1. 2025-09-17 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 Tongyi Lab , Alibaba Group https://tongyi-agent.github.io/blog https://github.com/Alibaba-NLP/DeepResearch Abstract This paper tackles open-ended deep research (OEDR), a complex challenge where AI agents must synthesize vast web-scale information into insightful re- ports. Current approaches are plagued by dual-fold limitations: static research pipelines that decouple planning from evidence acquisition and one-shot gen- eration paradigms that easily suffer from long-context failure issues like "loss in the middle" and hallucinations. To address these challenges, we introduce 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, source-grounded outline linking to a memory bank of evidence. The writer then executes a hierarchical retrieval and writing process, composing the re- port section by section. By performing targeted retrieval of only the necessary evidence from the memory bank for each part, it effectively mitigates long- context issues. Our framework establishes a new state-of-the-art across major OEDR benchmarks, including DeepResearch Bench, DeepConsult, and Deep- ResearchGym. These results validate our human-centric, iterative methodology, demonstrating that adaptive planning and focused synthesis are crucial for producing high-quality, reliable, and well-structured reports. 7.0 44 44.64 44.34 42 45.00 6.64 95.0 6.5 46.45 46 48 97.5 6.70 6.0 5.42 5.5 5.00 5.0 4.60 4.5 92.5 96.77 96.02 95.07 91.27 90.0 87.5 85.0 82.5 80.0 4.0 49.71 6.96 50.62 50.58 DeepResearchGym 100.0 84.46 80.25 52 50 DeepConsult 7.5 DeepResearch Bench Figure 1: Performance of varying deep research agents on DeepResearch Bench, DeepConsult, and DeepResearchGym. The results on DeepResearch Bench are taken from the official leaderboard. Our proposed WebWeaver achieves state-of-the-art performance. Corresponding author. 1

2. 1 Introduction Large Language Models (LLMs) (OpenAI, 2025b; Qwen Team, 2025; Liu et al., 2024; DeepMind, 2025; anthropic, 2025) have demonstrated remarkable capabilities across a wide array of well-defined tasks, from factual question answering (Wei et al., 2025; Mialon et al., 2023) to document summarization (Zhang et al., 2025) and code generation (Jiang et al., 2024). Their success, however, has largely been confined to scenarios with clear instructions and ground-truth answers. The true frontier for autonomous AI lies in transcending these structured problems to tackle the complex, open-ended challenges that define human-level knowledge work—a process driven by curiosity, synthesis, and the discovery of novel insights. We term this challenge open-ended deep research (OEDR). Unlike tasks with ground-truth answers, OEDR requires an agent to independently navigate and digest a vast corpus of information, often exceeding 100 web pages and PDFs, to form a detailed report that offers unique, synthesized viewpoints. This represents a monumental challenge, and as shown in Fig. 1, most current agents fail dramatically on the recent benchmarks (Du et al., 2025; Consult, 2025; Coelho et al., 2025) designed to test this capability, highlighting a critical gap we aim to address. Current attempts to tackle OEDR fall into two main categories: proprietary and open-source solutions. While several powerful proprietary agents exist (OpenAI, 2025a; Research, 2025b;d;a), their prohibitively expensive APIs and restrictive quotas create significant barriers, limiting widespread adoption and hindering academic research. Consequently, the focus has shifted towards open-source alternatives, which predominantly follow two paradigms. As shown in Fig. 2, the first is a straightforward "search- then-generate" approach (Tao et al., 2025; Roucher et al., 2025), where the agent gathers all information before directly generating a report. This method often results in low-quality, incoherent outputs because it lacks a guiding outline to structure the synthesis. The second, more sophisticated approach first generates a static outline and then performs targeted searches for each section (Han et al., 2025; Research, 2025e;c). However, this strategy is critically flawed: the outline is fixed upfront, relying solely on the LLM’s internal and often outdated knowledge. This rigidity "fossilizes" the research process, preventing the agent from exploring unexpected but valuable avenues discovered during its search. Furthermore, feeding all retrieved materials into a single context for final generation is susceptible to well-known issues like “loss in the middle” (Liu et al., 2023) and increased hallucinations, compromising the report’s accuracy and depth (Bai et al., 2024; Wu et al., 2025c). The key, we believe, lies in abandoning rigid, machine-like pipelines and instead embracing the organic process of human intellect. Our approach is designed to do just that: it teaches the agent to research like a person. A human expert doesn’t finalize their entire plan before starting; they allow their outline to be a living document. We implement this principle through an agentic loop where actions of searching and outline optimization are provided. As the agent explores the web-scale information landscape, its discoveries continuously inform and reshape the outline. This allows for genuine exploration and adaptation, ensuring that the research is not confined to its initial, limited understanding. Then, when it is time to write, our agent avoids the brute-force method of "reading" everything at once. Just as a human writer would refer to specific notes for a specific chapter, our agent composes each section by focusing only on the most pertinent source materials. By doing so, it operates with clarity and precision, crafting a final report that is not just a summary of data but a well-structured and deeply considered piece of analysis. To this end, we propose WebWeaver by following the human-centric philosophy, a dual-agent framework comprising a planner and a writer. As shown in Fig. 2, the planner embodies the exploratory research phase, operating in a dynamic cycle that iteratively interleaves evidence acquisition with outline op- timization, culminating in a comprehensive, source-grounded research outline, where each section is explicitly linked via citations to a curated memory bank of source evidence. When it turns to the writing phase, to address the critical long context and attentional management challenge, the writer executes a memory-grounded, hierarchical synthesis process. It constructs the report section by section, performing 2

3. Query Planner Query Report generation Static Outline Section 1 Section 2 Search Planner (a) Search then generate Search (b) Outline-guided search then generate Outline Optimization Query Report generation Planner Hierarchical Writing Outline Report Section 1 Section 1 Writer Section 2 Search Section 2 Retrieve (c) WebWeaver (ours) Report Memory bank Figure 2: Paradigm comparison: (a) the search-then-generate paradigm first gathers information and then directly generates a report; (b) the paradigm initializes a static outline and then performs targeted searches for the outline; (c) WebWeaver not only enables a dynamic research cycle where the outline and search strategy co-evolve but allows hierarchical and attentional writing by retrieving only relevant evidence. targeted retrieval of only the relevant evidence from a structured memory bank for each subtask. This synergistic division of labor enables our agent to navigate complex information landscapes and produce reports that are both comprehensive in scope and meticulous in their evidentiary grounding. Extensive experiments demonstrate that WebWeaver achieves state-of-the-art (SOTA) performance and outperforms both the proprietary and open-source agent systems on three recent and challenging open- ended deep research benchmarks. Detailed discussion is produced to demonstrate the effectiveness of outline optimization and memory-grounded synthesis. Critically, to enhance the performance of the smaller models for practical use, we construct a high-quality SFT dataset, WebWeaver-3k, generated by our framework. The supervised finetuning experiments with WebWeaver-3k demonstrate that the complex skills of thinking, searching, and writing can be distilled and taught, enabling smaller, accessible models to achieve the expert-level performance previously confined to large-scale proprietary systems. 2 Preliminaries Problem definition. We consider the open-ended research question without the ground-truth answers. Given an open-ended question, the agents need to search relevant information and finally output a report or article. To achieve this, we implement a planning agent for collecting information, a memory to store materials, and a writing agent for report generation. For both the planning and writing agents, we adopt ReAct (Yao et al., 2023) as the agent’s framework. Upon receiving a question, the agents perform several iterations of thought-action-observation. Specifically, in each iteration, based on the existing context, the LLM generates a thought and executes a parsable action, then awaits the environment to return an observation. The planning and writing stages terminate with the output token of “<terminate>”. A complete trajectory with T iterations can be defined as H T = ( τ 0 , a 0 , o 0 , . . . , τ i , a i , o i , . . . , τ T , a T ) , (1) where τ i , a i , o i represent thought, action, and observation sampled from the planning or writing policy based on all previous context in the i-th round, respectively. 3

4. Actions. For the planner, the action space consists of search, write outline, and terminate. Given the search queries, the search engine returns titles, snippets, and corresponding URLs. To save context space, we further execute the actions of the URL selection, parsing the page via URL, summarizing relevant contents, and extracting evidence with LLMs following the searching queries. The search tool finally returns the selected URLs and their corresponding summaries. The action of “write outline” is to generate and optimize the outline, and the “terminate” action is to terminate the planning process. For the writer, the action space consists of retrieve, write, and terminate. Besides the terminate action, the retrieve action is to retrieve evidence from the memory bank by providing the grounded citations in the outline. The write action is provided to write the section of the report. Memory bank. Answering an open-ended question requires long-context input of the collected informa- tion and long-context output of the final report. To search sufficient materials, the planner often searches and parses more than 100 web pages, with more than 100k tokens. The writer often outputs more than 20k tokens to produce a comprehensive report. Prior open-sourced deep research agents (Roucher et al., 2025; Research, 2025e;c) include all the raw materials (e.g., web pages and PDF files) in the LLM context, leading to quality degradation due to attentional failures like the “lost in the middle” problem, poor coherence, and increased hallucinations (Liu et al., 2023; Li et al., 2024; Bai et al., 2024; Wu et al., 2025c). To this end, we introduce a memory to achieve context management for both planner and writer. Only a short summary of the web page or PDF file is included in the search context, and only necessary raw pages will be retrieved from the memory to write the corresponding sections. 3 Method Our methodology is embodied in a dual-agent framework, comprising a planner and a writer. The planner is responsible for the dynamic cycle of evidence acquisition and outline optimization, while the writer performs focused, section-by-section synthesis to construct the final report. This division of labor directly mirrors the cognitive workflow of a human researcher. 3.1 Overview of WebWeaver The entire workflow is visualized in Fig. 3. Tasked with evidence acquisition and outline optimization, the planner operates in a dynamic research cycle. It iteratively interleaves evidence acquisition from web searches with the continuous refinement and optimization of a report outline. The output of this exploratory phase is not just a collection of sources but a comprehensive, well-structured outline where each section is explicitly linked via citations to a curated memory bank of source evidence. Subsequently, the writer takes over for the synthesis phase. To circumvent the pitfalls of one-shot generation and long-context issues, the writer adopts a section-wise and memory-grounded synthesis approach. For each section of the outline, it performs targeted retrieval of only the pertinent evidence from the memory bank and composes the content. This division of labor ensures that the final report is not only coherent and well-organized but also deeply source-grounded, faithfully mirroring the rigor of human-led deep research. 3.2 Research Cycle: Iterative Evidence Acquisition and Outline Optimization Recent deep research agents often follow an "outline-guided search" paradigm (Han et al., 2025; Research, 2025e;c). By generating a static outline before any evidence is gathered, they create a rigid research path that is blind to emergent insights. This fundamental decoupling of planning from discovery limits the depth and breadth of the research. We address this by proposing a dynamic research cycle where the outline and search strategy co-evolve, allowing the agent to adapt and explore new findings. The core of our planner’s operation is a dynamic research cycle that iteratively interleaves evidence 4

5. Query Please conduct a study and prepare a report on the 'Construction and Application of a Sports Intelligent Tutoring and Learning Guidance System Driven by Multimodal Data Fusion’. Think for searching: 1. I need to conduct a compre- hensive study on "Construction and Application …” 2. … {"query": [ "sports intelligent tutoring system multimodal data fusion", … ]} Think Actions Memory Bank Write Outline <ID_1> Summary: The article offers a … Evidence: … Observations Summary <ID_2> Summary: The article offers a … Evidence: … Think Actions Retrieve Write Terminate Observations Summary: The article offers a … Think for writing: From the materials, we obtain ... Search Terminate Outline Outline Section Round 1: Outline Optimize Round 2: Outline 1. Construction and Application… 1.1 I. Introduction… 1.1.1 A. Definition <citation>id_1</citation> 1.1.2 B. Role of Multimodal Data Fusion Evidence Round 1: Write Section 1 1. Construction and Application… 1.1 I. Introduction… 1.1.1 A. Definition <citation>id_1</citation> 1.1.2 B. Evolution from Traditional Sports …<citation>id_2</citation> 1.1.3 C. Role of Multimodal Data Fusion <citation>id_3</citation> 1. Construction and Application… 1.1 Introduction… 1.1.1 Definition: Sports Intelligent Tutoring Systems (Sports ITS) <citation>id_1</citation> … Think for retrieving 1. I need to start writing a comprehensive report …” 2. I should begin with retrieving… { } "retrieve_id ": [ "id_1", "id_2", "id_3” ] Evidence: <id_1> A sports news web page… Round 2: Write Section 2 2. Theoretical Foundations 2.1 Intelligent Tutoring Systems (ITS) Framework 2.1.1 Core Components and Architecture The theoretical foundation of Sports Intelligent… Report Figure 3: The workflow of WeaWeaver. Left: The planner first iteratively collects evidence via the search tool and optimizes the outline until outputting a comprehensive and citation-grounded outline. Right: The writer performs hierarchical and attentional writing by retrieving relevant evidence with the grounded citation in the outline. acquisition with outline optimization. Unlike static approaches, our planner continuously adapts its strategy based on emergent findings. For each step, the planner selects one of the three actions: search, outline optimization, and terminate. Evidence acquisition. When there is still insufficient evidence or knowledge to make a comprehensive outline to answer the open-ended question, the planner will continue collecting evidence by executing the search action. Given any search queries, the planner begins by querying a web search engine, which returns the results that contain the raw URLs with corresponding snippets and titles. To combat the contextual noise and processing overhead from raw URLs, it employs a two-stage filtering process. First, we prompt LLMs to select only the relevant URLs based on titles and snippets. Then, for each parsed page of the selected URLs, we perform two critical actions: leveraging LLMs to (1) distill a query-relevant summary, which is fed back into the planner’s context to inform subsequent search iterations, and (2) extract verifiable, detailed evidence (e.g., quotes, data points), which is stored in a structured memory bank for the subsequent writing. Outline optimization. After acquiring some evidence, the planner revisits the report’s outline. This is not a one-time generation step but a process of continuous refinement and optimization. The planner uses the newly acquired information to expand sections, add new subsections, or even restructure the entire outline to better reflect a comprehensive understanding of the topic. Crucially, it populates the outline with citations, mapping each section to the specific evidence IDs in the memory bank. This citation mechanism is vital for ensuring source-groundedness and enabling the hierarchical writing process in the next stage. This iterative loop continues until the planner outputs a terminate action with a tag “<terminate>” when the outline is sufficiently comprehensive and well-supported by evidence. 3.3 Memory-Grounded Synthesis: Hierarchical Retrieval and Writing A pivotal challenge in generating long-form reports is not just information access but attentional man- agement. The prevailing approach of feeding all gathered evidence into a single context window for 5

