Deep Research: A Systematic Survey

1. Introduction

• Deep Research frames LLMs in an iterative workflow that decomposes complex problems, acquires evidence with tools, and synthesizes validated insights into long-form answers.The motivation is that many real-world tasks require critical thinking, grounded information, and self-contained responses beyond single-shot prompting or static parametric knowledge.Source: §1 · p. 5
• The survey addresses the lack of a comprehensive, systematic account of Deep Research components, technical details, and open challenges.It distinguishes Deep Research from conventional RAG and synthesizes representative systems, techniques, and evaluation methods.Source: §1 · p. 5
• The survey formalizes a three-stage roadmap and four system components: query planning, information acquisition, memory management, and answer generation.Answer generation produces comprehensive outputs with explicit source attribution.Source: Figure 1 · p. 5, §1 · p. 6
• The survey organizes optimization approaches into workflow prompting, supervised fine-tuning, and end-to-end reinforcement learning, and consolidates evaluation criteria and open challenges.It aims to guide future research and support comparable reporting.Source: §1 · p. 6

2.1. What is Deep Research

• Deep Research plans queries, acquires and filters evidence from heterogeneous sources, revises working memory, and synthesizes verifiable answers with attribution.It is described as an end-to-end research workflow intended to produce coherent reports with minimal human supervision.Source: §2.1 · p. 6

2.2. Understanding Deep Research from Three Phases

• Phase I prioritizes finding correct sources and extracting answers with minimal synthesis, emphasizing faithfulness, predictable runtime, and retrieval and answer accuracy.Systems reformulate queries, retrieve and rerank documents, and may filter or compress results before producing concise answers with citations.Source: §2.2 · pp. 6–8
• Phase II integrates heterogeneous evidence into coherent reports, trading increased compute and complexity for gains in clarity, coverage, and decision support.Its evaluation emphasizes factuality, verified citations, structural coherence, and coverage of key points.Source: §2.2 · p. 8
• Phase III extends evidence aggregation toward hypothesis generation, experimental validation, critique, and novel perspectives, with evaluation emphasizing novelty, insight, reproducibility, and calibrated uncertainty.The survey presents the phases as a capability trajectory from precise evidence to readable analyses and defensible insights.Source: §2.2 · pp. 6, 8

2.3. Comparing Deep Research with RAG

• Unlike conventional RAG’s static retrieval loop, Deep Research supports dynamic tool interaction, long-horizon planning, and mechanisms that ground open-ended language outputs in evidence.The survey identifies active interaction with digital environments, iterative workflow control, and more reliable language interfaces as challenges conventional RAG or parameter scaling alone cannot solve.Source: §2.3 · p. 9

3. Key Components in Deep Research System

• The workflow cycles through query planning, knowledge acquisition, memory management, and answer generation to produce coherent, verifiable responses with citations.These components respectively decompose questions, retrieve and filter information, maintain intermediate findings, and synthesize evidence while checking consistency.Source: §3 · p. 9

3.1.1. Parallel Planning

• Parallel planning generates independent sub-queries in one pass for efficient execution, but cannot adapt to intermediate evidence and may miss dependencies.RL-based approaches instead train planners against downstream answer accuracy or other task metrics.Source: §3.1.1 · p. 10

3.1.2. Sequential Planning

• Sequential planning iteratively uses prior outputs to reformulate queries and handle dependencies, but long reasoning chains can increase cost, latency, noise, and error propagation.It supports stepwise inference, disambiguation, and progressive information gathering.Source: §3.1.2 · p. 11

3.1.3. Tree-based Planning

• Tree-based planning organizes sub-queries as a tree or DAG, combining branching exploration with local parallelism, but robust training requires dependency modeling and careful speed-quality trade-offs.The structure can support fine-grained decomposition and search over reasoning paths.Source: §3.1.3 · p. 12

3.2.1. Retrieval Tools

• Retrieval tools span lexical, semantic, and commercial web search, while multimodal methods retrieve text, layout, visuals, or structured chart and table data for grounded evidence pointers.Multimodal retrieval can capture information absent from text-only methods, but adds computational cost and cross-modal alignment challenges.Source: §3.2.1 · pp. 13–14

