DeepWeaver: Bridging the Evidence Synthesis Gap in Open-Ended Question Answering

• The paper defines an evidence synthesis gap: noisy, fragmented, knowledge-dense evidence can obscure fine-grained details and leave claims underdeveloped when LLMs rely on internal reasoning alone.The authors argue that open-ended QA needs an explicit synthesis stage beyond retrieval.Source: §1 · p. 2
• DeepWeaver organizes evidence into Thought Block Chains (TBCs), using subordinate TBCs to inspect residual evidence and commit newly discovered information and claims to the main chain.Each thought block stores a claim, keywords, salient information, and supporting evidence fragments.Source: §1 · p. 2
• LoQA pairs 100 water-environment research questions with evidence from a Chinese expert knowledge base of 500 books and over 100M characters, testing comprehensive evidence use, domain-specific detail, and citations.Source: §1 · p. 2
• DeepWeaver improves content sufficiency, citation grounding, and detail preservation on LoQA, and achieves deeper insights and higher citation quality than advanced agents on DeepResearch Bench.Source: §1 · p. 2

2 Task Formulation

• The task tests whether QA systems can use a large, fragmented evidence pool to thoroughly answer open-ended research questions while remaining grounded in the evidence.The evidence may be scattered across sources and support multiple claims, so retrieval alone does not establish whether a system can fully exploit it.Source: §2.1 · p. 2, §2 · p. 2

2.1 Definition

• Answer quality is defined by content sufficiency, detail preservation, and citation grounding: covering major claims, retaining fine-grained information, and supporting claims with relevant evidence.Source: §2.1 · p. 2

2.2 The LoQA Benchmark

• LoQA draws on a Chinese water-environment knowledge base of 500 books and over 100M characters, chosen for dense technical knowledge, complex causal relations, and practical open-ended questions.Source: §2.2 · p. 3
• LoQA contains 100 research-style questions selected from 3,000 LLM-generated candidates using quality filters and retrieval similarity to favor questions with rich evidence.The filters assess retrieval comprehensiveness, concreteness, domain depth, and expression completeness.Source: §2.2 · p. 3
• Each question receives about 100 retrieved 1,024-token fragments plus 100 random fragments, creating a noisy context of about 205K tokens per question.Because gold answers for long-context, open-ended domain questions are difficult to produce and evaluate, LoQA assesses comprehensiveness along its three controlled dimensions.Source: §2.2 · p. 3
• For citation grounding, DeepSeek-V3.2 judges evidence relevance, and Citation Count (CC), Relevant Count (RC), and Relevant Ratio (RR) measure evidence use and avoidance of irrelevant citations.Source: §2.2 · p. 3
• Content sufficiency also uses Argument Sufficiency (AS) to check reference-derived arguments, while Detail Preservation (DP) tests recovery of cloze blanks targeting technical terms, concepts, and short phrases.Source: §2.2 · p. 3
• Content sufficiency uses five answers generated from relevant-evidence partitions as references, with Recall and Word Alignment (WA) measuring lexical overlap against them.Source: §2.2 · p. 3

3.1 Thought Block Chain

• Each TBC is an ordered sequence of blocks, each pairing a candidate claim with keywords, salient information, and supporting evidence fragments.This explicit claim-to-evidence mapping lets the TBC organize fine-grained claims while reducing context burden during generation.Source: §3.1 · pp. 3–4

3.2 Evidence Weaving

• In Draft, the LLM directly generates an answer from the question and evidence pool, then extracts an initial main TBC from the draft.The draft uses the LLM’s global view of the evidence pool to form an initial set of claims.Source: §3.2 · p. 4
• The Subordinate stage inspects residual evidence for omitted aspects, details, or alternative support, after excluding fragments directly mentioned or meeting the block-similarity coverage criterion.The criterion selects top-k fragments whose similarity to a claim plus salient information exceeds the average similarity of evidence in that block.Source: §3.2 · p. 4, M if it satisfies either: (1) it is… · p. 4
• In Commit, DeepWeaver merges overlapping claims and combines their keywords, salient information, and evidence, then discards irrelevant, redundant, or weakly supported claims.These operations integrate multidimensional explanations while preventing the TBC from over-expanding.Source: M if it satisfies either: (1) it is… · p. 4, S is then committed back to the main… · p. 4

DISCARD

• Across n rounds, DeepWeaver repeats subordinate and commit stages to refine woven claims and incorporate additional supporting evidence.Draft and subordinate stages randomly sample r evidence fragments from E and the residual set Rt, allowing inspection of the full pool across rounds without processing it all at once.Source: q, MERGE(T 0 · p. 5