6. one-shot generation is fundamentally flawed. This brute-force method saturates the model’s attentional capacity, leading to some long-context issues like “loss in the middle (Liu et al., 2023)”, where crucial details are overlooked, and “contextual bleeding (Liu et al., 2025)”, where information from one section incorrectly influences the synthesis of another. We argue that a successful synthesis process must mirror human cognition by breaking down the complex task of long-context, one-step writing into manageable subtasks of attentional writing with only relevant evidence. Therefore, we adopt a hierarchical, divide- and-conquer strategy, where the report is constructed sequentially, with the model’s focus constrained to only the most relevant evidence at each step. Upon completion of the planning phase, the writer is provided with the structured, source-grounded outline and access to the evidence memory bank. The composition of each section is not a single, monolithic action but a deliberate, intra-sectional reasoning cycle designed to ensure both accuracy and coherence. This cycle unfolds as follows: First, the writer identifies its immediate subtask, such as “Let’s write the first section.” It then executes a targeted retrieval action, pulling only the relevant evidence from the memory bank as indicated by the outline’s citations. Upon receiving the evidence, the writer enters a crucial internal reasoning phase with a think action. In this thinking step, it analyzes the retrieved content, synthesizes key insights, selects the most compelling pieces of evidence, and formulates a coherent narrative structure for the section. This internal monologue is critical for moving beyond simple summarization to genuine synthesis. Only after this internal synthesis plan is formed does the writer proceed to the writing action, composing the prose and encapsulating it within “<write>” tags. Once a section is complete, its corresponding source materials are explicitly pruned from the context window and replaced with a placeholder message. This dynamic retrieval-and-pruning mechanism is the cornerstone of our approach: it ensures the writer’s context remains highly relevant for the next cycle, mitigates context overflow, and prevents cross-sectional interference. This entire process repeats hierarchically for all sections until the writer outputs a final “<terminate>” token, signaling the completion of the full report. 4 Experiments In this section, we first evaluate WebWeaver on three recent and challenging benchmarks. Detailed discussion is then produced to demonstrate the effectiveness of outline optimization and memory- grounded synthesis. Furthermore, we curate a high-quality SFT dataset to improve the capabilities of thinking, searching, and writing for a smaller model to achieve expert-level performance. 4.1 Setup Benchmarks. To evaluate the performance of Deep Research systems, we use three open-ended bench- mark datasets: • DeepResearch Bench (Du et al., 2025) comprises 100 PhD-level complex research tasks metic- ulously formulated by domain experts across 22 distinct fields, such as Science & Technology, Finance & Business, Software Engineering, and Art & Design. • DeepConsult (Consult, 2025) is a specialized collection of prompts tailored for in-depth research within the business and consulting domains. The query set encompasses a wide range of topics, including marketing strategy, financial analysis, emerging technology trends, and business planning. • DeepResearchGym (Coelho et al., 2025) is used to assess performance on real-world, complex queries. This dataset contains 100 queries sampled from the extensive Researchy Questions dataset (Rosset et al., 2024), which includes approximately 96,000 authentic information-seeking queries. 6

7. RACE Agent systems FACT Overall Comp. Insight Inst. Read. Eff. c. C. acc. WebShaper (32B) 34.93 31.58 26.17 44.81 40.38 - - langchain-open-deep-research doubao-research kimi-research Claude-research openai-deepresearch Gemini-2.5-pro-deepresearch 43.44 44.34 44.64 45.00 46.45 49.71 42.97 44.84 44.96 45.34 46.46 49.51 39.17 40.56 41.97 42.79 43.73 49.45 48.09 47.95 47.14 47.58 49.39 50.12 45.22 44.69 45.59 44.66 47.22 50.00 52.62 - - 39.79 165.34 52.86 - - 75.01 78.3 WebWeaver (qwen3-30b-a3b-instruct-2507) WebWeaver (gpt-oss-120b) WebWeaver (qwen3-235b-a22b-instruct-2507) WebWeaver (Claude-sonnet-4-20250514) 46.77 48.11 50.62 50.58 45.15 48.03 51.29 51.45 45.78 47.20 51.00 50.02 49.21 48.94 49.98 50.81 47.34 48.11 48.89 49.79 26.74 64.88 166.73 200.75 25.00 66.14 78.25 93.37 Table 1: Performance of agents on DeepResearch Bench in terms of comprehensiveness (Comp.), insight, instruction-following (Inst.), readability (Read.), effective citations (Eff. c.), and citation accuracy (C. acc.). The best results are highlighted with green color, and the second-best results are highlighted with underlines. Metric. We use the official evaluation metrics with the recommended judge LLMs for each benchmark. • DeepResearch Bench. This benchmark utilizes two suites of metrics to evaluate different aspects of the system’s output: 1) RACE (Report Quality): It assesses the quality of the final generated report with a reference report across four dimensions, namely Comprehensiveness (Comp.), Insight/Depth (Insight), Instruction-Following (Inst.), and Readability (Read.). An overall score is then calculated as a weighted summation of these components. 2) FACT (Web Retrieval): It measures the effectiveness and reliability of the information retrieval process. This includes Citation Accuracy (C. Acc.) and the Average Effective Citations per Task (Eff. c.). We adopt Gemini-2.5-pro as the judgement model by following the benchmark. • DeepConsult. Performance on this benchmark is determined through a pairwise comparison against the openai-deepsearch baseline. The primary metrics are the win rate, tie rate, and loss rate, which are supplemented by a reported average quality score. The judgement model is gpt-4.1-20250414. • DeepResearchGym. An LLM acts as a judge to assess the generated report on several quality dimensions, including clarity, insightfulness, depth, balance, breadth, support, and an average quality score. The judgement model is gpt-4.1-mini-20250414. Compared systems. We benchmark the performance of WebWeaver against a range of state-of-the-art DeepResearch systems. These systems are categorized into two groups: • Open-Source Systems: For open-source counterparts, we compare against WebShaper-32B (Tao et al., 2025) and langchain-open-deep-research (LangChain, Inc., 2023). • Proprietary Systems: We include several leading commercial systems: doubao-research (Research, 2025a), kimi-research (Research, 2025d), Claude-research (anthropic, 2025), openai-deepresearch (OpenAI, 2025a), and Gemini-2.5-pro-deepresearch (Research, 2025b). Implementation details. The WebWeaver is compatible with various advanced LLMs. In the experi- ments, we utilize the following models: Qwen3-30b-a3b-instruct-2507 (Yang et al., 2025), GPT-oss-120b (Agarwal et al., 2025), Qwen3-235b-a22b-instruct-2507 (Yang et al., 2025), and Claude-sonnet-4-20250514 7

8. DeepConsult Agent systems WebShaper (32B) win tie lose DeepResearchGym Avg. score Cla. Depth Bal. Brea. Sup. Ins. Avg. score 3.25 3.75 93.00 1.63 64.70 63.00 59.30 66.50 9.40 59.90 53.80 doubao-research Claude-research openai-deepresearch Gemini-2.5-pro-deepresearch 29.95 25.00 0.00 61.27 40.35 38.89 100.00 31.13 29.70 36.11 0.00 7.60 5.42 4.60 5.00 6.70 68.85 86.67 84.90 90.71 93.12 96.88 98.10 99.90 83.96 84.41 89.80 93.37 93.33 96.56 97.40 99.69 84.38 26.77 88.40 95.00 83.12 90.22 89.00 97.45 84.46 80.25 91.27 96.02 WebWeaver (qwen3-30b-a3b-instruct-2507) WebWeaver (qwen3-235b-a22b-instruct-2507) WebWeaver (gpt-oss-120b) WebWeaver (Claude-sonnet-4-20250514) 28.65 54.74 65.31 66.86 34.90 28.61 11.22 10.47 36.46 16.67 23.47 22.67 4.57 6.47 6.64 6.96 71.88 89.16 89.78 90.50 85.51 97.58 100.00 99.87 75.80 87.68 91.91 94.30 84.78 96.21 99.66 100.00 63.77 95.26 94.94 98.73 81.88 92.85 95.06 97.22 77.27 93.14 95.07 96.77 Table 2: Performance of agents on DeepConsult in terms of win rate and average scores and on DeepRe- searchGym in terms of clarity (Cla.), depth, balance (Bal.), breadth (Brea.), support (Sup.), and insightful- ness (Ins.). The best results are highlighted with green color, and the second-best results are highlighted with underlines. (anthropic, 2025). We adopt Claude-sonnet-4-20250514 as the default agent model for ablation studies and discussion without any statements. We use GPT-oss-120b to select relevant URLs, perform query-relevant summaries, and extract evidence for the search action. We present the case studies in Appendix B. 4.2 Main Results Results on DeepResearch Bench. As presented in Table 1, our WebWeaver framework establishes a new state-of-the-art, consistently outperforming existing agents. This superior performance is a direct result of our dual-agent, iterative methodology. The high scores in comprehensiveness (Comp.) and insight stem from the planner’s dynamic research cycle, which iteratively expands the report’s scope based on emergent findings, unlike the rigid outline-first approaches. This process naturally leads to a higher number of effective citations (Eff. c.), as the planner is intrinsically motivated to seek more evidence to ensure that each section is well-supported. Furthermore, the remarkable citation accuracy (C. acc.) of 93.37% is achieved by the strong synergy between our agents: the planner embeds specific citation IDs into the outline, and the writer’s hierarchical synthesis process uses this structure for targeted retrieval. By focusing only on relevant evidence for each section, it drastically reduces context-bleeding and hallucinations, which also contributes to the enhanced readability (Read.) and Instruction-following (Inst.) scores. This demonstrates that by emulating human research patterns, our framework produces not just more thorough but also significantly more reliable and well-structured reports. Results on DeepConsult and DeepResearchGym. To validate the generalizability of our framework, we further evaluated WebWeaver on the DeepConsult and DeepResearchGym benchmarks, with results presented in Table 2. Our method demonstrates clear superiority on both, achieving the highest win rate (66.86%) on DeepConsult and the top average score (96.77) on DeepResearchGym. This success is rooted in our core design. The near-perfect scores in Depth (100.00) and Breadth (100.00) are a direct result of the planner’s iterative research cycle, which relentlessly expands the report’s scope beyond the limits of static planning. Concurrently, the writer’s hierarchical synthesis process ensures these comprehensive findings are well-organized, leading to outstanding scores in balance (94.30) and support (98.73). In essence, the quantitative dominance in structural metrics like depth and breadth on DeepResearchGym provides a clear explanation for the qualitative victories on DeepConsult, proving that our human-inspired, iterative process is a fundamentally more robust strategy for complex information synthesis tasks. 4.3 Analysis Statistics of planning and writing. The statistics in Table 3 provide a compelling quantitative narrative that not only justifies but also demonstrates the benefits of WebWeaver’s design. The planning task 8

9. Planning statistics DeepResearch Bench DeepResearchGym # Search step 15.71 16.65 # Outline token 4876.21 3732.87 # Outline optimization 2.16 2.20 # Saved page 112.25 102.55 Writing statistics # Search query 20.24 21.93 # Evidence token 67237 66301 # Summary token 14980 12543 # Output token 26127 26004 # Writing step 24.78 24.71 Table 3: The planning and writing statistics of Claude-sonnet-4-20250514 on DeepResearch Bench and DeepResearchGym. DeepResearch Bench 4 rounds 1 round 15.0% 5.0% DeepResearchGym 4 rounds 1 round 3 rounds 13.1% 4.0% 21.0% 3 rounds 25.3% 57.6% 59.0% 2 rounds 2 rounds Figure 4: Statistics of outline optimization of Claude-sonnet-4-20250514 on DeepResearch Bench and DeepResearchGym. involves an extensive exploratory phase with nearly 16 search steps and 21 unique search queries, proving that a simple, linear search is insufficient. The critical finding is that the outline undergoes more than two optimization cycles on average, expanding into a complex 4k-token outline. This empirically invalidates static-outline approaches and shows the tangible benefit of our iterative process: it produces a richer, more comprehensive plan that adapts to discovery. This deep planning phase amasses a staggering amount of information—over 100 saved pages, culminating in 67k evidence tokens and 15k summary tokens. This sheer volume makes a single-context approach computationally hard, thus mandating our memory-centric architecture with targeted retrieval as a foundational requirement, not just an optimization. Finally, the writer’s process of composing a 26k-token report in 25 discrete writing steps validates that our hierarchical synthesis is a practical way to maintain coherence over long outputs. In essence, the statistics of searching and writing affirm that each component of WebWeaver is a necessary and beneficial response to the inherent challenges of OEDR. DeepResearch Bench DeepResearchGym 52 51 50.82 50.50 50 48 49.55 49.84 100 98 49.24 48.85 48.73 48.72 48.35 48.15 49.26 49 49.65 Rounds Round 1 Round 2 Round 3 48.58 47.91 96 Rounds Round 1 Round 2 Round 3 100.00 99.55 99.33 98.33 97.33 96.32 96.00 95.68 95.91 93.18 92.67 46.33 90 46 45 overall scores Comprehensiveness Insight 88 Instruction following Readability Evaluation Metrics Figure 5: End-to-end scores with varying rounds of outline optimization on Deepresearch Bench. 95.33 95.42 95.00 94.17 94 92 47 99.58 99.09 99.33 92.00 91.36 90.42 overall scores Clarity Depth Balance Breadth Evaluation Metrics Support Insightfulness Figure 6: End-to-end scores with varying rounds of outline optimization on DeepresearchGym. End-to-end benchmark comparison for varying rounds of outlines. To isolate and quantify the benefits of outline optimization, as reported in Fig. 5, 6, we conducted an ablation study by evaluating the end-to-end benchmark performance. We collect the samples with three-round outline optimization from 9

