OWL and QwenLM/Qwen-Agent show specialized optimization can make accessible models effective.Source: §3.1.1 · pp. 14–15

3.1.2 Tool Integration and Environmental Adaptability Comparison.

  • Nanobrowser specializes in web navigation and extraction, while dzhng/deep-research and nickscamara/open-deep-research add document processing; Manus and AutoGLM cover broader environments.Broad platforms may lack focused tools' domain performance; n8n is flexible for API integration but more limited in direct web and document interaction.Source: §3.1.2 · p. 15

3.1.3 Task Planning and Execution Stability Comparison.

  • OpenAI/AgentsSDK uses hierarchical task decomposition and adaptive execution for complex research workflows with reliable completion rates; Agent-RL/ReSearch develops robust execution strategies through reinforcement learning; smolagents/open_deep_research and TARS focus on multi-agent collaboration.Flowith/OracleMode offers advanced planning optimized for research tasks, though with more limited error recovery capabilities.Source: §3.1.2 · pp. 15–16

3.1.4 Knowledge Synthesis and Output Quality Comparison.

  • OpenAI/DeepResearch emphasizes report structure, Perplexity/DeepResearch citation practices, HKUDS/Auto-Deep-Research interactive exploration, and grapeot/deep_research_agent information evaluation.Both commercial platforms assess source credibility and reliability; mshumer/OpenDeepResearcher focuses on well-formatted reports, while grapeot prioritizes content assessment over presentation.Source: §3.1.2 · p. 16

3.2.1 Academic Research Scenario Adaptability Assessment.

  • OpenAI/DeepResearch demonstrates exceptional suitability for academic research through academic-database access, comprehensive literature review, methodological rigor, and citation quality; PaperQA and Scite offer complementary capabilities.Source: §3.2.1 · p. 16

3.2.2 Enterprise Decision-Making Scenario Adaptability Assessment.

  • Gemini/DeepResearch demonstrates exceptional suitability for enterprise decision-making through strong information currency, analytical capabilities, and actionable output formats; Manus offers similarly strong performance in information acquisition and analysis.Source: §3.2.2 · p. 17

3.2.3 Personal Knowledge Management Adaptability Assessment.

  • Perplexity/DeepResearch offers strong accessibility for personal knowledge management.Source: §3.2.3 · p. 18
  • Consumer-friendly services offer free access but limited personalization, whereas open-source systems enable local customization and infrastructure tools integrate with existing knowledge workflows.Open-source implementations include nickscamara/open-deep-research and OpenManus; infrastructure examples include Nanobrowser and Jina-AI/node-DeepResearch.Source: §3.2.3 · p. 18
  • More complex frameworks offer sophisticated capabilities but may challenge non-technical users.The passage identifies smolagents/open_deep_research as an example.Source: §3.2.3 · p. 18

3.3.1 Quantitative Evaluation Metrics.

  • OpenAI/DeepResearch leads across benchmark categories, especially Humanity’s Last Exam, while Gemini is comparable and Gemini 2.5 Pro Experimental showed higher user preference across four metrics.The supplied passages do not provide a complete metric-value pair for the preference comparison.Source: §3.3.1 · p. 18

3.3.2 Qualitative Assessment Frameworks.

  • Commercial systems generally score strongly on output coherence and factual accuracy: OpenAI produces structured reports, and Gemini shows similar strengths.OpenAI/DeepResearch is also described as moderately innovative in connecting disparate sources; Gemini/DeepResearch places slightly less emphasis on this.Source: §3.3.1 · p. 20
  • Open-source systems show distinct qualitative strengths: Agent-RL/ReSearch emphasizes insight novelty, while grapeot/deep_research_agent emphasizes factual accuracy.These strengths are associated with exploration and information verification, respectively.Source: §3.3.1 · p. 20

3.3.3 Efficiency and Resource Utilization Metrics.

  • Perplexity/DeepResearch combines competitive output quality with quick response times and high token efficiency, while cloud services depend on external infrastructure and incur costs.Commercial cloud services are also described as having moderate response times.Source: §3.3.1 · p. 20
  • Open-source efficiency varies: AutoGLM-Research and QwenLM/Qwen-Agent need substantial compute but allow local control, while lighter implementations can be slower and less token-efficient.nickscamara/open-deep-research is cited as a lighter-weight example.Source: §3.3.1 · pp. 20–21

3.3 Performance Metrics and Benchmarking

  • Overall, commercial systems lead standard benchmarks, while open-source alternatives provide competitive capabilities in specialized uses and advantages in customization, control, and potential cost efficiency.These advantages are identified for specialized applications.Source: §3.3.1 · p. 21

4 Implementation Technologies and Challenges

This section surveys Deep Research architectures, agent frameworks, and implementation approaches, including coding assistance and deployment considerations. It examines trade-offs and challenges in reasoning, retrieval, resource use, integration, coordination, reliability, privacy, transparency, and accessibility.

4.1 Architectural Implementation Patterns

  • Monolithic systems centralize control and shared memory around a core reasoning engine, supporting coherence but limiting extensibility and parallelization; OpenAI/DeepResearch is an example.Specialized modules integrate directly with the controller, and operations follow structured sequences.Source: §4.1 · pp. 21–22

4.1.2 Pipeline-Based Architecture Pattern.

  • Pipeline architectures sequence specialized stages through defined interfaces, enabling component reuse and workflow customization but making iterative refinement across components difficult; n8n exemplifies this approach.Advanced implementations can route tasks conditionally based on intermediate results.Source: §4.1.2 · pp. 22–23
  • Multi-agent systems distribute research among specialized autonomous agents using coordination protocols, enabling parallel work and scaling while risking inconsistent reasoning; smolagents/open_deep_research is an example.Roles can include searcher, analyst, and critic, with task allocation based on agent capabilities and workload.Source: §4.1.2 · pp. 23–24

4.1.4 Hybrid Architecture Pattern.

  • Hybrid systems combine architectural patterns across system levels and adapt execution to task or domain needs, offering flexibility while increasing implementation and integration complexity; Perplexity/DeepResearch is an example.They can combine centralized reasoning with distributed information gathering and specialized processing pipelines.Source: §4.1.4 · p. 25

4.1.5 Emerging Agent Framework Ecosystems.

  • Agent frameworks standardize components and orchestration: LangGraph provides graph-based control and state management, while CrewAI supports role-based collaboration.Other frameworks include Google ADK for tool integration, planning, and monitoring, Agno for autonomy and meta-reasoning, and TapeAgents [19] for iterative refinement.Source: §4.1.5 · pp. 26–27