3.3 Evidence-Grounded Answer Generation

• For each refined TBC block, DeepWeaver generates an answer section using the block metadata and its linked evidence subset as focused context.The final answer is generated from the refined TBC T n.Source: §3.3 · p. 5
• The final answer links the generated sections; block-wise generation reduces context pressure and improves citation grounding through smaller, relevant evidence subsets.Source: §3.3 · p. 5

4.1 Datasets

• LoQA pairs 100 practical water-environment research questions with long-context evidence from a 500-book knowledge base; DeepResearch Bench covers 100 PhD-level tasks across 22 fields.DeepResearch Bench evaluates answer quality through RACE criteria and citation accuracy and effective-citation counts through FACT.Source: §4.1 · p. 5

4.2 Baselines

• Comparisons span direct RAG with full, relevant, or sampled evidence; outline- and plan-based generation; divide-and-conquer and evidence-preprocessing methods; and advanced web agents.The DeepResearch Bench comparisons also include WebWeaver.Source: §4.2 · p. 5

4.4 Main Results

• With Qwen3-30B-A3B-Instruct-2507, DeepWeaver surpasses E-RAG by 15.5% in Argument Sufficiency, +14.7 Relevant Citations, 14.5% in Relevant Citation Ratio, and 16.6% in Detail Preservation.The LoQA results show improvements across content sufficiency, citation grounding, and detail preservation.Source: Table 1 · p. 6, §4.4 · p. 6
• E-RAG does not outperform E120-RAG despite receiving more evidence, and even oracle ER-RAG lags behind DeepWeaver, indicating that filtering noise alone is insufficient.The authors suggest that more retrieved chunks can increase context burden and cause useful information to be overlooked.Source: §4.4 · p. 6
• SoT, PaS, CoA, and LongRefiner show limitations including lost fine-grained detail, unreliable planning, insufficient claims, or missing details, supporting the need for claim-level evidence synthesis.CoA and LongRefiner improve citation use in some cases by reducing context burden or refining evidence.Source: §4.4 · p. 6

4.5 Ablation Study

• The ablation study examines TBC subordination and commitment; without subordination, the system directly expands the initial TBC.The supplied passage describes the ablation setup but does not report its results.Source: §4.5 · p. 6

M into the final answer. In w/o commit, the system

• Removing subordination hurts all three evidence-synthesis dimensions, while removing commitment yields comparable AS and DP but more blocks, indicating overlapping and redundant claims.The authors attribute the subordination result to incomplete initial TBCs that need refinement to recover overlooked evidence.Source: M into the final answer. In w/o… · p. 6, Table 2 · p. 6
• Two refinement rounds improve content sufficiency, citation grounding, and detail preservation over one, while a third yields no substantial gains; n = 2 is the best overall performance and cost–performance trade-off.The authors use n = 2 by default, stating that excessive refinement may add only marginally useful information.Source: Figure 3 · p. 7, M into the final answer. In w/o… · p. 7

4.6 Cross-Model Generalizability

• On DeepSeek-V4-Flash, DeepWeaver raises Recall from 24.2 to 31.6, AS from 75.4% to 86.5%, RC from 31.0 to 44.6, and DP from 60.0% to 67.4%.Gains are consistent across backbone models; DeepWeaver with Qwen3-30B-A3B-Instruct-2507 also outperforms direct generation with the strong long-context DeepSeek-V3.2.Source: Table 3 · p. 7, §4.6 · p. 7

4.7 Extension to Web-Based Deep Research

• For web deep research, DeepWeaver uses WebWeaver’s collected pages and section titles, then weaves fine-grained evidence chunks into detailed claim-level structures rather than webpage-level outline titles.Retrieval is not the focus of DeepWeaver, so the comparison reuses WebWeaver’s planner outputs and divides each question into section-level writing tasks.Source: §4.7 · p. 7
• DeepWeaver avoids a summary model for webpage-level evidence extraction during collection, substantially reducing the planning-stage cost bottleneck.The paper points to §A.3 for the cost comparison.Source: §4.7 · p. 7
• DeepWeaver achieves the best overall RACE score and improves effective citation and citation accuracy over WebWeaver on FACT, though denser citations and deeper claims may affect readability.The authors note that the evaluation metric may favor outline-driven writing and leave readability optimization for future work.Source: §4.7 · p. 7, Table 4 · p. 8