10. DeepResearch Bench and DeepResearchGym, adopting the same writing strategy for them. The benefits of this iterative refinement are evident across both benchmarks. On DeepResearch Bench, the overall score steadily climbs, driven primarily by significant gains in comprehensiveness (48.85 → 50.82) and insight (46.33 → 48.35). This directly validates our hypothesis that each optimization round allows the planner to build a more detailed and logically structured outline. This enhanced structure is further reflected in DeepResearchGym’s metrics, where later rounds achieve near-perfect scores in depth (100) and breadth (99.58), indicating a more exhaustive topic coverage. Crucially, this is not just about adding more content; the steady rise in support (95.91 → 98.33) demonstrates that a more refined outline creates a better-scaffolded structure, enabling the writer to more tightly link claims to evidence. In summary, this analysis empirically demonstrates that iterative outline optimization is not a redundant step but a critical mechanism for elevating a report from a simple summary to a deep, insightful, and well-supported piece of research. DeepResearch Bench 98.10 95 94.29 92.30 90.63 90 89.05 89.52 88.57 87.62 86.67 81.90 83.81 81.43 89.6 88.2 92.0 93.6 92.8 Rounds Round 1 Round 2 Round 3 95.2 94.4 87.2 87.6 85.2 89.6 81.2 78.8 77.2 71.2 70 81.6 73.6 66.4 76.19 75 60 70 rall ove 86.2 80 80 65 98.4 90 86.19 84.29 DeepResearchGym 100 Rounds Round 1 Round 2 Round 3 95.71 90.14 85 95.71 94.29 Scores 96.67 69.52 res sco In c stru ing llow fo tion th Dep Bal e anc h adt Bre t por Sup 51.2 50 s nes tful gh Insi s ore ll sc ra ove Evaluation Metrics Figure 7: LLM-judged scores for varying rounds of outline optimization on Deepresearch Bench. Ins ing llow fo tion truc th Dep e anc Bal h adt Bre t por Sup ess fuln ght Insi Evaluation Metrics Figure 8: LLM-judged scores for varying rounds of outline optimization on DeepresearchGym. LLM judgement for varying rounds of outlines. To directly evaluate whether our optimization truly improves outline quality, we utilized an LLM-as-a-judge (Zheng et al., 2023) to assess the outlines from each of the three optimization rounds using gpt-4.1-mini-2025-04-14 in terms of instruction following, depth, balance, breadth, support, and insightfulness. The judgment prompt is provided in Appendix A. The results in Fig. 7, 8 provide a resounding confirmation of our iterative approach. On both benchmarks, the overall score for the outline quality shows a significant, monotonic increase, jumping from 81.9 to 92.3 on DeepResearch Bench and from 77.2 to 88.2 on DeepResearchGym. This improvement is driven by clear gains in structural quality; the near-perfect scores in Depth (up to 95.71) and Breadth (up to 98.4) provide direct evidence that each optimization cycle successfully expands the research’s scope. Crucially, this is not mere expansion. The substantial increase in the Support score (e.g., from 51.2 to 73.6 on DeepResearchGym) is particularly revealing, indicating that later-round outlines are more effectively grounded with a stronger mapping between planned sections and available evidence. This enhanced grounding and structure culminate in a plan that is itself more insightful (improving by 10-15 scores on both benchmarks). Therefore, this direct assessment confirms that our iterative planner is not just adding content but is actively forging a superior, more coherent, and better-supported blueprint—the foundational prerequisite for a high-quality final report. Hierarchical retrieval and writing vs. brute-force writing. To empirically validate our hierarchical writing process, we conducted a critical ablation study comparing our hierarchical writer against a brute-force baseline that attempts to include the entire memory bank to generate the final report, which is similar to the workflow of LongWriter (Bai et al., 2025). The results are unequivocal: our hierarchical approach dramatically outperforms the brute-force method across every metric, confirming that a “divide and conquer” strategy is essential. The most striking improvements are in insight (40.97 → 50.02) and readability (42.29 → 49.79), which directly validates our hypothesis on attentional management; by 10

11. DeepResearch Bench 52 50.81 50.02 Writing Type 105 Brute-force Writing (LongWriter) Hierarchical Writing 51.45 50.58 50 Brute-force Writing (LongWriter) Hierarchical Writing 99.87 100 49.79 48.80 47.89 48 46 45.24 44 98.59 98.18 96.77 DeepResearchGym Writing Type 54 95 100.00 98.73 97.22 94.30 91.52 91.82 90.50 90 90.10 90.10 42.29 42 85 40.97 40 38 res r ove co all s ess Insi he pre Com en nsiv ght Inst llow n fo ctio ing ru 80.30 80 ity bil ada Re s ore ll sc ra ove Evaluation Metrics Figure 9: Performance comparison between hierar- chical writing and brute-force writing (LongWriter) on DeepResearch Bench. rity th e anc Dep Cla Bal h adt Bre Evaluation Metrics t por Sup ess fuln ght Insi Figure 10: Performance comparison between hi- erarchical writing and brute-force writing (Long- Writer) on DeepResearchGym. focusing the model on a curated context for each section, it can perform deeper reasoning rather than shallow summarization. This is further substantiated by the leap in support on DeepResearchGym (91.82 → 98.73), proving that our targeted retrieval-and-pruning mechanism effectively prevents “contextual bleeding” and ensures claims are correctly grounded. In conclusion, these results provide definitive evidence that emulating the human cognitive process of focused, section-by-section writing is not merely a beneficial choice but a fundamental requirement for generating coherent, insightful, and reliable long-form reports. Agentic Finetuning 100 2.4% 15.0% 28.2% 3 rounds 4 rounds 1 round Model Version Qwen3-30b-a3b-Instruct Qwen3-30b-a3b-Instruct (SFT) 85.90 WebWeaver-3k 90.89 77.27 80 60 46.77 48.11 40 25.00 20 54.4% 2 rounds 4.57 0 DeepResearch Bench (RACE) Citation Accuracy Benchmark Figure 11: Round statistics of outline opti- mization on WebWeaver-SFT. 6.09 DeepConsult DeepResearchGym Figure 12: Performance improvement of agentic finetun- ing on DeepResearch Bench. Agentic finetuning. While 30B-scale LLMs (e.g., Qwen3-30b-a3b-instruct-2507) possess strong founda- tional capabilities, they often exhibit deficiencies in stability and instruction-following when executing complex, multi-turn tool-calling sequences over long contexts. To bridge this critical gap, we constructed a high-quality Supervised Fine-Tuning (SFT) dataset: WebWeaver-3k. The process began by sourcing a diverse set of queries crawled from the web, which were then processed by a powerful, tier teacher model, instantiated within our WebWeaver agent framework. A stringent filtering protocol was applied to the resulting end-to-end research trajectories, retaining only those where the agent successfully executed the entire workflow and strictly adhered to the predefined action format. This quality control yielded a curated dataset of 3.3k high-fidelity planning trajectories and 3.1k writing trajectories. As detailed in Table 4 and Fig. 11, these trajectories encapsulate the profound complexity of the OEDR task, with an average case involving approximately 15 search steps, over two outline optimizations, and the processing 11

12. Planning statistics WebWeaver-SFT # Search step 14.67 # Outline token 4148.57 # Outline optimization 2.18 # Saved page 106.65 # Search query 18.8 Writing statistics # Evidence token 62637 # Summary token 14155 # Output token 22637 # Writing step 22.76 Table 4: The planning and writing statistics of training data on WebWeaver-SFT. of over 62,000 evidence tokens. By fine-tuning our base model on this data, we explicitly imbued it with the requisite long-sequence reasoning and tool-use capabilities to master our framework. The efficacy of our SFT strategy is quantitatively demonstrated by the significant performance gains across all benchmarks on Fig. 12, which directly reflect the model’s acquisition of our framework’s core competencies. The most dramatic validation is the leap in citation accuracy from a nearly unusable 25% to a reliable 85.90%. This provides direct, empirical evidence that the model has mastered the intricate mechanics of our Writer agent, learning to execute precise tool calls for evidence retrieval and faithfully write according to the source-grounded outline. Furthermore, the substantial increase in overall report quality, evidenced by the score on DeepConsult (4.57 → 6.09) and the massive jump on DeepResearchGym (77.27 → 90.89), reflects the successful acquisition of the planner’s more abstract abilities. These holistic improvements indicate that the model has learned the core loop of thinking (iteratively optimizing the outline) and searching (adaptively acquiring evidence), which is a prerequisite for generating a comprehensive and insightful final report. Ultimately, these results offer a powerful dual validation: they prove that our WebWeaver framework is a potent data generation engine, capable of deconstructing the formidable OEDR task into learnable demonstrations of thinking, searching, and writing, thereby enabling a smaller model to achieve expert-level performance. 5 Related Work Deep Research. Deep Research Agents have garnered significant attention for their powerful capa- bilities in information seeking, integration, and reasoning. Proprietary systems, such as DeepResearch (OpenAI, 2025a), Gemini Deep Research (google, 2025), and Claude Research (anthropic, 2025), have demonstrated performance comparable to human experts in domains like fact-checking and report writing. However, their opaque internal architectures and workflows hinder broader research and development. In the open-source community, many studies (Li et al., 2025b; Tao et al., 2025; Su et al., 2025; Qiao et al., 2025; Fang et al., 2025; Li et al., 2025a; Wu et al., 2025b;a) have been developed to tackle complex research benchmarks such as BrowseComp and GAIA by exploring methods like synthetic uncertain Question-Answering (QA) and formalized QA synthesis. Nevertheless, these solutions are primarily tailored for short-answer research queries and lack the capability to generate comprehensive, long-form reports on open-domain topics. Other open-source systems like OpenDeepResearch (Research, 2025e), GPT Researcher (Research, 2025c), and TTD-DR (Han et al., 2025) address long-form generation by first drafting a static framework, then retrieving content, and finally composing the report. This approach, characterized by a fixed structure and one-step generation, often leads to textual incoherence and hallucinations. In contrast, our method emphasizes the outline optimization and hierarchical writing processes to ensure the report’s fluency and factual accuracy. Long Writing. Ensuring the coherence and accuracy of LLM-generated long-form text is a persistent challenge. Previous work has explored methods like recursive prompting for story extension (Yang et al., 2022) and structured task decomposition to improve consistency (Yang et al., 2023; Wang et al., 2025). More recently, agent-based frameworks have become a mainstream solution. Systems like LongWriter (Bai et al., 2025) and CogWriter (Wan et al., 2025) employ a "plan-then-write" strategy, where a Planner Agent first creates an outline, and a Generation Agent then conditions on this plan to produce the full text. However, these methods rely on a static initial plan and a brute-force writing strategy by feeding 12

13. all the evidence into LLMs. In contrast, our approach uniquely enables the outline to be dynamically optimized in tandem with the evidence acquisition process, allowing for a comprehensive, source- grounded research outline. Furthermore, our proposed hierarchical writing process with only relevant evidence also mitigates the long-context issues from the brute-force writing strategy. 6 Conclusion In this paper, we introduced WebWeaver, a novel dual-agent framework designed to overcome the fundamental flaws of static, machine-like pipelines in open-ended deep research (OEDR). By emulating the human cognitive process that integrates the planner’s dynamic research cycle with the writer’s hierarchical retrieval and writing process, WebWeaver consistently outperforms both proprietary and open-source systems, establishing a new state-of-the-art. Beyond its superior performance, the true significance of WebWeaver lies in the new paradigm it offers the community for tackling complex, information-intensive tasks. It reframes the intractable challenge of long-context reasoning, demonstrating that it can be successfully deconstructed into a structured problem of system-level information management, orchestrated through a series of precise actions. Both the planner and writer are embodiments of this principle: they use tools to dynamically explore, structure, and write, rather than passively processing it in a single pass. This work does not just present a better agent system; it presents a new blueprint for building the agent system that masters intensive knowledge through deliberate actions, not just brute-force attention. 13