3.2.2. Retrieval Timing

• Adaptive retrieval triggers searches when the model lacks sufficient knowledge, with strategies based on confidence, response consistency, internal states, verbalized uncertainty, or reasoning policies.Retrieval timing matters because retrieval adds overhead and low-quality or irrelevant documents may mislead the model.Source: §3.2.2 · pp. 14, 16

3.2.3. Information Filtering

• Information filtering selects documents, compresses context, or cleans retrieved content to reduce noise, but added filtering can increase latency or remove useful evidence.Document selection includes point-wise, pair-wise, and list-wise strategies; compression includes lexical and embedding-based methods.Source: §3.2.3 · pp. 17–18

3.3.1. Memory Consolidation

• Consolidation transforms transient interactions into durable memories, either as concise summaries or as structured databases, graphs, and trees.Unstructured approaches summarize histories or events; structured approaches encode relationships and organize queryable memory.Source: §3.3.1 · p. 20

3.3.2. Memory Indexing

• Indexing builds navigable retrieval pathways over consolidated memories using auxiliary signals and structures that can capture semantic, relational, and temporal context.Examples include metadata-enhanced, graph-based, and timeline-based indexing.Source: §3.3.2 · p. 21

3.3.3. Memory Updating

• Non-parametric updating modifies external memory entries through operations such as adding, revising, merging, or reflecting on stored knowledge.These explicit operations incorporate new information and resolve inconsistencies.Source: §3.3.3 · pp. 21–22
• Parametric updating changes model-internal memory through global training, localized fact edits, or dedicated modules that can avoid modifying base-model parameters.Global updating can be computationally expensive and prone to catastrophic forgetting; modular updating avoids that risk and can reduce retrieval latency.Source: §3.3.3 · p. 22

3.3.4. Memory Forgetting

• Forgetting removes or suppresses outdated, irrelevant, or erroneous content, with passive decay and active deletion or invalidation as distinct mechanisms.Passive approaches include FIFO transfer to long-term storage; active forgetting can target external stores or model parameters.Source: §3.3.4 · p. 23

3.3. Memory Management

• These four interdependent operations govern context across long-horizon investigations and support the coherence needed for an iterative research loop.Memory management distinguishes DR agents from simple RAG systems by supporting consistency, adaptability, and self-evolution.Source: §3.3 · pp. 19–20, §3.3.4 · p. 24

3.4. Answer Generation

• Answer generation combines upstream information into a coherent, comprehensive, well-supported response that reflects the user’s intent.Its challenges include resolving conflicting evidence, maintaining long-range coherence, and providing transparent reasoning and citations.Source: §3.4 · p. 24

3.4.1. Integrating Upstream Information

• Upstream integration combines planned sub-queries, ranked and potentially conflicting evidence, and evolving memory; some systems use that state to guide later planning.Self-RAG adaptively retrieves passages and uses reflection tokens, while Plan-on-Graph and MCTS-OPS connect planning with evolving memory.Source: §3.4.1 · pp. 24–25, Figure 6 · p. 25

3.4.2. Synthesizing Evidence and Maintaining Coherence

• Conflicting evidence calls for reliability-sensitive synthesis, while long-form answers must preserve accuracy and information density without losing coherence.Approaches include credibility-aware attention, multi-agent deliberation, factuality rewards, longer SFT examples, reflection, and length-adaptive rewards.Source: §3.4.2 · pp. 25–26

3.4.3. Structuring Reasoning and Narrative

• Structured reasoning makes answers more interpretable, using prompted intermediate steps, explicit answer plans, or dynamically invoked external tools.RAPID uses an outline refined through evidence discovery and plan-guided writing; SuperWriter separates reasoning from text production.Source: §3.4.3 · pp. 26–27

3.4.4. Presentation Generation

• Presentation generation extends beyond text to visualizations, synchronized audio, editable slide decks, and video-grounded storytelling.Most DR systems still focus on cited text, while only a handful are described as supporting comprehensive multimodal output.Source: §3.4.4 · p. 27

4. Practical Techniques for Optimizing Deep Research Systems

• Workflow prompting uses a carefully designed pipeline to guide DR agents.The paper also describes supervised fine-tuning and agentic reinforcement learning as optimization paradigms.Source: §4 · pp. 27–28
• Table 2 compares output capabilities across contemporary representative deep research systems.Its caption indicates supported capabilities but does not report specific comparisons in the supplied passage.Source: Table 2 · p. 28
• Supervised fine-tuning and agentic reinforcement learning aim to train DR agents to reason, retrieve information, and generate high-quality answers.These paradigms differ from workflow prompting, which guides agents through a designed pipeline.Source: §4 · pp. 27–28

