Source-linked AI summary
WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research
Zijian Li, Xin Guan, Bo Zhang, Shen Huang, Houquan Zhou, Shaopeng Lai, Ming Yan, Yong Jiang, Pengjun Xie, Fei Huang, Jun Zhang, Jingren Zhou
TL;DR
Open-ended deep research requires agents to synthesize web-scale information without ground-truth answers, while existing systems separate planning from discovery or overload generation with irrelevant context. WebWeaver addresses this with a dual-agent planner–writer framework that co-evolves evidence and outlines, then uses citation-guided retrieval for focused report synthesis. It achieves state-of-the-art results across major open-ended deep research benchmarks.
Problem
Open-ended deep research asks agents to synthesize vast web-scale information into reports without ground-truth answers, while existing pipelines decouple planning from evidence acquisition and generation.
Method
WebWeaver uses a planner that iteratively interleaves evidence acquisition with outline optimization and a writer that performs hierarchical, citation-guided retrieval and section-wise synthesis.
Results
WebWeaver achieves state-of-the-art performance and outperforms proprietary and open-source agents across three challenging open-ended deep research benchmarks.
Takeaways & Limitations
The results support adaptive planning and focused, memory-grounded synthesis as an effective framework for producing comprehensive, trusted reports.
Abstract
from arXiv · showhide
This paper tackles \textbf{open-ended deep research (OEDR)}, a complex challenge where AI agents must synthesize vast web-scale information into insightful reports. Current approaches are plagued by dual-fold limitations: static research pipelines that decouple planning from evidence acquisition and monolithic generation paradigms that include redundant, irrelevant evidence, suffering from hallucination issues and low citation accuracy. To address these challenges, we introduce \textbf{WebWeaver}, a novel dual-agent framework that emulates the human research process. The planner operates in a dynamic cycle, iteratively interleaving evidence acquisition with outline optimization to produce a comprehensive, citation-grounded outline linking to a memory bank of evidence. The writer then executes a hierarchical retrieval and writing process, composing the report section by section. By performing targeted retrieval of only the necessary evidence from the memory bank via citations for each part, it effectively mitigates long-context issues and citation hallucinations. Our framework establishes a new state-of-the-art across major OEDR benchmarks, including DeepResearch Bench, DeepConsult, and DeepResearchGym. These results validate our human-centric, iterative methodology, demonstrating that adaptive planning and focused synthesis are crucial for producing comprehensive, trusted, and well-structured reports.
1 Introduction
Open-ended deep research requires agents to synthesize web-scale information without ground-truth answers, but existing pipelines decouple searching, outlining, and writing. WebWeaver addresses this with a human-centric planner–writer framework that co-evolves research and outlines, then produces citation-grounded reports through targeted evidence retrieval.
- Open-ended deep research asks agents to search relevant information and produce reports for questions without ground-truth answers.
- On DeepResearch Bench, WebWeaver achieves state-of-the-art RACE performance and exceeds reference answers, while also leading effective citations and citation accuracy on FACT.
- Existing approaches either search before generating or rely on static outlines, limiting coherence, adaptability, or research scope.
- WebWeaver’s planner interleaves evidence acquisition with outline optimization, allowing discoveries to reshape subsequent searches and sections.
- The writer performs section-level hierarchical synthesis by retrieving only evidence linked through outline citations from a structured memory bank.
- WebWeaver achieves state-of-the-art performance against proprietary and open-source agents on three open-ended deep research benchmarks.
2 Preliminaries
The framework models open-ended research as repeated thought, action, and observation for separate planning and writing agents. A planner gathers and structures evidence in memory, while a writer retrieves cited evidence to compose the report.
- Open-ended research uses a planner, memory, and writer because questions lack ground-truth answers and require information search followed by report generation.
- Both agents follow ReAct-style iterations in which thoughts produce parsable actions and observations update subsequent context.
- The planner can search, write or optimize an outline, and terminate; search processing also selects URLs, parses pages, summarizes content, and extracts evidence.
- The writer retrieves memory-bank evidence through grounded outline citations, writes report sections, and terminates when the report is complete.
- The memory bank addresses long-context demands because planning may process over 100 web pages and 100k tokens, while reports can exceed 20k tokens.
3 Method
WebWeaver combines a planner’s adaptive research cycle with a writer’s memory-grounded, section-wise synthesis. The planner continually updates evidence and outlines, while the writer retrieves and prunes context to maintain focused, coherent generation.
- 3.1 Overview of WebWeaver: WebWeaver divides work between a planner that acquires evidence and optimizes outlines and a writer that performs evidence-grounded, section-by-section synthesis.
- 3.1 Overview of WebWeaver: The planner interleaves web search with outline refinement, producing a structured outline whose sections cite evidence stored in a curated memory bank.
- 3.2 Research Cycle: Iterative Evidence Acquisition and Outline Optimization: Unlike static one-way paradigms, the research cycle lets outline and search strategy co-evolve in response to new findings.
- 3.2 Research Cycle: Iterative Evidence Acquisition and Outline Optimization: During evidence acquisition, search results are summarized for subsequent planning and detailed evidence is extracted into the memory bank for writing.
- 3.2 Research Cycle: Iterative Evidence Acquisition and Outline Optimization: The planner continuously expands, cites, or restructures the outline using newly acquired information rather than treating outline generation as one-time planning.