14. A Prompt Template for Outline Judgement The detailed prompt template and judgement criteria in terms of instruction following, depth, balance, breadth, support, and insightfulness are shown as follows: Judgement Criteria { "name": "Instruction following", "description": "Evaluate how well the outline follows the user’s instructions for an outline. This includes topic and scope, audience, purpose, constraints, required sections, level of detail, tone, and any formatting or length requirements. Check outline-specific expectations: clear hierar- chical structure (e.g., H1/H2/H3 or bullet levels), logical ordering, consistent granularity across sections, numbering if requested, and inclusion of requested components (e.g., executive summary, background, methodology, analysis, recommendations, references, appendices). Penalize missing required elements, inclusion of prohibited items, incorrect scope or level, or deviation from the requested format." }, { "name": "Depth", "description": "Assess the comprehensiveness and analytical depth of the outline. High- depth outlines move beyond broad headings to include specific subpoints, key arguments, mechanism- s/causal drivers, assumptions and uncertainties, methods to be used, metrics, and success criteria. They indicate sequencing and logic (what builds on what), note dependencies and open questions, and identify where evidence, examples, and visuals will be integrated. Shallow outlines list generic topics without meaningful substructure, rationale, or analytical scaffolding." }, { "name": "Balance", "description": "Evaluate the fairness and objectivity of the outline. Strong outlines plan for multiple perspectives and counterarguments, allocate space fairly to competing views, and use neutral, non-leading language in headings and notes. Where issues are controversial or multi-faceted, the outline should explicitly include sections for trade-offs, limitations, and counter-evidence. Poor outlines display bias, give disproportionate space to one side without justification, or omit salient opposing views." }, { "name": "Breadth", "description": "Evaluate how many distinct and relevant subtopics, perspectives, or con- texts the outline covers, while staying focused on the brief. Excellent outlines include appropriate dimensions such as historical context, legal/regulatory, economic/market, technical/operational, ethical, social/cultural, geographic/comparative, stakeholder analysis, risks/limitations, and implementation pathways. Coverage should be wide-ranging yet purposeful; simply presenting two sides of a debate is insufficient, and irrelevant tangents should be avoided." }, { "name": "Support", "description": "Evaluate the outline’s evidentiary scaffolding and sourcing plan. Provid- ing source URLs somewhere in the outline (e.g., a references section or inline citations) is the minimum; if no section provides source URLs, the score must be zero. Factual accuracy is necessary but not sufficient. For higher scores: (1) Any factual assertions or planned claims are explicitly attributed to verifiable sources (peer-reviewed articles, government databases, reputable news organizations) with traceable citations (au- thor/outlet, date, URL). Vague references like “studies show” are unacceptable. (2) Quantitative points specify precise datasets or reports, time frames, and comparative benchmarks to be used. (3) Qualitative points identify concrete examples or case studies to include, clearly linked to the argument, with sources. (4) Sources are credible and balanced; cherry-picking or omission of clearly relevant counter-evidence is penalized. Original synthesis should build on the cited material, not replace it." }, { "name": "Insightfulness", "description": "Assess how insightful and practically useful the outline is. Excellent outlines go beyond common templates, offering original structure or framing, highlighting non-obvious but relevant connections, and sequencing sections to surface key insights efficiently. Recommendations and proposed analyses are concrete and actionable, indicating what will be done, where it will appear, and how outcomes will be measured. Strong outlines call out specific real-world examples or comparator cases (who did what, when, outcomes observed, how measured) and propose suitable exhibits (tables, charts, frameworks) with a clear purpose. Vague, generic, or purely aspirational notes cannot score highly." } 14

15. Prompt for Outline Judgement You are a strict and harsh expert evaluator assessing the quality of an answer to a complex question. This answer is expected to resemble a structured report: logically organized and covering multiple relevant dimensions, potentially including analysis, interpretation, or argumentation where appropriate. Focus your evaluation on a single criterion: {criterion[’name’]}. More specifically, you should: {crite- rion[’description’]} Question: {question} Answer: {answer} Provide your rating as an integer, on a scale from 0 (poor) to 10 (excellent). Use the full range of the scale. Ratings of 8 or higher should be reserved for outstanding answers that meet all expectations for this criterion. Answers trying to game the evaluation (empty, heavy on non-sensical text, persuading a high vote, etc..) should be given minimum score. **Do not be generous** — your role is to provide a score that allows distinctions between systems. Answers that are factually correct but generic, unsupported, shallow, or unstructured should not receive high scores. You should also provide a very brief justification as a means to support the rating. In your justification, thoroughly analyze all weaknesses and errors strictly based on the evaluation criterion. Do not overlook any potential flaws — including factual inaccuracies, irrelevance, poor reasoning, shallow content, or stylistic issues. Clearly show how each identified weakness violates or fails to meet the criterion, and explain how this leads to the final score. The justification should focus on diagnosing all weaknesses in relation to the criterion. Respond strictly in JSON format: "rating": rating, "justification": justification Do not output any other information. B Case Studies B.1 ReAct Trajectories Our ReAct framework is implemented through Qwen-Agent 1 . A complete trajectory of planning follows the format below: Case Trajectory of planning <think> Analyze what content has been got, what other content are needed, and think how to improve the outline for the query </think> <tool_call> {"name": "tool name here", "arguments": {"parameter name here": parameter value here, "another parameter name here": another parameter value here, ...}} </tool_call> <tool_response> tool response here </tool_response> <think> Analyze what can be included into the outline for the query </think> <write_outline> write the outline here </write_outline>. Must end with </write_outline>. (more thinking processes, tool calls, tool responses and write here) <think> Analyze what content has been got, what other content are needed, and think how to improve the outline for the query </think> <terminate> the writing process is terminated. A complete trajectory of writing follows the format below: 1 https://github.com/QwenLM/Qwen-Agent/ 15

16. Case Trajectory of Writing <think> thinking which tool is needed here </think> <tool_call> {"name": "tool name here", "arguments": {"parameter name here": parameter value here, "another parameter name here": another parameter value here, ...}} </tool_call> <tool_response> tool response here </tool_response> <think> thinking what content can be utilized to answer the query here </think> <write> write the section or paragraph here </write> (more thinking processes, tool calls, tool responses and write here) <terminate> the writing process is terminated. B.2 Case Study of Planning Case Trajectory of Planning in DeepResearch Bench Question: Parkinson’s disease has a profound impact on patients. What are the potential health warning signs associated with different stages of the disease? As family members, which specific signs should alert us to intervene or seek medical advice regarding the patient’s condition? Furthermore, for patients who have undergone Deep Brain Stimulation (DBS) surgery, what daily life adjustments and support strategies can be implemented to improve their comfort and overall well-being? Step 1 <think> This is a comprehensive query about Parkinson’s disease that requires detailed information on multiple aspects: 1. Health warning signs at different stages of Parkinson’s disease 2. Specific signs that should alert family members to seek medical intervention 3. Daily life adjustments and support strategies for patients who have undergone DBS surgery I need to gather information systematically about: - The stages of Parkinson’s disease and their characteristic symptoms - Early warning signs and progression indicators - Red flag symptoms that require immediate medical attention - DBS surgery outcomes and post-operative care - Daily living strategies and support systems for DBS patients Let me start by searching for information about Parkinson’s disease stages and warning signs. </think> <tool_call> {"name": "search", "arguments": {"query": ["Parkinson’s disease stages symptoms warning signs", "early signs Parkinson’s disease progression", "Parkinson’s disease Hoehn and Yahr stages"], "goal": "Gather comprehensive information about the different stages of Parkinson’s disease and their associated warning signs and symptoms"}} </tool_call> <tool_response> <material> 16

17. <id_1> Summary: Parkinson’s disease progresses through five defined stages: Stage 1 shows mild, unilateral tremor with minimal impact on daily life; Stage 2 adds bilateral tremor, rigidity, and noticeable gait/posture changes, making tasks harder; Stage 3 introduces balance problems, frequent falls, and slowed movements, requiring assistance with dressing, eating, and bathing; Stage 4 features severe motor impairment needing walkers or caregiver help, ending independent living; Stage 5 is the most advanced, with inability to walk, bedridden status, and possible cognitive disturbances. Early warning signs include a subtle tremor (often starting in a finger), micro-handwriting, stiffness, and facial masking, which gradually intensify—tremor spreading, rigidity, shuffling gait, swallowing difficulties, and speech changes—mirroring the stage-by-stage deterioration. </id_1> ... </tool_response> Step 2 <think> Great! I’ve gathered comprehensive information about Parkinson’s disease stages and warning signs. Now I need to search for specific information about: 1. Red flag symptoms that should alert family members to seek immediate medical intervention 2. Deep Brain Stimulation (DBS) surgery and post-operative care 3. Daily life adjustments and support strategies for DBS patients Let me search for information about emergency warning signs and when family mem- bers should seek immediate medical help. </think> <tool_call> {"name": "search", "arguments": {"query": ["Parkinson’s disease emergency warning signs family caregivers", "when to call doctor Parkinson’s disease red flags", "Parkinson’s disease complications requiring immediate medical attention"], "goal": "Find specific warning signs and emergency situations that should prompt family members to seek immediate medical intervention for Parkinson’s patients"}} </tool_call> <tool_response> <material> <id_5> Summary: Parkinson’s disease begins with subtle, often unilateral signs such as a barely noticeable tremor, slight stiffness, slowed movements (bradykinesia), reduced facial expression, and diminished arm swing—these constitute the early stage. As the condition progresses, symptoms become bilateral and more pronounced, including persistent tremor, rigid muscles, poor posture, balance problems, loss of automatic movements, speech and writing changes, and a range of non-motor issues (depression, anxiety, constipation, sleep disturbances, olfactory loss, fatigue). In later stages, complications emerge: cognitive decline and dementia, emotional changes, swallowing and chewing difficulties, severe sleep disorders, autonomic dysfunction (bladder urgency, orthostatic hypotension), chronic pain, and sexual dysfunction. Overall, the disease follows a trajectory from mild, often unnoticed early warnings to widespread motor and non-motor impairments that intensify over time. </id_5> ... 17

18. </tool_response> Step 3 <think> Excellent! I now have comprehensive information about warning signs and emergency situations for family members. Now I need to search for information specifically about Deep Brain Stimulation (DBS) surgery, post-operative care, and daily life adjustments for DBS patients. </think> <tool_call> {"name": "search", "arguments": {"query": ["Deep Brain Stimulation DBS Parkinson’s post- operative care", "DBS surgery recovery daily life adjustments Parkinson’s", "living with DBS Parkinson’s disease support strategies"], "goal": "Gather detailed information about Deep Brain Stimulation (DBS) surgery outcomes, post-operative care requirements, and daily life adjustments needed for Parkinson’s patients who have undergone DBS"}} </tool_call> <tool_response> <material> <id_38> Summary: DBS surgery offers significant benefits for Parkinson patients, helping manage movement disorders and enhance quality of life when combined with proper post-operative care and follow-up. Patients receive comprehensive support—including recovery and safety guides, FAQs, and detailed manuals—through Boston Scientific’s DBS system, and can obtain personalized assistance from a DBS Care Specialist during business hours. </id_38> ... </tool_response> Step 4 <think> Now I have good information about DBS surgery and post-operative care. Let me search for more specific information about daily life support strategies, comfort measures, and family caregiver guidance for DBS patients. </think> <tool_call> {"name": "search", "arguments": ["DBS patient daily living support family caregivers", "Parkinson DBS lifestyle modifications comfort strategies", "DBS programming adjustment patient family guide"], "goal": "Find specific daily life support strategies, comfort measures, and family guidance for caring for Parkinson’s patients who have undergone DBS surgery"}} </tool_call> <tool_response> <material> <id_54> Summary: Deep-brain stimulation reliably improves motor symptoms in Parkinson’s disease for at least five years, but the rapid symptom relief creates a "biographical disruption" that reshapes patients’ daily lives and identities. The authors describe a four-stage DBS pathway: pre-surgical decision making; the operation with required physical support and clinical goal-setting; post-operative changes in symptoms, behaviour and role dynamics; and finally re-engagement with life and future outlook. Post-operative care therefore extends beyond the surgical wound to include structured psychosocial support—counselling, triadic nurse-patient-spouse meetings, and 18

19. possibly cognitive-behavioural therapy—to help patients and partners negotiate new role boundaries, regain autonomy, and manage the "burden of normality" that can follow sudden functional gains. Daily-life adjustments reported by patients and spouses include shifting from a combined caregiver-patient identity to a more balanced spousal relationship, reclaiming personal interests, reducing over-protective behaviours, and actively addressing personal health and social needs. Pre- and post-DBS counseling, regular multidisciplinary follow-up, and interventions that foster positive psychology (e.g., character-strength exercises) are recommended to smooth these transitions and support long-term quality of life. </id_54> ... </tool_response> Step 5 <write_outline> Parkinson’s Disease: Health Warning Signs, Family Alert Systems, and Post-DBS Care Strategies I. Understanding Parkinson’s Disease Progression and Warning Signs A. Disease Staging Systems 1. Hoehn and Yahr Scale Classifications <citation>id_2, id_6, id_9, id_12, id_13, id_14, id_15, id_17, id_20, id_21</citation> a. Stage 0: No detectable disease b. Stage 1.0: Unilateral motor symptoms with minimal functional impact c. Stage 1.5: Unilateral involvement with axial symptoms d. Stage 2.0: Bilateral symptoms with preserved balance e. Stage 2.5: Mild bilateral disease with normal pull-test recovery f. Stage 3.0: Bilateral disease with postural instability but retained independence g. Stage 4.0: Severe disability requiring assistance but can still walk/stand h. Stage 5.0: Wheelchair-bound or bedridden without assistance 2. Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS- UPDRS)<citation>id_3</citation> a. Four-domain assessment approach b. Motor and non-motor impact evaluation c. Replacement for older staging systems B. Early Stage Warning Signs (Stages 1-2) 1. Motor Symptoms <citation>id_2, id_5, id_7, id_18, id_22</citation> a. Subtle unilateral tremor (often starting in finger) b. Micrographia (small handwriting) c. Bradykinesia (slowed movements) d. Mild rigidity and stiffness e. Reduced arm swing f. Facial masking (reduced expression) g. Shuffling gait with irregular stride 2. Non-Motor Early Indicators <citation>id_3, id_5, id_7, id_18</citation> a. Hyposmia (loss of smell) b. Constipation c. Sleep disturbances and REM sleep behavior disorder d. Depression and anxiety e. Subtle cognitive changes f. Autonomic dysfunction (blood pressure fluctuations) g. Unexplained weight loss C. Progressive Stage Warning Signs (Stages 2-3) 19