Retrieval-Augmented and Attributed Generation

• Retrieval-augmented generation incorporates external knowledge through sparse term matching or dense vectors, with newer methods improving retrieval control, evidence organization, reasoning, and planning.Sparse retrieval is discussed by Robertson and Zaragoza (2009), Wang et al. (2023b), and Schick et al. (2023); dense retrieval by Lewis et al. (2021), Hofstätter et al. (2022), and Liu et al. (2025). Recent advances include work by Guu et al. (2020), Ram et al. (2023), Xu et al. (2024), Asai et al. (2024b), Han et al. (2024), Wang et al. (2024d), and Hu et al. (2025).Source: §5 · p. 8

Long-Context Evidence Integration

• Long-context methods refine retrieved evidence or divide it into segments, but evidence synthesis remains underexplored.Context refinement includes RECOMP, Chain-of-Note, and LongRefiner; divide-and-conquer methods split long contexts and aggregate intermediate outputs. These approaches reduce context burden through coarser evidence representations.Source: §5 · p. 8

Long-Form Grounded Writing and DeepResearch Agents

• Grounded writing and deep-research agents address long-form generation, attribution, and end-to-end QA through outlining, reflection, search interaction, evidence collection, and agent training.Survey-generation work includes Yan et al. (2025), Skarlinski et al. (2024), Shao et al. (2024), and Wu et al. (2025); citation-oriented work includes Huang et al. (2024) and Ye et al. (2024). Deep-research agents include Li et al. (2025a), Jin et al. (2025a), Li et al. (2025b), Wu et al. (2026), and Zheng et al. (2025).Source: §5 · p. 8

6 Conclusion

• DeepWeaver organizes noisy evidence into Thought Block Chains that connect fine-grained evidence fragments with evidence-grounded claims.Rather than treating retrieved evidence as a flat long-context input, it uses explicit evidence organization.Source: §6 · p. 8
• Through iterative evidence weaving, DeepWeaver identifies overlooked evidence, refines incomplete claims, and improves detail preservation and citation grounding.Source: §6 · p. 8
• Experiments on knowledge-base QA and web-based deep research show that DeepWeaver consistently improves answer quality across multiple LLM backbones.The conclusion also introduces LoQA, a high-density benchmark for evaluating evidence-grounded answer generation.Source: §6 · p. 8
• The findings suggest that open-ended QA requires explicit evidence synthesis between retrieval and generation, especially when evidence is noisy and fragmented.Source: §6 · p. 8

7 Limitations

• LoQA focuses on Chinese water-environment questions, so it does not cover all languages, domains, or document types.The authors describe the domain as providing dense technical evidence and complex practical problems.Source: §7 · p. 9
• DeepResearch Bench provides a complementary evaluation setting, while expanding LoQA’s metrics and assessment protocols remains future work.Source: §7 · p. 9
• DeepWeaver mainly improves evidence synthesis after evidence collection; the web evaluation reuses pages and section titles collected by WebWeaver’s planner.Source: §7 · p. 9
• Designing web-search agents better aligned with DeepWeaver remains an open problem.Source: §7 · p. 9
• DeepWeaver currently focuses on text-only evidence, leaving evidence weaving for figures, charts, images, and structured data as future work.Source: §7 · p. 9
• Even with improved evidence synthesis, model-generated content still requires careful verification.Source: §7 · p. 9

A.1.1 Data Statistics

• 100 LoQA questions each pair with an average of 206K evidence tokens and 91.79 relevant chunks, supporting fine-grained evaluation over noisy evidence.The benchmark includes 2,045 atomic scoring points and 4,779 cloze blanks, averaging 20.45 scoring points and 47.79 blanks per question; reference answers average 9,858.64 tokens per question.Source: §A.1.1 · p. 12, Table 5 · p. 12

A.1.2 Construction Details

• LoQA draws on a water-environment corpus: 30,000 professional PDFs were filtered to 800 relevant books, with 500 forming the knowledge base and the remainder serving as question-generation seeds.PDF extraction produced approximately 360 MB of clean text.Source: §A.1.2 · p. 12
• Three LLMs generated 2,700 candidate questions from seed fragments; DeepSeek-V3.2 scored them on four dimensions, and 100 were retained using embedding similarity to retrieved chunks.Questions were required to be decontextualized, concise, generally answerable, and dependent on multisource domain knowledge.Source: §A.1.2 · p. 12
• Each question receives roughly 90–100 top-similarity fragments plus about 100 randomly sampled fragments, yielding 200 fragments and approximately 200K evidence tokens.DeepSeek-V3.2 also assesses fragment relevance to support citation-grounding evaluation.Source: §A.1.2 · pp. 12–13
• For each question, DeepSeek-V3.2 generates five references: one from each of four disjoint partitions of relevant evidence and one from the full relevant-evidence set.Partition-based references support local, fine-grained coverage under lower context burden, while the full-evidence reference provides global synthesis.Source: §A.1.2 · p. 13
• Atomic arguments from the references measure content sufficiency, while cloze blanks from the full-evidence reference target domain-specific details; together with relevance annotations, they assess three dimensions.The dimensions are citation grounding, content sufficiency, and detail preservation.Source: §A.1.2 · p. 13
• LoQA relies on localized extraction and verification tasks—not holistic subjective scoring—including relevance judgments, argument matching, and cloze recovery.The paper argues these constrained decisions fall within recent advanced LLMs’ reliable capability range, supporting DeepSeek-V3.2’s use for evaluation-signal construction.Source: §A.1.2 · p. 13