4.1.6 Architectural Pattern Comparison.

  • Table 13 provides a comparison of architectural pattern characteristics across key performance dimensions.Source: §4.1.6 · p. 27, Table 13 · p. 27
  • Monolithic designs favor coherent reasoning and simplicity, pipelines extensibility and reuse, and multi-agent systems parallelization and fault tolerance; benchmarking still requires standardized tasks and environments.Hybrid architectures balance these characteristics for flexible optimization.Source: §4.1.6 · p. 27

4.2.1 Distributed Reasoning Architectures.

  • Distributed reasoning splits complex queries into parallel paths; OpenAI/DeepResearch coordinates execution threads, while open-source systems demonstrate parallelization under constrained resources.These architectures must also maintain memory coherence and state; implementations range from hierarchical coordination to centralized knowledge repositories.Source: §4.2.1 · pp. 27–28
  • Concurrent search accelerates gathering, while adaptive query planning, deduplication, source diversification, caching, and fingerprinting reduce redundancy and broaden coverage.Perplexity/DeepResearch dispatches dozens of queries; open-source coordination can use straightforward caching and result fingerprinting.Source: §4.2.1 · pp. 28–29
  • Systems allocate resources according to task complexity and progressively refine outputs, providing early findings or previews while deeper analysis continues.Perplexity/DeepResearch uses staged analysis, while open-source approaches show that progressive processing need not require sophisticated infrastructure.Source: §4.2.1 · p. 29

4.3 System Integration and Interoperability

  • Deep Research implementations have largely incompatible interfaces; MCP and A2A could support standardized, interchangeable components, but broader adoption is needed.MCP structures model-tool interaction, whereas A2A standardizes agent-to-agent communication.Source: §4.3 · pp. 29–30
  • Shared protocols, task and result structures, compatibility layers, and interoperability evaluations could let specialized components work within unified frameworks and accelerate innovation.Source: §4.3 · p. 30

4.3.2 Tool Integration Frameworks.

  • Monitoring frameworks track tool execution and failures, while open implementations such as Agent-RL/ReSearch use simpler tracking and retries for common failure modes.The survey notes that effective monitoring can be achieved with relatively straightforward implementations.Source: §4.3.2 · pp. 30–31

4.3.3 Cross-Platform Compatibility.

  • Platform abstraction separates core reasoning from environment-specific integrations, while container configurations support consistent deployment across infrastructures, as in TARS and OpenManus.Source: §4.3.3 · p. 31

4.3.4 Research-Oriented Coding Assistance Integration.

  • Coding assistants range from IDE-native completion such as GitHub Copilot [20] to conversational code generation such as ChatGPT-based tools [309].Source: §4.3.4 · p. 31
  • CodeCompose saved 17% of keystrokes through optimized latency solutions despite initial usability challenges [72].Source: §4.3.4 · p. 32
  • Coding assistant evaluations report comprehension gaps despite generative capabilities, while acceptance and effectiveness vary with task characteristics and user expertise [260].Source: §4.3.4 · p. 32
  • Coding-assistant integration can reduce technical barriers between research design and computation, but effectiveness depends on research-specific requirements and reproducible workflows.These requirements include documentation standards, reproducibility, and field-specific libraries and frameworks.Source: §4.3.4 · p. 33

4.4.1 Hallucination Control and Factual Consistency.

  • Straightforward grounding and contradiction handling can improve factual reliability, while uncertainty modeling and conflicting-evidence presentation help when definitive conclusions are unavailable.The passages describe source attribution and verification, alongside contradiction identification for user review.Source: §4.4.1 · pp. 33–34
  • Privacy safeguards include tenant isolation, local deployment via OpenManus, and controlled source access via Flowith/OracleMode; CI-Bench [56] evaluates adherence to contextual norms.Controlled access respects authentication requirements and source limitations.Source: §4.4.1 · p. 34

4.4.3 Explainability and Transparency.

  • Scientific applications require transparent, reproducible reasoning, and systems expose reasoning traces and source-linked citations to support understanding and verification.OpenAI/DeepResearch documents analytical steps, while Perplexity/DeepResearch links claims to original sources.Source: §4.4.3 · p. 34
  • Across the ecosystem, open-source systems can achieve effective research capabilities through pragmatic approaches, while strategies reflect tradeoffs among capability, efficiency, reliability, and accessibility.Commercial platforms benefit from extensive infrastructure and specialized components.Source: §4.4.3 · pp. 34–35

5 Evaluation Methodologies and Benchmarks

This section surveys how Deep Research systems are evaluated, spanning task completion, retrieval, synthesis, operational performance, reliability, usability, accessibility, and cross-domain benchmarks. It also covers emerging evaluations and system-selection methods, noting benchmark gaps and the need to match criteria to application needs despite heterogeneous data.

5.1.1 Task Completion Capability Assessment.

  • AutoGLM records 55.2% success on VAB-WebArena-Lite, 96.2% on OpenTable tasks, 36.2% on AndroidLab, and 89.7% on common tasks in popular Chinese apps; AutoGLM [154] improves WebArena-Lite success from 55.2% to 59.1% on a second attempt.The survey cautions that benchmark performance provides comparative indicators, while practical effectiveness depends on task characteristics and domain specifics.Source: §5.1.1 · p. 36
  • Retrieval evaluation uses precision, recall, and F1; Perplexity/DeepResearch is described as strong in recall, but the survey found no evidence that surveyed systems were formally evaluated on TREC [214].Source-diversity and source-spectrum measures assess breadth across domains, perspectives, and publication types beyond simple relevance metrics.Source: §5.1.1 · p. 36
  • OpenAI/DeepResearch achieved 26.6% accuracy on Humanity’s Last Exam (HLE) [212], while Perplexity/DeepResearch attained 21.1% on the same benchmark; TruthfulQA [151] figures for both were not publicly available.Source: §5.1.1 · p. 37