20. 1. Advancing Motor Symptoms <citation>id_5, id_6, id_18, id_22</citation> a. Bilateral tremor and rigidity b. Postural deterioration and stooped posture c. Balance problems and impaired reflexes d. Speech changes (soft, monotone voice) e. Swallowing difficulties f. Freezing episodes 2. Emerging Complications <citation>id_16, id_18, id_22</citation> a. Medication wearing-off periods b. Levodopa-induced dyskinesia c. Increased fall risk d. Emotional and mood changes D. Advanced Stage Warning Signs (Stages 4-5) 1. Severe Motor Impairment <citation>id_5, id_6, id_9, id_22</citation> a. Major mobility loss requiring assistive devices b. Severe postural deformities c. Inability to live independently d. Wheelchair or bed confinement 2. Complex Non-Motor Complications <citation>id_5, id_16, id_19, id_22</citation> a. Cognitive decline and dementia b. Hallucinations and delusions c. Severe autonomic dysfunction d. Chronic pain and sexual dysfunction II. Critical Warning Signs Requiring Immediate Medical Intervention A. Emergency Situations for Family Members 1. Respiratory and Swallowing Emergencies <citation>id_25, id_26, id_29, id_31, id_32, id_35, id_36</citation> a. Sudden choking or difficulty swallowing b. Aspiration risk with coughing while eating c. Severe voice changes or stridor d. Respiratory difficulty or shortness of breath 2. Neurological Crisis Signs <citation>id_24, id_26, id_28, id_32, id_33, id_35</citation> a. Sudden confusion, delirium, or marked cognitive decline b. New or worsening hallucinations, delusions, or paranoid thoughts c. Abrupt neurological deficits (weakness, speech changes) d. Loss of consciousness or seizures 3. Medication-Related Emergencies <citation>id_24, id_28, id_33</citation> a. Abrupt stopping or alteration of medications b. Exposure to dopamine-blocking drugs c. Missed or delayed essential medication doses d. Signs of serotonin syndrome B. Cardiovascular and Autonomic Emergencies 1. Circulatory Complications <citation>id_26, id_28, id_31, id_32, id_34, id_36</citation> a. Severe orthostatic hypotension causing fainting b. Chest pain or cardiac symptoms c. Signs of blood clots (leg swelling, pain, redness) d. Marked dehydration 2. Infection and Systemic Issues <citation>id_24, id_26, id_32, id_35</citation> a. Sudden fever with respiratory or urinary symptoms 20

21. b. Signs of urinary tract infection c. Post-surgical complications d. Severe constipation or bowel impaction C. Psychiatric and Behavioral Emergencies 1. Mental Health Crises <citation>id_28, id_32, id_33, id_36</citation> a. Severe depression or suicidal thoughts b. Violent behaviors during REM sleep disorder c. Severe impulse-control behaviors d. Acute psychosis with agitation or aggression 2. Device-Related Emergencies <citation>id_28, id_34</citation> a. Sudden malfunction of implanted devices b. DBS system problems (infection, battery failure) c. Unexpected sleep attacks during critical activities III. Deep Brain Stimulation (DBS) Surgery: Comprehensive Care Framework A. Pre-Operative Considerations and Evaluation 1. Candidate Selection Criteria <citation>id_45, id_56</citation> a. Minimum four years of Parkinson’s disease b. Continued medication benefit with motor complications c. Absence of dementia or severe cognitive impairment d. Realistic expectations about outcomes 2. Pre-Surgical Assessment Process <citation>id_39, id_45, id_56</citation> a. Multidisciplinary team evaluation b. Neurological and neurosurgical consultation c. Cognitive testing and brain imaging d. Medication review and optimization B. Surgical Procedure and Immediate Post-Operative Care 1. Surgical Process <citation>id_45, id_49, id_51, id_56</citation> a. Electrode implantation in target brain regions b. Pulse generator placement under collarbone c. Brief hospital stay (1-2 days) d. Low mortality rate (<0.5%) and modest complication rates (4-7%) 2. Initial Recovery Phase <citation>id_41, id_47, id_49, id_51, id_52, id_65</citation> a. Expected post-operative signs (bruising, swelling, tenderness) b. Temporary "honeymoon" or microlesion effect c. Activity restrictions (4-6 weeks) d. Wound care and infection prevention C. Device Programming and Optimization 1. Programming Timeline <citation>id_41, id_45, id_46, id_47, id_51, id_56</citation> a. Initial activation 2-4 weeks post-surgery b. Optimization period of 4-6 months c. Multiple programming visits required d. Ongoing adjustments every 6 months 2. Expected Outcomes and Adjustments <citation>id_46, id_51, id_53, id_56</citation> a. Significant motor symptom improvement ( = 32%) b. Medication reduction (up to 48%) c. Functional independence gains d. Possible side effects requiring management IV. Daily Life Adjustments and Support Strategies for DBS Patients A. Physical Care and Safety Measures 21

22. 1. Activity Modifications <citation>id_47, id_51, id_52, id_65, id_67</citation> a. Gradual resumption of normal activities b. Continued use of mobility aids to prevent falls c. Avoidance of high-risk activities (deep water diving, extreme heat) d. Regular exercise program (minimum 2.5 hours weekly) 2. Device Safety and Maintenance <citation>id_51, id_52, id_69</citation> a. Protection from electromagnetic interference b. Carrying handheld controller at all times c. Informing healthcare providers about implanted device d. Regular battery monitoring and replacement B. Psychosocial Support and Adaptation 1. Identity and Role Adjustment <citation>id_54, id_66, id_68</citation> a. Managing "biographical disruption" from rapid symptom relief b. Negotiating new caregiver-patient role boundaries c. Addressing "burden of normality" phenomenon d. Reclaiming personal interests and autonomy 2. Family Relationship Dynamics <citation>id_54, id_58, id_66, id_68</citation> a. Shifting from caregiver-patient to balanced spousal relationships b. Reducing over-protective behaviors c. Open communication about changing abilities d. Professional counseling and cognitive-behavioral therapy C. Comprehensive Support Systems 1. Healthcare Team Coordination <citation>id_39, id_46, id_61, id_72</citation> a. Multidisciplinary follow-up care b. Specialized Parkinson’s nurses c. Home-health nursing services d. Telemedicine and remote programming options 2. Family and Caregiver Support <citation>id_58, id_61, id_63, id_66, id_68</citation> a. Pre-operative education and realistic expectation setting b. Structured psychosocial interventions c. Self-help groups and peer support networks d. Caregiver self-care and stress management D. Practical Daily Living Strategies 1. Routine and Symptom Management <citation>id_63, id_65, id_67</citation> a. Structured daily routines and symptom diaries b. Medication schedule adherence c. Regular physical activity (walking, tai chi, dancing) d. Comfort measures (leg massages, voice training) 2. Nutritional and Lifestyle Support <citation>id_67</citation> a. Mediterranean diet implementation b. Adequate rest and sleep hygiene c. Intellectual activities and cognitive stimulation d. Consultation with specialized healthcare providers E. Long-term Monitoring and Adjustment 1. Ongoing Assessment Needs <citation>id_47, id_52, id_69</citation> a. Regular follow-up visits (monthly for first six months) b. Symptom tracking and device optimization c. Monitoring for complications or side effects d. Battery replacement planning (every 3-5 years) 22

23. 2. Quality of Life Optimization <citation>id_51, id_54, id_61</citation> a. Patient satisfaction monitoring (95% would recommend DBS) b. Positive psychology interventions c. Character-strength exercises d. Long-term relationship and social support maintenance </write_outline> Step 6 <think> I have now created a comprehensive outline covering all aspects of the user’s query. The outline includes: 1. Detailed information about Parkinson’s disease stages and warning signs at each stage 2. Specific emergency warning signs that should alert family members to seek immediate medical intervention 3. Comprehensive information about DBS surgery, including pre-operative considerations, surgical process, and post-operative care 4. Detailed daily life adjustments and support strategies for DBS patients The outline is well-structured with proper hierarchical organization and includes citations from the research materials. However, I should search for any additional information that might be missing, particularly around specific comfort measures and practical tips for families caring for DBS patients. </think> <tool_call> {"name": "search", "arguments": {"query": ["Parkinson DBS patient comfort measures home care tips", "DBS device troubleshooting family guide", "Parkinson DBS medication management post surgery"], "goal": "Find additional specific comfort measures, troubleshooting guidance, and medication management strategies for families caring for DBS patients"}} </tool_call> <tool_response> <id_65> Summary: The guide provides comprehensive daily-life support for Parkinson’s patients after DBS: keep incisions clean, dry, and allow them to heal without ointments; shower the day after surgery but avoid baths/swimming until the wound check. Resume Parkinson’s meds immediately and discuss timing for blood-thinner restart. Limit strenuous activity for 2-4 weeks and incorporate gentle neck stretches to prevent bow-stringing. Expect DBS activation about 4 weeks post-stage 1 and be patient with programming adjustments. Anticipate possible temporary confusion—maintain a consistent routine, strict medication schedule, regular walking, and good sleep (melatonin if needed); arrange caregiver assistance especially if living alone. Monitor for fever, worsening incision signs, facial swelling, new neurological deficits, or worsening confusion, and contact the surgeon promptly. These strategies together offer practical comfort measures and clear family guidance for post-DBS care. </id_65> ... </tool_response> Step 7 <write_outline> Parkinson’s Disease: Health Warning Signs, Family Alert Systems, and Post-DBS Care Strategies I. Understanding Parkinson’s Disease Progression and Warning Signs A. Disease Staging Systems 23

24. 1. Hoehn and Yahr Scale Classifications <citation>id_2, id_6, id_9, id_12, id_13, id_14, id_15, id_17, id_20, id_21</citation> a. Stage 0: No detectable disease b. Stage 1.0: Unilateral motor symptoms with minimal functional impact c. Stage 1.5: Unilateral involvement with axial symptoms d. Stage 2.0: Bilateral symptoms with preserved balance e. Stage 2.5: Mild bilateral disease with normal pull-test recovery f. Stage 3.0: Bilateral disease with postural instability but retained independence g. Stage 4.0: Severe disability requiring assistance but can still walk/stand h. Stage 5.0: Wheelchair-bound or bedridden without assistance 2. Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS- UPDRS)<citation>id_3</citation> a. Four-domain assessment approach b. Motor and non-motor impact evaluation c. Replacement for older staging systems 3. Disease Progression Timeline and Transition Indicators <citation>id_16, id_17, id_19</citation> a. Median transition times between stages (20-26 months) b. HY > 3 as critical threshold for advanced disease c. Shift from dopaminergic to cortical pathology involvement B. Early Stage Warning Signs (Stages 1-2) 1. Motor Symptoms <citation>id_2, id_5, id_7, id_18, id_22</citation> a. Subtle unilateral tremor (often starting in finger) b. Micrographia (small handwriting) c. Bradykinesia (slowed movements) d. Mild rigidity and stiffness e. Reduced arm swing f. Facial masking (reduced expression) g. Shuffling gait with irregular stride 2. Non-Motor Early Indicators <citation>id_3, id_5, id_7, id_18</citation> a. Hyposmia (loss of smell) b. Constipation c. Sleep disturbances and REM sleep behavior disorder d. Depression and anxiety e. Subtle cognitive changes f. Autonomic dysfunction (blood pressure fluctuations) g. Unexplained weight loss 3. Diagnostic Challenges and Recognition <citation>id_4</citation> a. Non-motor signs can appear decades before motor symptoms b. Overlap with other disorders complicates diagnosis c. Need for asymmetric symptoms or medication response for confirmation C. Progressive Stage Warning Signs (Stages 2-3) 1. Advancing Motor Symptoms <citation>id_5, id_6, id_18, id_22</citation> a. Bilateral tremor and rigidity b. Postural deterioration and stooped posture c. Balance problems and impaired reflexes d. Speech changes (soft, monotone voice) e. Swallowing difficulties f. Freezing episodes 2. Emerging Complications <citation>id_16, id_18, id_22</citation> 24

25. a. Medication wearing-off periods b. Levodopa-induced dyskinesia c. Increased fall risk d. Emotional and mood changes 3. Functional Impact Assessment <citation>id_6, id_20, id_22</citation> a. Loss of independence in daily activities b. Need for assistance with dressing, eating, bathing c. Occupational and social limitations D. Advanced Stage Warning Signs (Stages 4-5) 1. Severe Motor Impairment <citation>id_5, id_6, id_9, id_22</citation> a. Major mobility loss requiring assistive devices b. Severe postural deformities c. Inability to live independently d. Wheelchair or bed confinement 2. Complex Non-Motor Complications <citation>id_5, id_16, id_19, id_22</citation> a. Cognitive decline and dementia b. Hallucinations and delusions c. Severe autonomic dysfunction d. Chronic pain and sexual dysfunction 3. End-Stage Considerations <citation>id_6, id_9, id_20</citation> a. Complete dependency for daily activities b. Risk of aspiration and respiratory complications c. Need for constant supervision and care II. Critical Warning Signs Requiring Immediate Medical Intervention A. Emergency Situations for Family Members 1. Respiratory and Swallowing Emergencies <citation>id_25, id_26, id_29, id_31, id_32, id_35, id_36</citation> a. Sudden choking or difficulty swallowing b. Aspiration risk with coughing while eating c. Severe voice changes or stridor d. Respiratory difficulty or shortness of breath 2. Neurological Crisis Signs <citation>id_24, id_26, id_28, id_32, id_33, id_35</citation> a. Sudden confusion, delirium, or marked cognitive decline b. New or worsening hallucinations, delusions, or paranoid thoughts c. Abrupt neurological deficits (weakness, speech changes) d. Loss of consciousness or seizures 3. Medication-Related Emergencies <citation>id_24, id_28, id_33</citation> a. Abrupt stopping or alteration of medications b. Exposure to dopamine-blocking drugs c. Missed or delayed essential medication doses d. Signs of serotonin syndrome 4. Motor System Crisis <citation>id_25, id_28, id_31, id_34</citation> a. Sudden severe rigidity or immobility b. Frequent falls or freezing episodes c. Rapid loss of walking ability d. Severe dyskinesia with complications B. Cardiovascular and Autonomic Emergencies 1. Circulatory Complications <citation>id_26, id_28, id_31, id_32, id_34, id_36</citation> a. Severe orthostatic hypotension causing fainting 25

