WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research

The field of deep research automation has evolved from simple keyword-based retrieval to sophisticated multi-turn, knowledge-augmented systems. Early methods like keyword search and rule-based extraction faced limitations in relevance and coherence. Recent advances include retrieval-augmented generation models such as REALM and RAG, which integrate external knowledge bases. However, these systems often struggle with long-context management, source traceability, and hallucination issues. The rise of large language models (LLMs) like GPT-3 and PaLM has enabled more fluent text generation, but static pipelines and full-input methods still suffer from information overload and hallucinations. Recent efforts focus on structured, source-grounded outputs, but lack dynamic, iterative mechanisms to adapt to new findings. WebWeaver addresses these gaps by introducing a feedback loop that continuously refines research blueprints and evidence retrieval, inspired by human research behavior.

Current deep research systems predominantly rely on static workflows, where the research plan is fixed before evidence collection, or on one-pass approaches that gather all information upfront. These methods are limited by outdated internal knowledge, inability to adapt to new discoveries, and excessive information overload, leading to incoherent reports and low citation accuracy. The core challenge is to develop a system that can dynamically adjust its research strategy, selectively retrieve relevant evidence, and maintain source traceability over long contexts. Achieving this requires balancing exploration and exploitation, managing long-term dependencies, and reducing hallucination risks, all within computationally feasible frameworks.

WebWeaver’s key innovations include: 1) a dynamic, iterative research cycle where the planner refines the outline based on newly acquired evidence; 2) a citation-grounded, hierarchical retrieval mechanism that ensures each report section is supported by relevant, verifiable sources; 3) integration of ReAct-based multi-turn reasoning to enable flexible, adaptive decision-making. These elements collectively enable the system to mimic human research behaviors—exploring, hypothesizing, and summarizing—while maintaining source fidelity. The approach also leverages small, fine-tuned models trained on a high-quality dataset (WebWeaver-3k), making it accessible and scalable.

• �� The planner begins with an initial outline and enters a multi-round loop: • Action 1: Search—queries web engines, filters URLs based on relevance, extracts summaries and evidence, and updates the memory bank. • Action 2: Outline optimization—refines the research blueprint, adding citations and restructuring sections based on new evidence. • Action 3: Terminate—stops when the outline is comprehensive. • The writer then takes the refined outline and evidence memory to generate the report: • For each section, it retrieves relevant evidence using citations, performs internal reasoning (“think”), then writes (“<write>”) the content. • After each section, evidence is pruned from context to prevent overload. • The process repeats until the entire report is completed, ensuring source-grounded, coherent output.

The evaluation employed three benchmarks: DeepResearch Bench (100 complex tasks across 22 fields), DeepConsult (business and consulting prompts), and DeepResearchGym (real-world queries). Metrics included RACE (comprehensiveness, insight, readability) and FACT (citation accuracy). Baselines included static pipelines and full-input models. WebWeaver outperformed all, with scores of 50.62 (RACE) and 78.25 (FACT). Ablation studies confirmed the importance of outline optimization and layered retrieval. Fine-tuning on WebWeaver-3k enabled small models (e.g., GPT-oss-120b) to reach near-expert levels, validating the approach’s scalability and robustness across domains.

WebWeaver achieved the highest scores on all benchmarks, notably 50.62 on RACE and 78.25 on FACT, with citation accuracy over 93%. It demonstrated superior ability to adapt research strategies, identify gaps, and produce source-grounded reports. The layered retrieval significantly reduced hallucinations and information loss, especially in multi-domain tasks. Fine-tuned small models matched large models’ performance, indicating high efficiency. These results underscore the effectiveness of dynamic, iterative research cycles over static or one-pass methods.

WebWeaver’s framework suits automated scientific literature review, policy analysis, and enterprise intelligence. It can assist researchers in rapidly synthesizing complex data, support decision-makers with trustworthy reports, and facilitate knowledge management. Its ability to adapt research paths and ensure source fidelity makes it valuable for any domain requiring rigorous, comprehensive analysis. Future integration with multimodal data and reasoning systems could further enhance its utility, enabling autonomous scientific discovery and real-time decision support.

Despite its strengths, WebWeaver faces challenges in environments with unreliable or biased sources, where evidence quality impacts report trustworthiness. Its multi-round process incurs high computational costs, limiting real-time applications. The reliance on high-quality evidence sources makes it vulnerable to misinformation. Future work should focus on improving robustness, efficiency, and multimodal integration, as well as addressing ethical concerns related to automated research.