- 3.3 Memory-Grounded Synthesis: Hierarchical Retrieval and Writing: The writer retrieves only citation-linked evidence for each subtask, reasons over it, writes the section, and prunes its source materials before the next cycle.
4 Experiments
WebWeaver is evaluated on three open-ended deep research benchmarks through standard metrics, with experiments examining overall performance, iterative outline optimization, hierarchical writing, and agentic finetuning. Across these analyses, the framework achieves strong benchmark results, iterative refinement improves report structure and support, focused writing manages context, and WebWeaver-3k improves smaller-model performance.
- Main Results: WebWeaver establishes state-of-the-art performance across DeepResearch Bench, DeepConsult, and DeepResearchGym.It achieves 93.37% citation accuracy on DeepResearch Bench, a 66.86% win rate on DeepConsult, and a 96.77 average score on DeepResearchGym.
- Analysis: The planner averages nearly 16 search steps, 21 unique queries, and more than two outline-optimization cycles, producing an outline of about 4k tokens.These statistics indicate that the evaluated research process involves extensive exploration and repeated outline refinement.
- Analysis: Iterative outline optimization improves comprehensiveness, insight, depth, breadth, and evidence support across the evaluated benchmarks.DeepResearch Bench comprehensiveness rises from 48.85 to 50.82 and insight from 46.33 to 48.35, while DeepResearchGym reaches depth 100, breadth 99.58, and support 98.33.
- Analysis: Hierarchical writing supplies smaller, focused contexts at each writing step than brute-force writing and produces a substantially longer, more detailed report.The comparison attributes the richer output to providing relevant information at the appropriate stage rather than overwhelming the model with all available information.
- Analysis: Finetuning on WebWeaver-3k raises citation accuracy from nearly 25% to 85.90% and improves DeepConsult and DeepResearchGym scores from 4.57 to 6.09 and 77.27 to 90.89.The reported gains are associated with learning the framework’s long-sequence reasoning, tool-use, iterative planning, evidence acquisition, and writing capabilities.
5 Related Works
Prior deep-research systems use static plans, brute-force evidence conditioning, or parallel section writing, creating adaptation and coherence limitations.
- Open-ended research agents include proprietary systems with strong performance but costly, restricted APIs that limit adoption and academic study.
- Static outline-and-write methods fix the plan before evidence evolves and feed all evidence into the model during generation.
- Parallel section writing can improve decomposition, but title-based retrieval and separate generation risk content and style incoherence.
6 Conclusion
WebWeaver combines dynamic planning with hierarchical retrieval and writing, outperforming proprietary and open-source systems while framing long-context reasoning as structured information management.
- WebWeaver consistently outperforms proprietary and open-source systems, establishing a new state-of-the-art.
- Its planner and writer dynamically explore, structure, retrieve, and compose information through deliberate actions rather than single-pass processing.
- The framework reframes long-context reasoning as a structured system-level information-management problem orchestrated through precise actions.
A Prompt Template for Outline Judgement
The prompt evaluates outline quality across instruction following, depth, breadth, support, and insightfulness. It requires structured, balanced, evidence-grounded, and practically useful outlines.
- The rubric assesses instruction following, including scope, audience, purpose, constraints, required sections, hierarchy, ordering, granularity, and formatting.
- Strong outlines provide analytical depth through specific subpoints, mechanisms, assumptions, uncertainties, methods, metrics, dependencies, and open questions.
- The rubric rewards broad but focused coverage, balanced perspectives, traceable source support, concrete examples, and original actionable synthesis.
B.1 ReAct Trajectories
The paper presents separate planning and writing trajectories implemented with explicit ReAct-style tool calls, responses, reasoning, and termination steps. Planning repeatedly improves the outline, while writing retrieves evidence and composes sections.
- The framework separates planning and writing into distinct trajectories, with planning focused on outline improvement and writing focused on report composition.
- Planning interleaves analysis, tool calls, tool responses, and outline writing before terminating the process.
- Writing follows a ReAct format that combines reasoning, tool use, retrieved responses, section writing, and eventual termination.
B.2 Case Study of Planning
The planning case study addresses a multifaceted Parkinson’s disease query by decomposing it into disease progression, emergency warning signs, DBS care, and daily-life support. The resulting outline is described as comprehensive and logically organized.
- The case study decomposes the query into disease stages, warning signs, emergency intervention, DBS outcomes and care, and daily-life support strategies.
- Planning searches sequentially for Parkinson’s stages, emergency signs, DBS postoperative care, and family or daily-living guidance.
- Retrieved evidence covers subtle unilateral symptoms, progressive bilateral impairment, DBS support resources, and psychosocial adjustments after treatment.
- The resulting outline is judged to have four hierarchical levels and to provide comprehensive, practical, logically organized coverage of the requested issues.
B.3 Case Study of Writing
The writing case study retrieves evidence section by section to explain Parkinson’s progression, staging, assessment, warning signs, and care implications. It highlights Stage 2-to-3 transition as a major safety threshold and identifies urgent warning signs for families.
- The generated article explains progression from subtle unilateral symptoms to severe bilateral impairment requiring constant care, with emergency signs including choking, falls, medication disruption, neurological changes, and psychiatric crises.
- The transition from Stage 2 to Stage 3 marks the emergence of balance problems and a substantially altered safety and care profile.
- MDS-UPDRS provides a more nuanced assessment of motor and non-motor effects than stage-based classification alone.