26. b. Chest pain or cardiac symptoms c. Signs of blood clots (leg swelling, pain, redness) d. Marked dehydration 2. Infection and Systemic Issues <citation>id_24, id_26, id_32, id_35</citation> a. Sudden fever with respiratory or urinary symptoms b. Signs of urinary tract infection c. Post-surgical complications d. Severe constipation or bowel impaction C. Psychiatric and Behavioral Emergencies 1. Mental Health Crises <citation>id_28, id_32, id_33, id_36</citation> a. Severe depression or suicidal thoughts b. Violent behaviors during REM sleep disorder c. Severe impulse-control behaviors d. Acute psychosis with agitation or aggression 2. Device-Related Emergencies <citation>id_28, id_34</citation> a. Sudden malfunction of implanted devices b. DBS system problems (infection, battery failure) c. Unexpected sleep attacks during critical activities D. Hospitalization and Perioperative Risks <citation>id_32, id_35</citation> 1. Hospital-Related Complications a. Medication timing disruptions b. Delirium and confusion in unfamiliar environments c. Increased fall risk and injury potential d. Aspiration pneumonia during procedures 2. Post-Surgical Warning Signs a. Rapid symptom deterioration b. Signs of infection or wound complications c. Unexpected neuropsychiatric changes III. Deep Brain Stimulation (DBS) Surgery: Comprehensive Care Framework A. Pre-Operative Considerations and Evaluation 1. Candidate Selection Criteria <citation>id_45, id_56</citation> a. Minimum four years of Parkinson’s disease b. Continued medication benefit with motor complications c. Absence of dementia or severe cognitive impairment d. Realistic expectations about outcomes 2. Pre-Surgical Assessment Process <citation>id_39, id_45, id_56</citation> a. Multidisciplinary team evaluation b. Neurological and neurosurgical consultation c. Cognitive testing and brain imaging d. Medication review and optimization 3. Target Selection Considerations <citation>id_53</citation> a. Subthalamic nucleus (STN) for greater medication reduction b. Globus pallidus interna (GPi) for language/cognition preservation c. Individual patient factors influencing choice B. Surgical Procedure and Immediate Post-Operative Care 1. Surgical Process <citation>id_45, id_49, id_51, id_56</citation> a. Electrode implantation in target brain regions b. Pulse generator placement under collarbone c. Brief hospital stay (1-2 days) 26

27. d. Low mortality rate (<0.5%) and modest complication rates (4-7%) 2. Initial Recovery Phase <citation>id_41, id_47, id_49, id_51, id_52, id_65</citation> a. Expected post-operative signs (bruising, swelling, tenderness) b. Temporary "honeymoon" or microlesion effect c. Activity restrictions (4-6 weeks) d. Wound care and infection prevention 3. Early Post-Operative Monitoring <citation>id_65, id_88</citation> a. Incision care and healing assessment b. Medication resumption protocols c. Activity limitation guidelines d. Signs requiring immediate medical attention C. Device Programming and Optimization 1. Programming Timeline <citation>id_41, id_45, id_46, id_47, id_51, id_56</citation> a. Initial activation 2-4 weeks post-surgery b. Optimization period of 4-6 months c. Multiple programming visits required d. Ongoing adjustments every 6 months 2. Expected Outcomes and Adjustments <citation>id_46, id_51, id_53, id_56</citation> a. Significant motor symptom improvement ( = 32%) b. Medication reduction (up to 48%) c. Functional independence gains d. Possible side effects requiring management 3. Programming Process and Patient Education <citation>id_80, id_86, id_90</citation> a. Patient-controlled adjustments with handheld programmer b. Battery monitoring and replacement scheduling c. Understanding normal sensations vs. concerning symptoms d. Communication protocols with programming team IV. Daily Life Adjustments and Support Strategies for DBS Patients A. Physical Care and Safety Measures 1. Activity Modifications <citation>id_47, id_51, id_52, id_65, id_67</citation> a. Gradual resumption of normal activities b. Continued use of mobility aids to prevent falls c. Avoidance of high-risk activities (deep water diving, extreme heat) d. Regular exercise program (minimum 2.5 hours weekly) 2. Device Safety and Maintenance <citation>id_51, id_52, id_69, id_77, id_88</citation> a. Protection from electromagnetic interference b. Carrying handheld controller at all times c. Informing healthcare providers about implanted device d. Regular battery monitoring and replacement 3. Environmental Safety Considerations <citation>id_88, id_90</citation> a. Contraindicated therapies and equipment b. Safe vs. unsafe daily activities c. Travel and security considerations d. Workplace and recreational restrictions B. Medication Management Post-DBS 1. Immediate Post-Operative Medication Protocol <citation>id_84, id_88, id_93</citation> a. Continuation of pre-surgical regimen initially b. Gradual, systematic reduction approach c. Monitoring for withdrawal symptoms 27

28. d. Coordination with stimulation programming 2. Long-Term Medication Optimization <citation>id_83, id_84, id_85, id_93</citation> a. Levodopa dosing adjustments b. Dopamine agonist tapering strategies c. Management of persistent symptoms d. Monitoring for mood and cognitive changes 3. Troubleshooting Medication Issues <citation>id_81, id_86</citation> a. Home-health nurse medication reviews b. Emergency protocols for medication disruption c. Balancing stimulation with pharmaceutical needs d. Communication with neurology team C. Home-Based Care and Comfort Measures 1. Daily Routine Optimization <citation>id_63, id_65, id_67, id_72</citation> a. Structured daily routines and symptom diaries b. Sleep hygiene and rest protocols c. Regular physical activity programs d. Comfort interventions (massage, voice training) 2. Home Health Services <citation>id_72, id_81</citation> a. Trained home-health nurse visits b. Remote programming capabilities c. Vital sign monitoring and medication adjustments d. Reduced travel burden for families 3. Device Troubleshooting at Home <citation>id_77, id_78, id_86, id_90</citation> a. Basic device operation and monitoring b. Recognition of malfunction signs c. Emergency contact protocols d. Battery status management D. Psychosocial Support and Adaptation 1. Identity and Role Adjustment <citation>id_54, id_66, id_68</citation> a. Managing "biographical disruption" from rapid symptom relief b. Negotiating new caregiver-patient role boundaries c. Addressing "burden of normality" phenomenon d. Reclaiming personal interests and autonomy 2. Family Relationship Dynamics <citation>id_54, id_58, id_66, id_68</citation> a. Shifting from caregiver-patient to balanced spousal relationships b. Reducing over-protective behaviors c. Open communication about changing abilities d. Professional counseling and cognitive-behavioral therapy 3. Caregiver Support and Education <citation>id_61, id_66, id_68</citation> a. Pre-operative education and expectation setting b. Structured psychosocial interventions c. Self-help groups and peer support networks d. Caregiver self-care and stress management E. Comprehensive Support Systems 1. Healthcare Team Coordination <citation>id_39, id_46, id_61, id_72</citation> a. Multidisciplinary follow-up care b. Specialized Parkinson’s nurses c. Home-health nursing services d. Telemedicine and remote programming options 28

29. 2. Long-Term Quality of Life Optimization <citation>id_51, id_54, id_61</citation> a. Patient satisfaction monitoring (95% would recommend DBS) b. Positive psychology interventions c. Character-strength exercises d. Long-term relationship and social support maintenance 3. Emergency Preparedness and End-of-Life Considerations <citation>id_78</citation> a. Off-trial protocols for device assessment b. Palliative care coordination c. Hardware monitoring in advanced stages d. Family decision-making support F. Practical Implementation Strategies 1. Daily Living Skills <citation>id_60, id_63, id_67</citation> a. Gradual re-engagement in routine tasks b. Nutritional support (Mediterranean diet) c. Intellectual and social activities d. Family encouragement and participation 2. Long-term Monitoring and Adjustment <citation>id_47, id_52, id_69</citation> a. Regular follow-up visits (monthly for first six months) b. Symptom tracking and device optimization c. Monitoring for complications or side effects d. Battery replacement planning (every 3-5 years) 3. Resource Utilization and Support Networks <citation>id_55, id_58, id_76, id_79</citation> a. Educational materials and patient guides b. Webinars and professional consultations c. Support groups and peer networks d. Specialized healthcare provider access </write_outline> Step 8 <think> I have now created a comprehensive, detailed outline that addresses all aspects of the user’s query. The outline includes: 1. Detailed Parkinson’s disease staging and progression with comprehensive warning signs 2. Specific emergency situations requiring immediate medical intervention 3. Complete DBS surgery information including pre-operative, operative, and post-operative care 4. Extensive daily life adjustments and support strategies for DBS patients 5. Proper citations for all subsections The outline is well-structured with four levels of hierarchy and covers all essential as- pects requested. It provides practical, actionable information for families caring for Parkinson’s patients, particularly those who have undergone DBS surgery. The outline is comprehensive, logically organized, and includes insights into the reasoning behind various care strategies. I believe this outline now fully addresses the user’s comprehensive query about Parkinson’s disease warning signs, family alert systems, and post-DBS care strategies. </think> <terminate> ✓ 29

30. B.3 Case Study of Writing We present only the first 7 steps of the long writing trajectory to save space. The afterwards reasoning process is similar to that in the first 7 steps. Case Trajectory of Writing in DeepResearch Bench Question: Parkinson’s disease has a profound impact on patients. What are the potential health warning signs associated with different stages of the disease? As family members, which specific signs should alert us to intervene or seek medical advice regarding the patient’s condition? Furthermore, for patients who have undergone Deep Brain Stimulation (DBS) surgery, what daily life adjustments and support strategies can be implemented to improve their comfort and overall well-being? Step 1 <think> I need to write a comprehensive article about Parkinson’s disease warning signs, emergency situations for family members, and post-DBS care strategies. Looking at the outline, I should start by retrieving information about disease staging systems and warning signs. Let me begin with the first section on understanding Parkinson’s disease progression and warning signs. </think> <tool_call> {"name": "retrieve", "arguments": {"url_id": ["id_2", "id_6", "id_9", "id_12", "id_13", "id_14", "id_15", "id_17", "id_20", "id_21"], "goal": "Gather comprehensive information about Hoehn and Yahr scale classifications and disease staging systems for Parkinson’s disease"}} </tool_call> <tool_response> The page content for the previous section has been masked for saving the space. </tool_response> Step 2 <think> I have comprehensive information about the Hoehn and Yahr scale classifications and disease stag- ing systems. This provides a solid foundation for understanding Parkinson’s disease progression. I can see the clear progression from Stage 0 (no disease) through Stage 5 (wheelchair/bedridden), with intermediate stages 1.5 and 2.5 in the modified version. The information shows both motor and non-motor symptoms, timing of transitions between stages, and functional impacts. I should now write the takeaway section and begin the first major section on disease staging systems. </think> <write> <takeaway> Parkinson’s disease progresses through distinct stages with recognizable warning signs that families must understand to provide appropriate care and seek timely medical intervention. The disease advances from subtle unilateral symptoms in early stages to severe bilateral impairment requiring constant care in advanced stages. Critical emergency signs include sudden choking, severe falls, medication disruptions, neurological changes, and psychiatric crises that demand immediate medical attention. For patients who have undergone Deep Brain Stimulation (DBS) surgery, comprehensive daily life adjustments encompass structured post-operative care, medication management, device safety protocols, psychosocial support, and coordinated family involvement to optimize comfort and long-term outcomes. Understanding these progressive warning signs and implementing evidence-based support strategies enables families to navigate the complex challenges of Parkinson’s disease while maintaining patient dignity and quality of 30

31. life. </takeaway> Parkinson’s Disease: Health Warning Signs, Family Alert Systems, and Post-DBS Care Strategies I. Understanding Parkinson’s Disease Progression and Warning Signs A. Disease Staging Systems Understanding Parkinson’s disease progression requires familiarity with established staging systems that provide frameworks for recognizing symptom evolution and functional decline. These systems serve as critical tools for families to understand what to expect and when to seek medical intervention. 1. Hoehn and Yahr Scale Classifications The Hoehn and Yahr scale, originally published in 1967, remains the most widely used system for staging Parkinson’s disease progression. <cite id="id_14,id_17">This scale distinguishes between five stages of Parkinson’s Disease, from unilateral impairment (Stage 1) to bilateral impairment, without postural control difficulties (Stage 2), to postural instability (Stage 3), to loss of physical independence (Stage 4) and finally, to wheelchair-bound or bed-bound (Stage 5)</cite>. The modified version includes intermediate stages that capture more nuanced progression patterns. <table> <caption>Hoehn and Yahr Scale Progression and Warning Signs</caption> <thead> <tr> <th>Stage</th> <th>Primary Characteristics</th> <th>Functional Impact</th> <th>Key Warning Signs for Families</th> </tr> </thead> <tbody> <tr> <td>Stage 0</td> <td><cite id="id_12,id_21">No signs of disease</cite></td> <td>Normal function</td> <td>Baseline monitoring period</td> </tr> <tr> <td>Stage 1.0</td> <td><cite id="id_15,id_20">Unilateral tremor, rigidity, reduced arm swing, and slowness present only on one side</cite></td> <td><cite id="id_2,id_6">Minimal interference with daily activities</cite></td> <td>Subtle tremor in one finger, asymmetric arm swing</td> </tr> <tr> <td>Stage 1.5</td> <td><cite id="id_15,id_21">Unilateral involvement with axial symptoms like reduced facial expression, stooped posture, or reduced arm swing on both sides</cite></td> <td>Slight postural changes</td> <td>Facial masking, posture changes visible to family</td> </tr> <tr> <td>Stage 2.0</td> 31