5.2 Non-Functional Evaluation Metrics

  • Response-time profiling reports Perplexity/DeepResearch at 2–5 minutes for moderate tasks and OpenAI/DeepResearch at 5–10 minutes for similar complexity.Open-source response times vary by implementation and deployment; the passage contrasts accessibility emphasis in nickscamara/open-deep-research with response-time optimizations in QwenLM/Qwen-Agent.Source: §5.2 · pp. 37–38
  • Resource profiling covers memory, computational requirements, and energy; commercial cloud systems obscure some measures, while Camel-AI/OWL and AutoGLM-Research provide more transparent resource profiles.Operational costs offer indirect resource indicators for managed systems, and resource differences affect deployment scenarios and accessibility.Source: §5.2 · p. 38
  • Reliability evaluation measures failure frequency across diverse scenarios, and the passage reports that commercial systems generally have lower error rates than open-source alternatives.Source: §5.2 · p. 38
  • Reliability assessment probes adversarial failure modes and error recovery, while extended-session and repeated-run consistency matters especially for open-source deployment environments.The section treats engineering quality as important beyond core algorithmic performance.Source: §5.2 · p. 38
  • Usability assessment can use SUS and task-completion time; commercial systems are described as stronger overall, with Perplexity/DeepResearch emphasizing intuitive interaction for nontechnical users.Open-source alternatives vary, and HKUDS/AutoDeep-Research incorporates interface enhancements for accessibility.Source: §5.2 · p. 38
  • Commercial systems generally have gentler learning curves, while open implementations vary from technically demanding n8n to simplified, more accessible alternatives.The assessment measures time-to-proficiency across users with different technical backgrounds and highlights an accessibility-sophistication tradeoff.Source: §5.2 · p. 39

5.3 Cross-Domain Evaluation Benchmarks

  • No dedicated literature-review benchmark is identified, and the authors propose using Nature Reviews papers as gold standards and citation networks to test retrieval and seminal-work identification.The proposed evaluation targets literature identification, synthesis, and research-gap detection.Source: §5.3 · p. 39
  • OpenAI/DeepResearch achieves 26.6% accuracy on HLE [212] and averages 72.57% on GAIA [172], while Perplexity/DeepResearch achieves 21.1% accuracy on HLE [212] and 93.9% on SimpleQA [290].SimpleQA evaluates short, fact-seeking questions; these indirect benchmarks offer insight into capabilities relevant to literature synthesis, while direct literature-review benchmarks remain underdeveloped.Source: §5.3 · pp. 39, 41
  • No benchmark specifically assesses research-methodology strengths and weaknesses; proposed evaluations include flaw detection, reproducibility, confounders, statistical power, and uncertainty.Expert-annotated papers could provide a gold standard for evaluating analytical capability.Source: §5.3 · p. 40
  • OpenAI/DeepResearch is described as handling complex multidomain business analysis with detailed recommendations, while Gemini/DeepResearch processes extensive datasets into concise, factual reports.Business benchmarks require integrating quantitative financial data with qualitative market dynamics and regulatory considerations.Source: §5.3 · p. 40
  • No Deep Research projects have published official FinEval [103] results, though OpenAI/DeepResearch is described as strong in quantitative financial analysis and n8n as competitive through financial-data and tool integration.The survey emphasizes domain-specific integrations and data accessibility beyond core language-model capability.Source: §5.3 · pp. 40–41
  • 93.9% accuracy is reported for Perplexity/DeepResearch on SimpleQA, while OpenAI and Gemini/DeepResearch have no publicly available SimpleQA accuracy figures.Simple factual-recall metrics are useful baselines but do not fully represent complex research workflows.Source: §5.3 · p. 41
  • Humanities and social-science tasks show more variable, generally lower accuracy than STEM, while commercial systems lead and specialized open alternatives can be competitive.MMLU covers historical analysis, ethical evaluation, and social-trend identification; the differences underscore challenges in contextual understanding and interpretation.Source: §5.3 · p. 41

5.4 Emerging Evaluation Approaches

  • Interactive evaluations address static benchmarks’ limits by testing iterative research and clarification, but no surveyed system has public QuestBench [141] results.QuestBench assesses identifying missing information and asking clarifying questions; systems have shown strong performance in other interactive evaluations.Source: §5.4 · p. 41
  • Multimodal evaluation tests integration across text, images, visualizations, and structured content; Gemini/DeepResearch excels on image-inclusive tasks, while open systems show emerging capabilities.Jina-AI/node-DeepResearch incorporates components for multimodal processing.Source: §5.4 · p. 42
  • Ethical evaluation considers awareness, bias detection, and fairness: OpenAI/DeepResearch uses explicit guidelines and mitigation, while grapeot/deep_research_agent emphasizes source transparency and attribution.These dimensions extend assessment beyond technical performance amid concerns that automated research can amplify existing information biases.Source: §5.4 · p. 42

5.5.1 Systems Selection Criteria.

  • Systems are selected for implementing at least two core Deep Research dimensions, adequate public documentation, recent activity, and balanced commercial, open-source, general-purpose, and specialized representation.The methodology aims to support systematic, consistent evaluation across systems.Source: §5.5.1 · pp. 42–43

5.5.2 Evaluation Dimensions and Metrics Application.

  • Evaluation applies consistent dimensions across systems, but benchmarks vary with system focus and available performance data; Table 15 presents coverage across representative systems.The supplied passage does not provide table cell values.Source: §5.5.2 · p. 43, Table 15 · p. 43

5.5.3 Data Collection Methods.

  • Data collection combines published benchmarks, documentation, repository analysis, and direct testing; unavailable benchmark results are reported rather than extrapolated.Comparisons are complicated by benchmark diversity, limited commercial-system transparency, rapid updates, and domain specialization, so qualitative architectural analysis complements quantitative metrics.Source: §5.5.3 · p. 43

6 Applications and Use Cases

Deep Research systems are described as supporting academic and interdisciplinary research through literature synthesis, hypothesis work, and cross-domain methods. The section highlights applications from review automation and gap identification to methodology transfer and knowledge translation.

6.1.1 Literature Review and Synthesis.

  • Medical research reviews using OpenAI/DeepResearch analyze thousands of publications to identify intervention efficacy patterns with significantly reduced human effort compared with traditional methods.Related systems are described as enabling rapid synthesis of research.Source: §6.1.1 · p. 45
  • Automated systematic reviews expand queries to retrieve relevant scholarly articles and reduce time and resource burdens [234].Implementations can customize search strategies, source weightings, and output formats for disciplinary conventions.Source: §6.1.1 · p. 45
  • Gemini/DeepResearch identifies interdisciplinary research opportunities, while HKUDS/Auto-Deep-Research detects methodological limitations and underexplored variables across research corpora.These systems extend literature synthesis toward identifying research gaps and directions.Source: §6.1.1 · p. 45
  • OpenAI/DeepResearch supports hypothesis generation from literature synthesis, while Camel-AI/OWL applies domain-specific constraints and validation criteria.The described implementations use researcher feedback to refine hypotheses against research objectives.Source: §6.1.1 · pp. 45–46

6.1.3 Interdisciplinary Research Support.

  • OpenAI/DeepResearch supports methodology transfer by explicitly identifying adaptation requirements and implementation guidance for applying techniques across fields, while Perplexity/DeepResearch maps concepts across disciplines to reduce unfamiliar-domain onboarding barriers.Source: §6.1.3 · p. 46

6.2 Scientific Discovery Applications

Deep Research systems support scientific work through data analysis, experimental planning and simulation, literature integration, and increasingly autonomous discovery workflows. Their use in open-ended discovery remains limited and calls for oversight and validation.