4.1.1. Deep Research System of Anthropic

• A lead orchestrator coordinates specialized worker agents in a workflow spanning planning, delegation, synthesis, and citation under a research budget.The budget controls agent count, tool usage, and reasoning depth.Source: §4.1.1 · p. 28, §4.1 · p. 28
• Agent effort scales with query complexity, ranging from 1–2 agents for factual lookups to up to 10 or more for multi-perspective analyses.The orchestrator assigns quotas and stopping criteria to support dynamic budget reallocation.Source: §4.1.1 · p. 28
• The workflow decomposes queries into prompted subtasks and uses tool-selection rules and evidence logging to support traceable attribution.Prompts specify objectives, output schemas, citation policies, and fallback actions; tool choices consider freshness, verifiability, and latency.Source: §4.1.1 · p. 28
• Figure 7 compares strong-to-weak distillation, multi-agent distillation, and iterative self-evolving through how agents perform tasks, learn, and refine abilities.The caption frames these as three data-synthesis approaches.Source: Figure 7 · p. 29
• The orchestrator integrates verified findings into a coherent report, linking claims to sources and checking schema compliance and factual grounding.The report includes transparent citation.Source: §4.1.1 · p. 29
• Anthropic’s system is characterized as a scalable, interpretable multi-agent paradigm built around modular delegation, explicit budgeting, and verifiable reasoning.The survey describes it as achieving high-quality synthesis.Source: §4.1.1 · p. 29

4.2. Supervised Fine-Tuning

• SFT commonly serves as a warm-up before online reinforcement learning, giving DR agents basic task-solving skills.Manually collecting expert trajectories is labor-intensive, costly, and difficult to scale, motivating automatic dataset construction.Source: §4.2 · p. 29
• The survey groups automatic SFT-data construction into strong-to-weak distillation and iterative self-evolution, in which models train on data they produce themselves.Self-evolution is described as a cyclic process of generating training data and fine-tuning, with DR work such as EvolveSearch selecting high-performing rollouts for SFT.Source: §4.2 · p. 29, §4.2.2 · pp. 30–31

4.2.1. Strong-to-weak Distillation

• Strong-to-weak distillation transfers decision trajectories from powerful teacher systems to smaller student models.Single-agent teachers offer a simple pipeline but may produce shallow trajectories; multi-agent teachers can provide more diverse, complex trajectories covering planning, decomposition, correction, and reflection.Source: §4.2.1 · pp. 29–30
• Multi-agent distillation can broaden behavioral coverage, but requires careful design, substantial inference, and infrastructure for logging and verification.Its data quality can also be sensitive to prompting.Source: §4.2.1 · p. 30

4.2.2. Iterative Self-Evolving

• Iterative self-evolution interleaves data generation and training, enabling scalable training with less reliance on external models or human annotations.The approach can explore data distributions beyond handcrafted knowledge.Source: §4.2.2 · p. 31
• Across iterations, distributional drift, reward hacking, and self-reinforcing errors can degrade data quality, while weak validation may yield premature convergence to narrow modes.These risks can potentially lead to training collapse or limited performance ceilings.Source: §4.2.2 · p. 31

4.3.1. Preliminary

• PPO uses a value-model-derived advantage for each output, whereas GRPO compares responses within a group, simplifying implementation and reducing resource requirements.PPO relies on accurate value estimation and requires resources to train the value model.Source: §4.3.1 · p. 33
• DR reward design commonly uses task-specific rule-based scores or an LLM judge that scores output quality against a rubric.Rule-based rewards suit short, well-defined answers but struggle with multi-answer and open-ended questions; judge criteria can include accuracy, completeness, citation quality, and clarity.Source: §4.3.1 · p. 33

4.3.2. End-to-end Optimization of a Specific Module

• Single-module optimization targets a component such as the planner, potentially improving credit assignment while reducing training-data and computational costs.Keeping other modules fixed restricts the optimization space and may leave their design or performance flaws unaddressed.Source: §4.3.2 · p. 34