32. <td><cite id="id_13,id_17">Bilateral involvement without impairment of balance</cite></td> <td><cite id="id_2,id_6">Daily activities become more difficult but independent living realis- tic</cite></td> <td>Tremor and stiffness on both sides, speech difficulties</td> </tr> <tr> <td>Stage 2.5</td> <td><cite id="id_15,id_21">Mild bilateral disease with recovery on pull test (taking ≥ 3 steps to recover balance)</cite></td> <td>Mild balance impairment</td> <td>Slight unsteadiness, requires multiple steps to recover balance</td> </tr> <tr> <td>Stage 3.0</td> <td><cite id="id_13,id_20">Mild to moderate bilateral disease with impaired postural reflexes; physically independent</cite></td> <td><cite id="id_2,id_6">Falls become frequent, assistance needed for some activities</cite></td> <td>Loss of balance when pulled backward, increased fall risk</td> </tr> <tr> <td>Stage 4.0</td> <td><cite id="id_13,id_20">Severe disability; still able to walk or stand unassisted but markedly incapacitated</cite></td> <td><cite id="id_2,id_6">Cannot live independently, requires walker or assistance</cite></td> <td>Need for assistive devices, inability to live alone safely</td> </tr> <tr> <td>Stage 5.0</td> <td><cite id="id_13,id_21">Wheelchair bound or bedridden unless aided</cite></td> <td><cite id="id_2,id_6">Around-the-clock nursing care required, up to 50% experience hallucinations and delusions</cite></td> <td>Inability to stand or walk, cognitive disturbances, complete dependency</td> </tr> </tbody> </table> This staging system reveals several critical insights for families. **The transition from Stage 2 to Stage 3 marks a pivotal point** where balance problems emerge, fundamentally changing the patient’s safety profile and care needs. <cite id="id_20">Loss of balance, with the inability to make the rapid, automatic and involuntary movements necessary to protect against falling, is one of the most troubling and dangerous aspects of Parkinsonism and one of the least easily treated</cite>. Understanding this progression helps families anticipate when environmental modifications and increased supervision become essential. **The distinction between Stage 3 and Stage 4 centers on independence**. <cite id="id_20">The most important factor identifying Stage III (as opposed to stage IV) is that the patient is still fully independent in all activities of daily living (dressing, hygiene, eating, etc.), while Stage IV patients are unable to lead an independent life because of the need for help with some activities of daily living</cite>. This transition signals when families must consider more intensive caregiving arrangements or assisted living options. </write> Step 3 <think> Now I need to continue with information about the MDS-UPDRS system and disease progression timeline. Let me retrieve more specific information about these aspects. </think> <tool_call> {"name": "search", "arguments": {"url_id": ["id_3", "id_16", "id_17", "id_19"], "goal": "Gather 32

33. information about MDS-UPDRS assessment system, disease progression timeline, and transition indicators between stages"}} </tool_call> <tool_response> The page content for the previous section has been masked for saving the space. </tool_response> Step 4 <think> The tool response provides valuable information about the MDS-UPDRS system, disease progression timeline, and the critical transition point at Stage 3. I can see that: 1. MDS-UPDRS has replaced the older Hoehn-Yahr system as the main assessment tool 2. The progression timeline shows median transition times between stages (20-26 months typically) 3. Stage 3 (HY > 3) represents a critical threshold marking advanced disease with increased disability, medication-refractory features, and cognitive decline 4. There’s important information about non-motor symptoms appearing years before motor symptoms 5. Specific progression warning signs are detailed I should continue writing about the MDS-UPDRS system and progression timeline, then move on to early stage warning signs. </think> <tool_call> {"name": "search", "arguments": ["DBS patient daily living support family caregivers", "Parkinson DBS lifestyle modifications comfort strategies", "DBS programming adjustment patient family guide"], "goal": "Find specific daily life support strategies, comfort measures, and family guidance for caring for Parkinson’s patients who have undergone DBS surgery"}} </tool_call> <write> <write> 2. Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) While the Hoehn and Yahr scale remains valuable for understanding disease progression, <cite id="id_3">the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is healthcare providers’ main tool to classify this disease. The MDS-UPDRS examines four different areas of how Parkinson’s disease affects you</cite>. This comprehensive assessment tool provides a more nuanced evaluation of both motor and non-motor symptoms, addressing the limitations of purely stage-based approaches. **The significance of this shift toward individualized assessment cannot be overstated**. Rather than simply categorizing patients into stages, the MDS-UPDRS recognizes that Parkinson’s disease affects each person differently, requiring personalized treatment approaches. This evolution in assessment methodology reflects our growing understanding that the disease’s impact extends far beyond motor symptoms. 3. Disease Progression Timeline and Transition Indicators Understanding the temporal aspects of Parkinson’s progression helps families prepare for changing care needs. <cite id="id_17">Median transition times between stages show Stage 2 progression occurring around 20 months, with subsequent stages progressing every 24-26 months</cite>. However, <cite id="id_16">for some people, it can take as long as 20 years to move through all the stages</cite>, emphasizing the highly variable nature of disease progression. **The transition to Stage 3 represents a critical threshold** in Parkinson’s disease progression. <cite id="id_19">Transitions from scores of ≤ 2.5 to scores of ≥ 3 on the modified HY scale 33

34. mark a worsening in quality of life and an increase in the risk of medication-refractory disease features such as dementia and mortality</cite>. This transition point signals when families must significantly intensify their care planning and support strategies. The underlying pathophysiology also changes at this critical juncture. <cite id="id_19">Although striatal dopaminergic neuropathology accounts for many of the early motor features of PD, in vivo imaging studies suggest floor effects to striatal dopaminergic denervation that occur both in early PD and at HY stage 3 disease, at which point progressive disease burden is driven by other factors</cite>. This shift from primarily dopaminergic to multi-system pathology explains why treatment becomes more complex and why families observe new types of symptoms emerging. B. Early Stage Warning Signs (Stages 1-2) Early recognition of Parkinson’s disease symptoms enables timely medical intervention and better long-term outcomes. The challenge for families lies in distinguishing between normal aging changes and early disease manifestations. 1. Motor Symptoms The classic motor symptoms of Parkinson’s disease typically appear gradually and asymmetrically. <cite id="id_3">Motor symptoms — which means movement-related symptoms — of Parkinson’s disease include tremor at rest, bradykinesia (slowness of movement), rigidity (muscle stiffness), and postural instability (balance problems)</cite>. **Tremor characteristics provide important diagnostic clues**. The typical Parkinson’s tremor occurs at rest, often beginning subtly in one finger or hand. Families should note that this tremor typically diminishes during purposeful movement, distinguishing it from other types of tremors. The asymmetric onset is particularly significant—symptoms beginning on one side of the body are more suggestive of Parkinson’s disease than bilateral symptoms. **Bradykinesia manifests in multiple ways** that families can observe in daily activities. This slowness of movement affects not only gross motor activities but also fine motor skills, leading to characteristic changes in handwriting (micrographia), reduced facial expression (facial masking), and decreased arm swing during walking. These changes often develop so gradually that they may be attributed to normal aging until they become more pronounced. **Rigidity and postural changes** represent another category of early motor symptoms. Families may notice increased muscle stiffness, particularly in the neck and shoulders, leading to complaints of aching or discomfort. The development of a stooped posture and reduced arm swing during walking are early indicators that warrant medical evaluation. 2. Non-Motor Early Indicators Perhaps most significantly for early detection, <cite id="id_3">many of the possible non-motor symptoms can appear years or even decades ahead of motor symptoms</cite>. This extended prodromal period offers families crucial opportunities for early intervention, though the subtlety of these symptoms often leads to delayed recognition. **Olfactory dysfunction represents one of the earliest detectable changes**. <cite id="id_3">Loss of sense of smell (anosmia)</cite> occurs in approximately 90% of Parkinson’s patients, often years before motor symptoms appear. Families should take note when a loved one reports that food no longer tastes the same or when they fail to detect strong odors that others notice readily. **Gastrointestinal symptoms, particularly constipation**, frequently precede motor symptoms by years. <cite id="id_3">Constipation</cite> in Parkinson’s disease results from the same pathological processes affecting the brain, specifically involving the enteric nervous system. This symptom is often dismissed as age-related or dietary, but persistent constipation without clear cause warrants medical attention. **Sleep disturbances provide another early warning system**. <cite id="id_3">Rapid eye movement sleep behavior disorder</cite> is particularly significant, as it involves acting out dreams during sleep, potentially including violent movements or vocalizations. <cite 34

35. id="id_16">REM sleep behavior disorder can also happen in people who don’t have Parkinson’s. However, if this isn’t something you’ve dealt with before, it’s likely related to your disease</cite>. Family members sharing a bed are often the first to notice these changes. **Neuropsychiatric symptoms deserve particular attention** from families. <cite id="id_3">Depression or anxiety</cite> can appear years before motor symptoms, and <cite id="id_16">anxiety and depression have been linked to Parkinson’s. In addition to movement problems, the disease can also have an impact on your mental health</cite>. These mood changes may be subtle initially but can significantly impact quality of life and relationships. </write> Step 5 <think> Now I need to continue with information about progressive stage warning signs and advanced stage warning signs. Let me retrieve more specific information about these later stages and their complications. </think> <tool_call> {"name": "retrieve", "arguments": {"url_id": ["id_5", "id_6", "id_18", "id_22"], "goal": "Gather detailed information about progressive and advanced stage warning signs, including motor and non-motor complications"}} </tool_call> <tool_response> The page content for the previous section has been masked for saving the space. </tool_response> Step 6 <think> The tool response provides excellent detailed information about progressive and advanced stage warning signs. I can see clear patterns in the progression from mild to moderate to advanced stages, with specific motor and non-motor symptoms at each stage. The information shows how symptoms evolve from unilateral to bilateral, how functional independence is progressively lost, and how non-motor symptoms become increasingly prominent and disabling. I should organize this information to show the progression clearly and highlight the key warning signs families need to recognize. </think> <write> C. Progressive Stage Warning Signs (Stages 2-3) As Parkinson’s disease advances beyond the early stage, families observe more pronounced symptoms that begin to significantly impact daily functioning. This progressive phase represents a critical period where proactive intervention can substantially influence quality of life outcomes. 1. Advancing Motor Symptoms **The bilateral spread of symptoms marks a fundamental shift** in disease progression. <cite id="id_18">As the disease progresses, it eventually affects both sides. However, the symptoms may still be more severe on one side than on the other</cite>. This bilateral involvement signals that families must adapt their support strategies to address more comprehensive functional limitations. <cite id="id_22">Often within 3 to 7 years, you’ll see more changes. Early on, you might have a little trouble with something like buttoning a shirt. At this point, you may not be able to do it at all</cite>. **This progression from mild difficulty to complete inability** represents a pattern families will observe across multiple daily activities, requiring gradual assumption of caregiving responsibilities. 35

36. **Speech and swallowing changes become particularly noticeable** during this stage. <cite id="id_5">You may speak softly or quickly, slur, or hesitate before talking. Your speech may be flat or monotone, without typical speech patterns</cite>. These changes affect not only communication but also safety, as <cite id="id_22">trouble swallowing</cite> increases the risk of aspiration and nutritional deficiencies. **Gait disturbances and freezing episodes** represent some of the most challenging aspects of progressive Parkinson’s disease. <cite id="id_18">People with Parkinson’s disease often develop a parkinsonian gait that includes a tendency to lean forward; take small, quick steps; and reduce swinging their arms. They also may have trouble initiating or continuing movement</cite>. The emergence of <cite id="id_22">freezing when you first start to walk or change direction, as if your feet are glued to the ground</cite> creates significant safety concerns and mobility limitations. 2. Emerging Complications **Medication effectiveness begins to fluctuate** during the progressive stage, creating new challenges for symptom management. <cite id="id_22">You might also find that the medicine you take starts to wear off between doses</cite>. This "wearing-off" phenomenon requires families to become more attentive to timing of medications and to recognize when symptoms return between doses. **Balance problems and fall risk increase substantially**. <cite id="id_6">At this stage, falls become more frequent due to balance and reflex problems</cite>, and <cite id="id_22">falls to be more likely</cite>. This increased fall risk necessitates environmental modifications and potentially the introduction of assistive devices. 3. Functional Impact Assessment **The distinction between Stages 2 and 3 centers on independence preservation**. <cite id="id_6">Due to the various issues this stage presents, daily tasks are much more difficult to do, but people can still live independently. A combination of medication and therapy can help to manage the symptoms mentioned above</cite>. Understanding this threshold helps families recognize when to increase support while still encouraging independence. D. Advanced Stage Warning Signs (Stages 4-5) The advanced stages of Parkinson’s disease present complex challenges that require comprehen- sive care planning and family coordination. These stages are characterized by severe functional limitations and the emergence of complications that significantly impact quality of life. 1. Severe Motor Impairment **Stage 4 marks the loss of independent living capability**. <cite id="id_6">The critical factor in separating people with stage 3 Parkinson’s and stage 4 is independence. Motor skills and deep brain stimulation are heavily impacted at stage 4 and as such, movement symptoms affect a person’s ability to retain their independence</cite>. This transition requires families to make difficult decisions about living arrangements and care provision. <cite id="id_22">At this point, you likely: Are limited to bed or a wheelchair, Can’t live on your own, Have severe posture issues in your neck, back, and hips, Need help with daily tasks</cite>. **The comprehensive nature of these limitations** means that families must coordinate multiple aspects of care, from basic hygiene to mobility assistance. **Stage 5 represents the most severe manifestation** of motor symptoms. <cite id="id_6">Severe stiffness can make it difficult, if not impossible, for a person to stand or walk. This is due to stiffness causing the legs to essentially freeze when the patient attempts to stand</cite>. This "freezing" phenomenon at the advanced stage differs qualitatively from earlier freezing episodes, representing a more permanent limitation rather than intermittent difficulty. 2. Complex Non-Motor Complications **Cognitive decline becomes a prominent feature** in advanced stages. <cite id="id_18">Over time, as the disease progresses, some people may develop dementia and be diagnosed with 36