6.2 Scientific Discovery Applications

  • Gemini/DeepResearch synthesizes hundreds of climate models and observational datasets to identify consistent patterns and outliers while integrating diverse formats coherently.Tool integration also extends analysis beyond language models’ native capabilities.Source: §6.2 · p. 47
  • SqlCompose [161] automates SQL authoring to reduce syntax barriers in large-scale data operations, while DataInquirer measures practitioner workflow variation and AI’s effect on novice-expert alignment [325].Interactive data-wrangling recommendations support transformation and cleaning [211].Source: §6.2 · p. 47
  • OpenAI/DeepResearch identifies unexpected drug-interaction patterns in clinical literature and proposes mechanistic explanations to guide further investigation.This combines pattern recognition with explanatory hypothesis generation; specialized anomaly tools also target statistical outliers and contextual explanations.Source: §6.2 · p. 47

6.2.2 Experiment Design and Simulation.

  • Systems support protocol generation and optimization, but Yu et al. [314] identify gaps in open-ended exploration, creative hypothesis generation, and experimental-design optimization that constrain autonomous discovery.Gemini/DeepResearch incorporates best practices, potential confounds, and control strategies; Agent-RL/ReSearch emphasizes statistical power and confound control.Source: §6.2.2 · pp. 47–48
  • OpenAI/DeepResearch integrates computational modeling tools to rapidly assess theoretical predictions against existing evidence, accelerating theory refinement by identifying empirical constraints and validation opportunities more efficiently than manual methods.Manus provides sophisticated orchestration of computational modeling and simulation tools within research.Source: §6.2.2 · p. 48
  • Gemini/DeepResearch synthesizes scientific figures, tables, and text into cohesive analyses, enabling broader use of multimodal literature content.Customizable open implementations also support multimodal scientific-content processing.Source: §6.2.2 · p. 48
  • Perplexity/DeepResearch guides resolution of conflicting scientific findings by identifying methodological differences and contextual factors rather than simply aggregating evidence.HKUDS/Auto-Deep-Research uses explicit evidence weighting and confidence estimation.Source: §6.2.2 · p. 48

6.2.4 Autonomous Scientific Discovery.

  • Oversight and validation are identified as a limitation for autonomous scientific discovery.Source: §6.2.4 · p. 49

6.3 Business Intelligence Applications

The section describes business-intelligence uses of Deep Research, including trend detection, investment assessment, risk analysis, and process improvement. These applications combine information across sources with domain-specific analytical components.

Emerging Trend Identification

  • OpenAI/DeepResearch analyzes temporal patterns across industry publications, startup activity, and technology indicators to identify emerging trends with greater lead time than manual methods alone.Flowith/OracleMode is described as incorporating trend modeling and weak-signal amplification.Source: §6.3 · pp. 49–50

Strategic Decision Support

  • Perplexity/DeepResearch integrates financial metrics, market positioning, competitive dynamics, and growth indicators to support investment evaluation.Open frameworks such as mshumer/OpenDeepResearcher emphasize structured evaluation and source integration.Source: §6.3 · p. 50
  • Gemini/DeepResearch identifies regulatory, competitive, technological, and market threats with associated impact and likelihood estimates.Manus incorporates risk categorization and prioritization mechanisms.Source: §6.3 · p. 50

Business Process Optimization

  • OpenAI/DeepResearch benchmarks processes against industry standards and adjacent-sector practices to identify optimization opportunities and support cross-industry learning.Gemini/DeepResearch also provides implementation planning with change-management, resource, and risk-mitigation considerations.Source: §6.3 · p. 50

6.4 Financial Analysis Applications

Deep Research is presented as supporting financial assessment through integrated asset, management, trend, and risk analysis. The section also notes potential limits in tasks requiring deep domain understanding and the possibility of redundant or inaccurate outputs.

6.4 Financial Analysis Applications

  • Perplexity/DeepResearch integrates financial metrics, market position, competitive dynamics, and growth indicators for asset evaluation, providing broader information integration than manual methods typically allow.Open implementations such as n8n connect specialized financial data sources and analytical tools.Source: §6.4 · p. 51
  • OpenAI/DeepResearch assesses management using historical performance, leadership approach, strategic consistency, and reputation across sources.The section says this can provide deeper leadership insights than standard financial analysis.Source: §6.4 · p. 51

6.4.2 Financial Trend Analysis.

  • Systems identify patterns across financial indicators and contextual factors, with trend analysis integrating more factors than manual analysis typically allows.The passage also describes specialized trend-analysis components for statistical pattern detection and causal-factor identification.Source: §6.4.2 · p. 51
  • The section notes that AI-system effectiveness may be limited for tasks requiring deep domain understanding and that generated outputs can contain redundancy or inaccuracies [254].These cautions qualify the described financial applications.Source: §6.4.2 · p. 52
  • Perplexity/DeepResearch combines market, credit, operational, regulatory, and systemic risks, while Gemini/DeepResearch supports stress testing using crisis patterns, risk models, and system dependencies.The described stress tests identify potential vulnerabilities.Source: §6.4.2 · p. 52

6.5 Educational Applications

Educational applications tailor research systems to learners through personalized pathways, explanations, content creation, and guided research practice. The section presents specialized adaptation as extending utility beyond general research capabilities.