4.3.3. End-to-end Optimization of an Entire Pipeline

• End-to-end pipeline optimization jointly trains the DR workflow, spanning query decomposition, search, reading, and report generation.The survey contrasts it with optimizing a single module while leaving the others fixed.Source: §4.3.3 · pp. 34–35
• Full-pipeline methods optimize reasoning, query rewriting, retrieval, tool invocation, and answer generation, with reward functions allowing different objectives to be emphasized.The survey cites work on multi-hop search, long-chain web search, and multi-tool scenarios as examples of this approach.Source: §4.3.3 · pp. 34–35
• Full-pipeline optimization faces sparse rewards, excessively long responses, and unstable training, requiring continued work on effectiveness, stability, and efficiency.The survey also summarizes methods using multi-reward designs and simulated or real-world web environments.Source: §4.3.3 · pp. 34–35

5. Evaluation of Deep Research System

• DR evaluation covers information seeking and report generation, with each scenario presenting distinct challenges.The survey frames these as representative application areas for DR systems.Source: §5 · p. 36

5.1. Agentic Information Seeking

• Information-seeking benchmarks progress from static single-hop retrieval toward multi-hop reasoning and complex interactions in dynamic web environments.The survey reviews benchmarks along query complexity and interaction-environment complexity.Source: §5.1 · p. 36, §5.1.1 · p. 36, §5.1.2 · p. 38

5.1.1. Complex Queries

• Many multi-hop benchmarks assess cross-document reasoning, but their structured, linear paths do not capture the ambiguity and branching exploration of real-world research.Newer datasets such as GAIA and HLE aim to better reflect ambiguity, branching evidence paths, and iterative synthesis.Source: §5.1.1 · p. 38

5.1.2. Interaction Environment

• Interactive evaluation increasingly uses authentic, dynamic web environments and tasks requiring persistent navigation or thorough information collection.Benchmarks also include reproducible environments and sandboxes for evaluating web research tasks.Source: §5.1.2 · pp. 38–39

5.2. Comprehensive Report Generation

• Comprehensive report generation tests integration of multiple sources into structured, coherent outputs, including aggregation, organization, factual consistency, and clarity.Survey-generation benchmarks assess content and citation quality, while long-report benchmarks use reference-based, citation-based, rubric, or factual assessments.Source: §5.2 · p. 39, §5.2.1 · p. 39, §5.2.2 · p. 39

5.2.2. Long-Form Report Generation

• Long-form report evaluation lacks a single gold-standard answer, so benchmarks commonly rely on LLM-as-judge methods to assess content, factual accuracy, citations, and structure.Acceptable content, reasoning approaches, and presentation styles vary.Source: §5.2.2 · p. 40

5.2.3. Poster Generation

• Poster generation combines multi-source synthesis with visual layout, and evaluation commonly considers factual completeness, visual communication, and readability.Paper2Poster and PosterGen use evaluation dimensions that include visual quality or design as well as content-related criteria.Source: §5.2.3 · p. 40

5.2.4. Slides Generation

• Slide-generation benchmarks assess information selection and organization while checking content integrity and factual accuracy.Some support both reference-based similarity assessment and reference-free evaluation of design quality.Source: §5.2.4 · pp. 40–41

5.3. AI for Research

• AI for Research covers idea generation, experimental execution, academic writing, and peer review, where evaluation must address novelty as well as output quality.The survey distinguishes these tasks from report generation because they seek new perspectives, conclusions, and knowledge.Source: §5.3 · p. 41

5.3.1. Idea Generation

• Idea novelty is difficult and resource-intensive for experts to assess, and LLM novelty judgments have shown lower agreement with expert reviewers than human evaluations.Approaches include human or LLM review, literature-informed LLM evaluation, and embedding-based density methods.Source: §5.3.1 · p. 41

5.3.2. Experimental Execution

• Experimental execution is evaluated through objective outcomes such as benchmark performance or compiler output, as well as subjective human or LLM judgments of design quality.Some evaluations compare code implementations against atomic research ideas or apply human-designed rubrics to code and related materials.Source: §5.3.2 · pp. 41–42

5.3.3. Academic Writing

• Academic-writing evaluation considers factual accuracy, coherence, clarity, alignment with ideas and results, citation integration, and visualization.Benchmarks include rubric-based assessment and factual evaluation of insights, citation faithfulness, and evidence coverage.Source: §5.3.3 · p. 42