37. Parkinson’s dementia, a type of Lewy body dementia. People with Parkinson’s dementia may have severe memory and thinking problems that affect daily living</cite>. This cognitive impairment adds another layer of complexity to caregiving responsibilities. **Psychiatric symptoms emerge as significant challenges**. <cite id="id_6">Furthermore, up to 50% of Parkinson’s patients in stages 4 and 5 experience confusion, hallucinations, and delusions. For clarity, hallucinations are seeing things that aren’t really there, and delusions are when a person believes something despite evidence to the contrary</cite>. These symptoms can be particularly distressing for families and may require specialized psychiatric intervention. **Swallowing and nutritional complications** become life-threatening concerns. <cite id="id_5">Late-stage Parkinson’s disease affects the muscles in the mouth. This causes trouble swallowing and chewing, which can lead to not getting enough nutrients in your diet. If food or saliva collects in the mouth, it can cause choking or drooling</cite>. These complications require careful monitoring and may necessitate alternative feeding methods. <table> <caption>Progressive Warning Signs by Disease Stage</caption> <thead> <tr> <th>Stage</th> <th>Motor Warning Signs</th> <th>Non-Motor Warning Signs</th> <th>Functional Impact</th> <th>Family Action Required</th> </tr> </thead> <tbody> <tr> <td>Early (1-2)</td> <td>Unilateral tremor, mild rigidity, reduced arm swing, micrographia</td> <td>Sleep disturbances, constipation, loss of smell, mood changes</td> <td>Minimal daily life disruption</td> <td>Medical evaluation, symptom monitoring</td> </tr> <tr> <td>Progressive (2-3)</td> <td>Bilateral symptoms, gait freezing, speech changes, balance problems</td> <td>Medication wearing-off, cognitive changes, swallowing difficulty</td> <td>Increased difficulty with daily tasks, still independent</td> <td>Environmental modifications, fall prevention, medication timing</td> </tr> <tr> <td>Advanced (4-5)</td> <td>Severe rigidity, wheelchair/bed bound, postural deformities</td> <td>Dementia, hallucinations, severe dysphagia, autonomic failure</td> <td>Complete dependency for daily activities</td> <td>24-hour care, safety monitoring, end-of-life planning</td> </tr> </tbody> </table> **The temporal progression of non-motor symptoms** follows a predictable pattern that families should understand. <cite id="id_22">What may show up early. You may have these issues years before any classic motor symptoms like tremors: Constipation, Depression, Loss of smell, Low blood pressure when you stand up, Pain, Sleep issues</cite>. Conversely, <cite id="id_22">What may show up later. Dementia and psychosis are two serious mental health issues that usually take a while to show up</cite>. This progression pattern reveals that **non-motor symptoms often become more disabling than motor symptoms** in advanced stages. <cite id="id_22">Almost everyone with Parkinson’s gets at least one of these. When severe, they’re more likely than motor issues to lead to a disability or make you move into a nursing home</cite>. This insight helps families understand why comprehensive care planning must address both motor and non-motor aspects of the disease. 37

38. 3. End-Stage Considerations **The advanced stage requires families to confront difficult realities** about prognosis and care goals. <cite id="id_22">After about 10 years, most people will have at least one major issue, like dementia or a physical disability</cite>. Understanding this timeline helps families engage in advance care planning while the patient can still participate in decision-making. **The variability in progression** means that some individuals may never reach the most severe stages, while others progress more rapidly. <cite id="id_22">Some people never reach this stage</cite> of advanced disability, emphasizing the importance of individualized care planning rather than assuming inevitable decline. </write> Step 7 <think> Now I need to move on to the second major section about critical warning signs requiring immediate medical intervention. This is crucial information for families to know when to seek emergency care. Let me retrieve information about emergency situations and warning signs. </think> <tool_call> {"name": "retrieve", "arguments": {"url_id": ["id_24", "id_25", "id_26", "id_28", "id_29", "id_31", "id_32", "id_33", "id_34", "id_35", "id_36"], "goal": "Gather comprehensive information about emergency warning signs and situations that require immediate medical intervention for Parkinson’s patients"}} </tool_call> <tool_response> The page content for the previous section has been masked for saving the space. </tool_response> Step 8 ... <terminate> ✓ 38

39. References Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025. anthropic. Meet claude, 2025. URL https://www.anthropic.com/claude. Yushi Bai, Jiajie Zhang, Xin Lv, Linzhi Zheng, Siqi Zhu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longwriter: Unleashing 10,000+ word generation from long context llms. arXiv preprint arXiv:2408.07055, 2024. Yushi Bai, Jiajie Zhang, Xin Lv, Linzhi Zheng, Siqi Zhu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longwriter: Unleashing 10, 000+ word generation from long context llms. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. URL https://openreview.net/forum?id=kQ5s9Yh0WI. João Coelho, Jingjie Ning, Jingyuan He, Kangrui Mao, Abhijay Paladugu, Pranav Setlur, Jiahe Jin, Jamie Callan, João Magalhães, Bruno Martins, et al. Deepresearchgym: A free, transparent, and reproducible evaluation sandbox for deep research. arXiv preprint arXiv:2505.19253, 2025. Deep Consult. Deep consult. 2025. URL https://github.com/Su-Sea/ydc-deep-research-evals. Google DeepMind. Gemini 2.5, 2025. URL https://blog.google/technology/google-deepmind/gemi ni-model-thinking-updates-march-2025/. Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. Deepresearch bench: A comprehensive benchmark for deep research agents. arXiv preprint arXiv:2506.11763, 2025. Runnan Fang, Shihao Cai, Baixuan Li, Jialong Wu, Guangyu Li, Wenbiao Yin, Xinyu Wang, Xiaobin Wang, Liangcai Su, Zhen Zhang, Shibin Wu, Zhengwei Tao, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. Towards general agentic intelligence via environment scaling, 2025. google. Try deep research and our new experimental model in gemini, your ai assistant, 2025. URL https://blog.google/products/gemini/google-gemini-deep-research/. Rujun Han, Yanfei Chen, Zoey CuiZhu, Lesly Miculicich, Guan Sun, Yuanjun Bi, Weiming Wen, Hui Wan, Chunfeng Wen, Solène Maître, George Lee, Vishy Tirumalashetty, Emily Xue, Zizhao Zhang, Salem Haykal, Burak Gokturk, Tomas Pfister, and Chen-Yu Lee. Deep researcher with test-time diffusion. CoRR, abs/2507.16075, 2025. doi: 10.48550/ARXIV.2507.16075. URL https://doi.org/10.48550/arX iv.2507.16075. Juyong Jiang, Fan Wang, Jiasi Shen, Sungju Kim, and Sunghun Kim. A survey on large language models for code generation. arXiv preprint arXiv:2406.00515, 2024. LangChain, Inc. LangChain: Building applications with LLMs through composability, 2023. URL https://python.langchain.com/. Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Yida Zhao, Liwen Zhang, Litu Ou, Dingchu Zhang, Xixi Wu, Jialong Wu, Xinyu Wang, Zile Qiao, et al. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning, 2025a. Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, Weizhou Shen, Junkai Zhang, Dingchu Zhang, Xixi Wu, Yong Jiang, Ming Yan, Pengjun Xie, Fei Huang, and Jingren Zhou. Websailor: Navigating super-human reasoning for web agent. CoRR, abs/2507.02592, 2025b. doi: 10.48550/ARXIV.2507.02592. URL https: //doi.org/10.48550/arXiv.2507.02592. 39

40. Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. Long-context llms struggle with long in-context learning. arXiv preprint arXiv:2404.02060, 2024. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. DeepSeek-V3 technical report. arXiv preprint arXiv:2412.19437, 2024. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172, 2023. Siyi Liu, Kishaloy Halder, Zheng Qi, Wei Xiao, Nikolaos Pappas, Phu Mon Htut, Neha Anna John, Yassine Benajiba, and Dan Roth. Towards long context hallucination detection. arXiv preprint arXiv:2504.19457, 2025. Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. In The Twelfth International Conference on Learning Representations, 2023. OpenAI. Deep research system card, 2025a. URL https://cdn.openai.com/deep-research-system-c ard.pdf. OpenAI. Introducing openai o3 and o4-mini, 2025b. URL https://openai.com/index/introducing-o 3-and-o4-mini/. Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, Wenbiao Yin, Xinyu Wang, Zhen Zhang, Baixuan Li, Huifeng Yin, Kuan Li, Rui Min, Minpeng Liao, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. WebResearcher: Unleashing unbounded reasoning capability in Long-Horizon Agents, 2025. Qwen Team. QwQ-32B: Embracing the power of reinforcement learning, March 2025. URL https: //qwenlm.github.io/blog/qwq-32b/. Doubao Deep Research. Doubao deep research. 2025a. URL https://www.doubao.com/chat/. Gemini Research. Gemini research. 2025b. URL https://gemini.google/overview/deep-research/. GPT Research. Gpt research. 2025c. URL https://github.com/assafelovic/gpt-researcher. Kimi Deep Research. Kimi deep research. 2025d. URL https://www.kimi.com/. Open Deep Research. Open deep research. 2025e. URL https://github.com/langchain-ai/open_dee p_research. Corby Rosset, Ho-Lam Chung, Guanghui Qin, Ethan C. Chau, Zhuo Feng, Ahmed Awadallah, Jennifer Neville, and Nikhil Rao. Researchy questions: A dataset of multi-perspective, decompositional questions for LLM web agents. CoRR, abs/2402.17896, 2024. doi: 10.48550/ARXIV.2402.17896. URL https://doi.org/10.48550/arXiv.2402.17896. Aymeric Roucher, Albert Villanova del Moral, merve, Thomas Wolf, and Clémentine Fourrier. Open- source deepresearch – freeing our search agents. 2025. URL https://huggingface.co/blog/open-d eep-research. Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, Chenxi Wang, Maojia Song, Xinyu Wang, Kuan Li, Jialong Wu, Xuanzhong Chen, Zile Qiao, Zhongwang Zhang, Huifeng Yin, Shihao Cai, Runnan Fang, Zhengwei Tao, Wenbiao Yin, et al. Scaling agents via continual pre-training, 2025. 40

41. Zhengwei Tao, Jialong Wu, Wenbiao Yin, Junkai Zhang, Baixuan Li, Haiyang Shen, Kuan Li, Liwen Zhang, Xinyu Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. Webshaper: Agentically data synthesizing via information-seeking formalization. CoRR, abs/2507.15061, 2025. doi: 10.48550/A RXIV.2507.15061. URL https://doi.org/10.48550/arXiv.2507.15061. Kaiyang Wan, Honglin Mu, Rui Hao, Haoran Luo, Tianle Gu, and Xiuying Chen. A cognitive writing perspective for constrained long-form text generation. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.), Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025, pp. 9832–9844. Association for Computational Linguistics, 2025. URL https://aclanthology.org/2025.findings-acl.511/. Qianyue Wang, Jinwu Hu, Zhengping Li, Yufeng Wang, Daiyuan Li, Yu Hu, and Mingkui Tan. Generating long-form story using dynamic hierarchical outlining with memory-enhancement. In Luis Chiruzzo, Alan Ritter, and Lu Wang (eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2025 - Volume 1: Long Papers, Albuquerque, New Mexico, USA, April 29 - May 4, 2025, pp. 1352–1391. Association for Computational Linguistics, 2025. doi: 10.18653/V1/2025.NAACL-LONG.63. URL https: //doi.org/10.18653/v1/2025.naacl-long.63. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516, 2025. Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Yong Jiang, Pengjun Xie, et al. Webdancer: Towards autonomous information seeking agency. arXiv preprint arXiv:2505.22648, 2025a. Weiqi Wu, Xin Guan, Shen Huang, Yong Jiang, Pengjun Xie, Fei Huang, Jiuxin Cao, Hai Zhao, and Jingren Zhou. Masksearch: A universal pre-training framework to enhance agentic search capability. 2025b. URL https://arxiv.org/abs/2505.20285. Yuhao Wu, Ming Shan Hee, Zhiqiang Hu, and Roy Ka-Wei Lee. Longgenbench: Benchmarking long-form generation in long context llms. In The Thirteenth International Conference on Learning Representations, 2025c. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. Re3: Generating longer stories with recursive reprompting and revision. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pp. 4393–4479. Association for Computational Linguistics, 2022. doi: 10.18653/V1/2022.EMNLP-MAIN.296. URL https://doi.org/10.18653/v1/2022.emnlp-main.296. Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian. DOC: improving long story coherence with detailed outline control. In Anna Rogers, Jordan L. Boyd-Graber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pp. 3378–3465. Association for Computational Linguistics, 2023. doi: 10.18653/V1/2023.ACL-LONG.190. URL https://doi.org/10.18653/v1/2023.acl-long.190. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. Re- act: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. Haopeng Zhang, Philip S Yu, and Jiawei Zhang. A systematic survey of text summarization: From statistical methods to large language models. ACM Computing Surveys, 57(11):1–41, 2025. 41

42. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. In NeurIPS, 2023. 42