6.5 Educational Applications

  • Data-science learners face prompt-formulation and code-adaptation difficulties similar to those of data scientists; adoption challenges include user resistance and ineffective utilization.The section mentions active support during initial use and clear communication of system capabilities [252].Source: §6.5 · p. 53

6.5.1 Personalized Learning Support

  • OpenAI/DeepResearch develops learning plans that map knowledge structures and prerequisites to resources tailored to learners’ interests, gaps, styles, and objectives.The section characterizes these journeys as more personalized than standardized approaches and notes OpenManus emphasizes interest-driven exploration and adaptive difficulty.Source: §6.5 · p. 53
  • Perplexity/DeepResearch adjusts explanation detail and terminology to learner background, while Auto-Deep-Research uses knowledge modeling and explanation generation.The section describes this scaffolding as more targeted than generic responses.Source: §6.5 · p. 53

6.5.2 Educational Content Development

  • Gemini/DeepResearch supports curricula integrating learning-science principles, domain structure, and diverse resources; open frameworks emphasize progression modeling and resource alignment.The section frames this integration as more comprehensive than typically practical for individual educators.Source: §6.5 · p. 53
  • OpenAI/DeepResearch creates learning materials combining explanatory text, visualizations, examples, and assessments aligned with learning outcomes.Qwen-Agent is described as modeling learning objectives and generating multiple content formats.Source: §6.5 · p. 54

6.5.3 Academic Research Training

  • Perplexity/DeepResearch teaches research methods through demonstrations, rationale, and structured feedback on learner attempts.Jina-AI/node-DeepResearch supports research-practice environments with guidance and feedback mechanisms.Source: §6.5 · p. 54
  • Drosos et al. [71] report that carefully designed “provocations” can help restore critical thinking in AI-assisted knowledge work.The section describes guided source evaluation, evidence weighing, and analytical practice as ways systems support critical-evaluation training.Source: §6.5 · p. 54

6.6 Personal Knowledge Management Applications

Personal knowledge-management applications include tailored information abstraction, curiosity-driven learning, development planning, and support for individual decisions and life plans. Across these applications, the section identifies domain adaptation, specialized components, and workflow integration as common patterns.

6.6.1 Information Organization and Curation

  • OpenAI/DeepResearch turns complex sources into overviews, detailed analyses, and conceptual maps tailored to comprehension preferences.The section presents this as more targeted than generic summaries; Nanobrowser is described as using knowledge distillation and representation generation.Source: §6.6 · p. 55

6.6.2 Personal Learning and Development

  • Gemini/DeepResearch supports curiosity-driven discovery by identifying connections, extensions, and practical applications related to individual interests.OpenManus emphasizes interest mapping and discovery facilitation.Source: §6.6 · p. 55
  • Perplexity/DeepResearch plans skill development around learning resources, progression, practice, individual objectives, and constraints.TARS is described as using skill modeling and development-path generation.Source: §6.6 · p. 55

6.6.3 Decision Support for Individual Users

  • OpenAI/DeepResearch supports personal decision analysis using multiple criteria, preference weighting, and consequence projection tailored to individual values and constraints.Agent-RL/ReSearch emphasizes preference elicitation and consequence modeling.Source: §6.6 · p. 56
  • Gemini/DeepResearch integrates career, financial, health, and personal considerations into life plans tailored to individual values and objectives.Flowith/OracleMode is described as incorporating value modeling and multi-domain integration.Source: §6.6 · p. 56

6.6 Personal Knowledge Management Applications

  • Across applications, the section identifies domain adaptation, specialized component design, and integration with existing workflows as recurring patterns.It presents thoughtful alignment with domain requirements and user needs as the route by which technical capabilities translate into practical value.Source: §6.6 · p. 56

7 Ethical Considerations and Limitations

This section examines the ethical responsibilities and limitations involved in developing and deploying Deep Research systems, focusing on information accuracy, privacy, source attribution and intellectual property, and accessibility. It considers responsible information handling and safeguards alongside barriers to accessible design.

7.1 Information Accuracy and Hallucination Concerns

  • Deep Research systems face reliability challenges, including unclear uncertainty boundaries and misleading conversational responses that can amplify initial inaccuracies [55] [113].The concern is especially salient in research contexts where uncertainty is contested and workflows iteratively refine earlier outputs.Source: §7.1 · p. 56, §7.1.1 · p. 56

7.1.1 Factual Verification Mechanisms.

  • OpenAI/DeepResearch uses graduated confidence scores based on evidence quality, source consistency, and reasoning reliability to distinguish supported conclusions from speculation.The section also calls for better uncertainty representation, workflow design, and interfaces that communicate confidence boundaries [107] [270].Source: §7.1.1 · pp. 57–58
  • Verification approaches include cross-checking information across independent sources and automatically checking key claims against trusted references.Other implementations use citations linking claims to sources or specialized modules that assess source credibility and content consistency.Source: §7.1.1 · pp. 57–58

7.1.3 Quality Control Frameworks.

  • Gemini Deep Research verifies consistency, sources, and reasoning before delivering research outputs, while OpenAI/DeepResearch incorporates verified accuracy feedback into system improvements, supporting reliability through error identification and correction.Source: §7.1.3 · p. 59

7.2 Privacy and Data Security

  • Deep Research systems must protect sensitive information throughout the research process.The following subsection introduces comprehensive user-information protection mechanisms but provides no further details in this part.Source: §7.2 · p. 59
  • Strict session and tenant isolation, deployment-dependent controls, and data minimization limit exposure, with local deployment enabling stronger organizational isolation than cloud-dependent implementations.Collecting only information needed for service provision and limiting retention reduces potential exposure through security incidents or authorized access.Source: §7.2 · pp. 59–60

7.2.2 Sensitive Information Handling

  • Perplexity/DeepResearch detects and redacts personal identifiers from outputs unless they are relevant to research objectives, reducing inadvertent exposure.Open approaches vary, from basic detection of common identifiers to configurable sensitivity controls.Source: §7.2.2 · p. 60
  • Safeguards for health, financial, and other regulated information support domain-specific compliance, while jurisdiction-specific adaptations and configurable controls address differing regional requirements.The passages name GDPR and CCPA as examples of regional privacy frameworks.Source: §7.2.2 · p. 60
  • Transparency and user-control mechanisms support compliance and trust; Osogami [204] argues autonomous-system regulation should consider action sequences, and detailed access logs can support oversight.The action-sequence perspective is especially relevant to complex, multi-step research workflows.Source: §7.2.2 · pp. 60–61