5.3.4. Peer Review

• Peer-review evaluation is challenging because reviews are lengthy and subjective, so benchmarks combine structured datasets with automatic metrics and human or LLM judgments.Datasets capture review dimensions, scores, and in some cases rebuttal-stage discussions.Source: §5.3.4 · pp. 42–43

5.4. Software Engineering

• Software-engineering DR shifts evaluation from isolated code snippets toward agents resolving real-world issues and executing complex development tasks.SWE-Bench evaluates resolution of real-world GitHub issues but does not yet cover full end-to-end software development.Source: §5.4 · p. 43

6.1. Retrieval Timing

• Answer-correctness rewards can cause over- or under-retrieval, and answering without relevant evidence risks misleading users, especially in safety-critical domains.The survey proposes stepwise rewards for knowledge gaps and retrievability, plus checks of post-retrieval correctness and final uncertainty.Source: §6.1 · p. 43

6.2. Memory Evolution; 6.2.1. Proactive Personalization Memory Evolution

• Memory is often a passive archive of interactions, while future systems are envisioned as predictive user models that support proactive reasoning and personalization.Examples include profiles combining preferences and past interactions, and systems that use intent or dynamic profiles to guide conversations.Source: §6.2 · p. 43, §6.2.1 · p. 44

6.2.2. Cognitive-Inspired Structured Memory Evolution

• Flat text-chunk stores lack explicit relations and temporal updates; knowledge graphs and bi-temporal models offer structures for relational reasoning and preserving historical context.The survey proposes continuously transforming, expanding, pruning, and reorganizing structured memory as information arrives.Source: §6.2.2 · pp. 44–45

6.2.3. Goal-Driven Reinforced Memory Evolution

• Heuristic memory retention may not align stored information with task goals, motivating reinforcement learning policies for memory operations based on final task outcomes.Extending this approach across storage, retrieval, and use remains difficult because long-term credit assignment must connect final outcomes to intermediate decisions.Source: §6.2.3 · p. 45

6.3. Instability in Training Algorithms; 6.3.1. Existing Solutions

• Multi-turn reinforcement learning can collapse through void turns or the Echo Trap, while trajectory filtering methods are proposed to stabilize training.SimpleTIR filters trajectories with task-irrelevant turns; StarPO-S filters for trajectories showing meaningful exploration.Source: §6.3 · p. 45, §6.3.1 · p. 46

6.3.2. Future Directions

• Future training directions include cold-start methods that preserve exploration and denser, smoother rewards, particularly for GRPO-style algorithms.SFT may reduce output entropy, while GRPO is more sensitive to reward variance and extreme values than PPO, whose critic smooths rewards.Source: §6.3.2 · p. 46

6.4. Evaluation of Deep Research System; 6.4.1. Logical Evaluation

• Long-form DR evaluation remains difficult: models may fail to connect evidence into coherent arguments, and existing logical benchmarks do not capture consistency across extended contexts.The survey suggests assessing coherence at sentence, paragraph, and document levels.Source: §6.4 · pp. 46–47, §6.4.1 · p. 47

6.4.2. Boundary between Novelty and Hallucination

• Novel outputs can contain unverifiable claims or fabricated connections, so novelty scores and judge assessments alone do not establish verifiability.Suggested checks include linking claims to sources, testing false information, and validating claims against sources published after a cutoff date.Source: §6.4.2 · pp. 47–48

7.1. Creativity

• LLM creativity may extend beyond recombination toward genuine innovation, but proposed links between hallucination and creativity require distinguishing productive divergence from erroneous output.The survey notes that mechanisms such as insight or sudden restructuring of mental representations may not be readily explained by probabilistic models.Source: §7.1 · p. 48

7.2. Fairness

• Autonomous research agents could amplify academic biases by favoring mainstream fields, methods, or prominent researchers and overlooking emerging or non-mainstream contributions.The discussion calls for fairness checks across decision steps because early bias may shape later exploration and perpetuate inequities.Source: §7.2 · p. 49

7.3. Safety and Reliability

• To support safety and reliability, deep research systems should provide traceable evidence chains, transparent reasoning, and validation mechanisms to curb hallucinated science.Source: §7.3 · p. 49

• The survey consolidates advances in deep research architectures, optimization, and evaluation into a roadmap intended to facilitate future research.It plans continued updates covering emerging paradigms such as multimodal reasoning, self-evolving memory, and agentic reinforcement learning.Source: §8 · p. 49