7.3.1 Citation Generation and Verification

  • Commercial systems generate citations in standard academic formats, while citation-coverage checks can flag unsupported claims needing attribution.These mechanisms aim to keep significant claims connected to sources.Source: §7.3.1 · p. 61
  • Ethical attribution extends beyond facts to acknowledge conceptual contributions, including intellectual frameworks and theoretical approaches.The passages describe concept-level attribution as recognizing contributions beyond simple factual information.Source: §7.3.1 · p. 61
  • Attributing synthesized insights requires crediting multiple contributing sources while distinguishing sourced information from system-generated connections.The passages describe this as a challenge for cross-source synthesis.Source: §7.3.1 · pp. 61–62
  • Copyright handling remains uneven: commercial approaches evaluate purpose, nature, amount, and market impact, while the described open approach emphasizes acknowledgment.The fair-use factors are presented as supporting legal compliance while enabling legitimate research use.Source: §7.3.1 · p. 62

7.3.1 Citation Generation and Verification.

  • License-aware processing adapts source use to license terms, while clear output-rights frameworks address complex intellectual-property questions.Open implementations also use simplified license categories for common frameworks such as Creative Commons and commercial restrictions.Source: §7.3.1 · p. 62
  • Research systems must assess whether outputs are derivative works; commercial frameworks evaluate source transformation, while Qwen-Agent uses a basic assessment of reorganization and analytical additions.These assessments help align downstream use with source licenses.Source: §7.3.1 · p. 62

7.4.1 Technology Access Disparities.

  • Resource-intensive systems can exclude users lacking computing access; cloud services reduce local requirements but introduce costs, while efficient open systems can run on limited hardware.Camel-AI/OWL emphasizes efficiency to support broader deployment on constrained hardware.Source: §7.4.1 · p. 63
  • Financial requirements create socioeconomic access disparities: some commercial systems offer limited free access with premium tiers, while open-source alternatives remove direct fees but may impose technical hurdles.The passage names Perplexity/DeepResearch, HKUDS/Auto-Deep-Research, and nickscamara/open-deep-research as examples.Source: §7.4.1 · p. 63

7.4.2 User Expertise Requirements.

  • Technical complexity and domain-knowledge prerequisites can exclude users, but simplified deployment and domain guidance can reduce those barriers.OpenManus emphasizes simplified local deployment, while commercial and open-source systems increasingly offer domain guidance for users with limited background knowledge.Source: §7.4.2 · p. 63
  • Multilingual features can reduce language barriers, but quality disparities persist; extensible designs such as Flowith/OracleMode enable community-driven support beyond dominant languages.Language limitations particularly affect non-dominant language communities.Source: §7.4.2 · p. 64
  • Commercial systems increasingly provide screen-reader compatibility, keyboard navigation, and alternative formats, while uneven open-source accessibility leaves work to ensure equitable access.The passages identify continued community development as important across implementation contexts.Source: §7.4.2 · p. 64
  • Responsible development must address factual verification, attribution quality, privacy protection, intellectual-property respect, and accessible design across commercial and open-source systems.The section characterizes these responsibilities as extending beyond technical performance.Source: §7.4.2 · p. 64

8 Future Research Directions

The section surveys future research directions for Deep Research systems, identifying advanced reasoning architectures, multimodal integration, domain specialization, and human-AI collaboration with standardization as four key areas.

8 Future Research Directions

  • Zheng et al. [329] propose scaling deep research capabilities via reinforcement learning in real-world environments.The section cites this work as recent research relevant to future directions.Source: §8 · p. 64
  • Wu et al. [297] explore using tools to enhance LLM reasoning capabilities for deep research applications.The section cites their work as one example of research on deep research capabilities.Source: §8 · p. 64
  • Anthropic’s comprehensive framework for building effective agents offers design principles that could inform future Deep Research systems.The section presents it as an additional source of design principles.Source: §8 · p. 64
  • The four identified research areas are advanced reasoning architectures, multimodal integration, domain specialization, and human-AI collaboration with standardization.The section presents these as directions that could enhance capabilities, address current limitations, and expand practical impact across domains.Source: §8 · p. 64

8.1 Advanced Reasoning Architectures

This section outlines research directions in advanced reasoning for research systems, including improved context management, hybrid symbolic-neural reasoning, causal analysis, and richer uncertainty representation with probabilistic belief updates. These mechanisms are intended to support better information handling, reliability, domain-specific analysis, trust calibration, accuracy, and explainability.

8.1.1 Context Window Optimization and Management.

  • External memory and adaptive context management could address context limits while supporting richer knowledge integration and more transparent, reliable reasoning.Proposed mechanisms include hierarchical summaries, relevance-based context allocation, structured memory hierarchies, and retrieval integrated into reasoning.Source: §8.1.1 · pp. 65–66

8.1.2 Hybrid Symbolic-Neural Approaches.

  • Hybrid symbolic-neural systems could pair neural flexibility with formal logic to improve reasoning reliability and explainability.Suggested approaches include logical verification, task-dependent selection of reasoning methods, and translation between symbolic and neural representations.Source: §8.1.1 · p. 66
  • Dynamic knowledge graphs could be updated by research findings and used in reasoning, improving precision on tasks that require structured relationship understanding.Proposed mechanisms include graph-aware attention and hybrid graph traversal with neural generation.Source: §8.1.1 · p. 66

8.1.3 Causal Reasoning Enhancement.

  • Specialized causal reasoning could assess evidence and alternative explanations, supporting higher-quality research in medicine, social sciences, and policy analysis.The section proposes causal graphs that model interventions and counterfactuals, causal uncertainty quantification, and structured prompts.Source: §8.1.3 · pp. 66–67
  • Explicit intervention and counterfactual modeling could support what-if analysis, action planning, and outcome prediction for decision support.Proposed directions include counterfactual generation, intervention optimization, and domain-specific intervention templates.Source: §8.1.3 · p. 67

8.1.4 Uncertainty Representation and Reasoning.

  • Current systems’ uncertainty representations are relatively simplistic and inadequately capture different uncertainty types.The passage identifies multi-dimensional uncertainty modeling as a future research direction but provides no further details in this part.Source: §8.1.3 · p. 67

8.1.3 Causal Reasoning Enhancement

  • Uncertainty frameworks could distinguish epistemic, aleatoric, and model uncertainty; Perplexity/DeepResearch already distinguishes source uncertainty from integration uncertainty.More comprehensive approaches could support nuanced, reliable uncertainty communication.Source: §8.1.3 · p. 67
  • Propagation across reasoning chains, user-facing visualization, and uncertainty-aware planning could improve system reliability and calibrate user trust.Planning should balance different uncertainty types in decision contexts.Source: §8.1.3 · p. 67

Bayesian Reasoning Integration

  • Bayesian components could update beliefs using evidence strength and prior knowledge; grapeot/deep_research_agent provides an early example through basic evidence weighting.More sophisticated integration could support principled uncertainty handling across domains while enhancing accuracy and explainability.Source: §8.1.3 · pp. 67–68
  • Research directions include scalable Bayesian inference for large language models, explainable belief updates, and domain-specific priors for uncertain or evidence-limited fields.Field-specific background knowledge could inform common analysis patterns and improve reasoning quality in these domains.Source: §8.1.3 · p. 68

8.2 Multi-Modal Deep Research

The section identifies multimodal integration as an important direction for Deep Research, spanning visual, video, and audio sources. It proposes cross-modal reasoning, verification, and explanation methods to integrate these sources into research workflows.

Visual Information Integration

  • Current systems have limited capabilities for extracting and interpreting scientific images, while multimodal integration remains a key missing capability in AI research systems.Proposed work includes chart-data extraction, diagram understanding, and visual analysis tailored to fields such as medicine and astronomy.Source: §8.2 · p. 68

Multimodal Source Analysis

  • Video is an underutilized information source, and comprehensive audio integration into research workflows remains limited.Proposed components would extract temporal information from presentations or demonstrations, and spoken content from sources such as podcasts and lectures, while maintaining attribution and context.Source: §8.2 · p. 69

Cross-Modal Reasoning Techniques

  • Current reasoning typically operates within single modalities, motivating reasoning chains that transfer information across text, visual, and numerical analysis.The section describes Gemini/DeepResearch as an early step, with basic visual incorporation in reasoning steps.Source: §8.2 · p. 69
  • Integrating modalities introduces consistency challenges; proposed verification would detect cross-modal conflicts and align uncertainty across evidence types.Gemini/DeepResearch is described as implementing basic cross-format validation.Source: §8.2 · p. 70
  • Future research outputs could coordinate text, visuals, and interactive components, with formats adapted to content to improve communication effectiveness.The section cites mshumer/OpenDeepResearcher as an early example of report visualization.Source: §8.2 · p. 70

8.3 Domain-Specific Optimization

The section argues for adapting Deep Research systems to domain-specific needs, including scientific workflows and specialized legal, regulatory, and medical tasks. It highlights varied researcher preferences for automation and continuing limits in regulatory compliance and patient-specific personalization.

8.3 Domain-Specific Optimization

  • Domain-specific implementations highlight the value of adapting research systems beyond general research capabilities.The section identifies field-tailored enhancement as a route to performance improvements for specialized domains.Source: §8.3.1 · p. 70, §8.3 · p. 70

8.3.1 Scientific Domain Adaptation.

  • Scientific adaptation could combine field-specific fine-tuning, architectural changes, or symbolic components with interfaces for research workflows.Workflows may include experimental design, data analysis, literature integration, and theory development.Source: §8.3.1 · pp. 70–71
  • User studies report varying automation needs across data-science and machine-learning workflows, with targeted rather than complete end-to-end automation aligning with researcher preferences [284].The section proposes assistants for experimental design and pipelines combining automated and human analysis.Source: §8.3.1 · p. 71
  • Legal systems struggle with the precision and structure of legal analysis; proposed specialization includes case-based reasoning, statutory interpretation, and doctrinal analysis.OpenAI/DeepResearch is described as incorporating basic legal language handling.Source: §8.3.1 · p. 71
  • Regulatory compliance requires comprehensive coverage and exceptional precision, while compliance optimization remains limited.Proposed components include tracking regulatory changes, extracting obligations, and mapping responsibilities to organizational processes.Source: §8.3.1 · p. 71
  • Medical research support calls for evidence synthesis across studies and guidelines, while comprehensive patient-specific personalization remains limited.Proposed approaches include evidence grading using GRADE, quantitative meta-analysis, and adapting findings to patient characteristics and preferences; the section emphasizes caution in this high-stakes domain.Source: §8.3.1 · p. 72

8.4 Human-AI Collaboration and Standardization

This section presents human-AI collaboration and ecosystem standardization as complementary directions for advancing Deep Research, covering interactive, expertise-aware workflows, shared knowledge creation and planning, system evaluation, interoperability, and broader participation.

8.4.1 Interactive Research Workflows

  • Research-question interfaces could iteratively clarify, expand, or focus questions using initial results and user feedback, including ambiguity detection and scope adjustment.The section also proposes perspective-diversification tools to surface alternative viewpoints.Source: §8.4 · pp. 72–73
  • Because systems typically provide static outputs, research interfaces could support dynamic navigation, adjustable detail, and transitions to sources for more exploratory discovery.The section describes OpenManus as an early example with basic exploration capabilities.Source: §8.4 · p. 73

8.4.2 Expertise Augmentation Models

  • Expertise-adaptive systems could personalize research approaches and explanations by assessing user knowledge, identifying knowledge gaps, and requesting expert help beyond autonomous capability.The section cites Perplexity/DeepResearch as an early example of basic terminology adjustment.Source: §8.4.2 · p. 73
  • Complementary designs could use AI for information processing while reserving subjective evaluation and contextual understanding for people, supported by explanations, confidence signals, and critique.Agent-RL/ReSearch is described as an early example of dividing analytical responsibilities.Source: §8.4.2 · pp. 73–74

8.4.3 Framework Standardization Efforts

  • Component standards can support modular substitution and consistent integration: OpenAI/AgentsSDK specifies agent interfaces, MCP standardizes model-tool interaction, and A2A standardizes agent communication.Source: §8.4.2 · p. 74
  • Evaluation practices vary widely, and comprehensive standardization remains limited despite early benchmarks such as HLE [212] and MMLU.The section calls for shared methodologies and multidimensional assessments beyond simple accuracy to support consistent comparison and improvement signals.Source: §8.4.2 · p. 74

8.4.4 Cross-Platform Research Protocols

  • Research systems produce incompatible outputs, and basic document formats have not yielded research-specific standards for sharing findings across platforms.Proposed formats include attributed findings with confidence metadata, shared evidence representations, and schemas for questions and objectives.Source: §8.4.2 · pp. 74–75
  • Research-specific coordination across systems remains limited; future frameworks could allocate tasks, integrate results, manage processes, and coordinate distributed search and verification.The section cites n8n as an early example of workflows enabling coordination.Source: §8.4.2 · p. 75

8.4.5 Joint Human-AI Knowledge Creation

  • Collaborative research could move beyond retrieval through transparent reasoning, prompt-as-instrument interfaces [232], and agents learning assistance behaviors by observation [127].The passages describe transparent reasoning as supporting collaboration and prompt-based instruments as a way to make interactions more intuitive.Source: §8.4.2 · p. 75
  • Writing support could extend generation with context-aware refinement [248] and iterative fill-in-the-middle co-editing [302], reducing friction in joint content development [116].The passages also envision seamless revision across human and AI contributions.Source: §8.4.2 · p. 75
  • Future writing systems could suggest context-based expansions, match document style, and use rejected suggestions as feedback while preserving original intent [271].The passages propose seamless revision interfaces for editing human and AI contributions.Source: §8.4.2 · p. 75

8.4.2 Expertise Augmentation Models

  • Distributed research workflows could improve through specialization and parallelization, supported by coordination protocols that limit communication overhead.The passage fragments also refer to maintaining consistency across distributed findings.Source: §8.4.2 · p. 75
  • Mixed-initiative frameworks could balance human preferences and AI-identified opportunities in research planning; smolagents/open_deep_research offers an early suggestion-mechanism example.The proposed direction is shared determination of research direction and approach.Source: §8.4.2 · p. 76
  • Research-planning tools could identify unexplored opportunities, visualize path trade-offs, elicit evolving preferences, and explain AI decision logic through reward mechanisms [239].These approaches aim to combine human insight with AI-identified opportunities in balanced partnerships.Source: §8.4.2 · p. 76
  • Future Deep Research progress is framed as complementary advances in reasoning, multimodal capabilities, domain specialization, collaboration, and standardization, with open-source and academic work supporting broad participation.The passages say commercial implementations will drive innovation while open-source alternatives and academic research help expand what is possible.Source: §8.4.2 · p. 76

9 Conclusion

The conclusion traces the emerging Deep Research ecosystem, its architectural patterns, implementation approaches, and application opportunities. It considers the systems’ potential role in research and knowledge work alongside their technical, ethical, collaboration, and access challenges.

9.1 Key Findings and Contributions

  • Effective systems show varied foundation-model strategies: commercial platforms use proprietary models with extensive context and reasoning, while open-source projects can deliver research capabilities with more accessible models.Examples include OpenAI/DeepResearch and Gemini/DeepResearch, alongside Camel-AI/OWL and QwenLM/Qwen-Agent.Source: §9.1 · p. 76
  • Environmental interaction ranges from specialized web navigation and extraction to broad access across multiple environments.Nanobrowser and dzhng/deep-research illustrate specialization; Manus and AutoGLM-Search provide broader interaction capabilities.Source: §9.1 · p. 77
  • Planning and execution approaches vary, while knowledge synthesis consistently emphasizes information evaluation but differs in presentation and interactivity.OpenAI/AgentsSDK and Flowith/OracleMode offer sophisticated planning; Agent-RL/ReSearch emphasizes execution reliability, and smolagents/open_deep_research collaborative approaches.Source: §9.1 · p. 77
  • Commercial systems emphasize optimized performance and comprehensive capabilities, whereas open-source systems offer architectural diversity and customization with greater deployment complexity and typically lower benchmark performance.Commercial platforms also entail costs and customization limitations; open-source alternatives can offer greater control and potentially lower operational costs for specialized or high-volume use.Source: §9.1 · p. 77
  • Application adaptations address field-specific needs, including literature coverage and citation quality in academic research, experimental design in science, and actionable insights in business and education.The survey identifies these adaptations as the means by which general deep-research capabilities translate into domain value.Source: §9.1 · pp. 77–78
  • Ethical priorities include accuracy, privacy, intellectual property, and accessibility, with implementation approaches varying across systems and constraints.Commercial systems such as OpenAI/DeepResearch and Perplexity/DeepResearch use multi-level verification and explicit attribution, while open-source approaches differ in sophistication and deployment model.Source: §9.1 · p. 78

9.2 Limitations and Outlook

  • The survey captures an early-2025 snapshot of a rapidly changing field, and limited transparency—especially for commercial systems—constrains complete architectural comparison.Its classification framework provides a structural foundation for continued assessment as capabilities and implementations evolve.Source: §9.2 · p. 78
  • Real-world impact remains difficult to assess because many systems are early in deployment; assessing realized effects requires longer-term observation beyond this survey’s scope.The survey notes promising initial capabilities across academic research, business intelligence, and education, while broader effects remain partly speculative.Source: §9.2 · p. 79

9.3 Broader Implications

  • Deep Research may enable new research methods, accelerate discovery, and complement human insight, pointing toward collaborative research rather than replacing researchers; overreliance can compromise data quality.Ashktorab et al. [15] note that users may append AI-generated responses even when they conflict with other information.Source: §9.3 · p. 79
  • Commercial access tiers and open-source implementations may broaden access to sophisticated research capabilities and reduce barriers to high-quality information access and analysis.The survey links this potential democratization to education, entrepreneurship, civic participation, and individual knowledge development, while noting that technical expertise and computational resources remain barriers.Source: §9.3 · p. 79
  • Deep Research may support collective intelligence by integrating knowledge, sharing insights, and enabling cross-organizational and cross-disciplinary discovery that addresses fragmentation in complex domains.The survey also identifies potential benefits for education, entrepreneurship, civic participation, and knowledge equity, although access barriers remain.Source: §9.3 · p. 80
  • Embedding these systems in collaborative knowledge ecosystems may enhance collective sense-making, evidence-based decisions, and shared understanding beyond individual productivity gains.These potential effects emphasize the social and organizational dimensions of Deep Research.Source: §9.3 · p. 80

9.4 Final Thoughts

  • The ecosystem spans diverse technical approaches and application domains, with competition, specialization, and complementary development contributing to its health.The survey describes this ecosystem as including commercial platforms, open-source alternatives, and specialized components.Source: §9.4 · p. 80
  • Research on advanced reasoning, multimodal capabilities, domain specialization, human-AI collaboration, and ecosystem standardization may build on current implementations and expand research capabilities across domains.The survey anticipates implications for how information is discovered, validated, synthesized, and used.Source: §9.4 · p. 80
  • Responsible development requires attention to information accuracy, privacy, intellectual property, and accessibility alongside technical advancement to realize benefits while minimizing harms.The survey identifies these as ethical considerations for the Deep Research ecosystem.Source: §9.4 · p. 80
  • The survey’s frameworks, analysis, and directions provide a foundation for further examination of Deep Research as a technical field and potentially transformative capability for practical knowledge work.The conclusion highlights implications for future information access and knowledge synthesis.Source: §9.4 · p. 80
Loading 2506.12594v1…