A.1.3 Comparison with Prior Benchmarks

• LoQA evaluates comprehensive, well-cited, detail-preserving answers to professional open-ended questions paired with fixed, extensive, noisy evidence pools.It is complementary to prior benchmarks, which emphasize explanatory QA, short-form answers, long-context understanding, or end-to-end web research.Source: §A.1.3 · pp. 13–14, Table 6 · p. 14

A.2 Details of Extending DeepWeaver to DeepResearch Bench

• DeepResearch Bench evaluates end-to-end research agents, so the extension tests whether DeepWeaver improves evidence-to-answer generation with the same search system and comparable inference cost.DeepWeaver has no search module, making this a controlled comparison of the generation stage.Source: §A.2 · p. 14
• WebWeaver’s planner searches pages, builds a hierarchical outline and evidence memory, but DeepWeaver receives only first-level outline titles and their raw web pages.This isolates evidence weaving from WebWeaver’s detailed outline and extracted evidence memory; WebWeaver is cited as (Li et al., 2025b).Source: §A.2 · pp. 14–15
• For each first-level outline node, the system rewrites it as a section task and routes critical or information-heavy sections to DeepWeaver, using direct generation otherwise.Raw webpages are truncated to 35K tokens and split into 1,024-token evidence fragments; rewriting and routing cost a few hundred tokens and are described as negligible.Source: §A.2 · p. 15
• Questions average 4.41 first-level nodes; 2.35 nodes (59.39%) route to DeepWeaver, and routed items average 98 chunks and 92.75K raw tokens.LoQA evidence is about 200K tokens per question; the passage presents this as evidence that DeepWeaver can handle realistic multi-webpage evidence-context lengths.Source: Table 7 · p. 15
• Lightweight polishing, an introduction, and a conclusion are added to match DeepResearch Bench’s answer format, while the core comparison concerns synthesized web evidence.The intended answer content is comprehensive, insightful, and well-cited.Source: §A.2 · p. 15

A.3 Cost Analysis

• At comparable cost, DeepWeaver passes up to 35K original tokens per webpage versus WebWeaver’s 24K, lowering cost per observed token from 1.26×10−5 to 8.94×10−6.Subordinate inspection is the main added cost, but its total token use is comparable to Draft; Commit, Discard, Polish, Intro, and Outro use only a small fraction.Source: §A.3 · p. 15, Table 8 · p. 16, Figure 6 · p. 16
• DeepWeaver costs $0.313 versus WebWeaver’s $0.303 and takes 14.2 versus 18.5 minutes of local inference, while using 94K versus 307K output tokens.Input-token use is 3.16M versus 2.35M. The comparison excludes web search and page access and uses the same searcher.Source: §A.3 · p. 15, Table 8 · p. 16

A.4 Prompt Templates

• The templates represent each TBC logic block with an argument, 3–5 keywords, key information, and reference chunks, while the subordinate stage generates supplementary blocks from uncovered evidence.The subordinate prompt asks that each block highlight its supplementary value to existing logic; reference chunks may come from displayed citations or other highly relevant passages.Source: §A.4 · pp. 18–19
• Commit templates screen cross-group main–subordinate overlaps for merging, then organize covered blocks with unique references; discard templates flag off-topic, redundant, incorrect, or low-impact blocks.Merge candidates must be clear overlaps, and the merge output may pair only one main block with one subordinate block. Discard rationales are returned in JSON, or as an empty array when none qualify.Source: §A.4 · pp. 20–22
• The report-writing prompts require focused discussion, preservation of valid inline citations, reduced repetition, and clear subsection structure.Direct-generation prompts also require using supplied evidence chunks and citing them with the provided local IDs.Source: §A.4 · pp. 24–25

A.5 Case Study

• In the LoQA case, DeepWeaver reframes wastewater treatment as a system for recovering water, energy, and valuable materials from wastewater.It links initially separate claims about reuse, energy and material recovery, ecological evaluation, and economic feasibility through subordinate claims about circular resource use and multidimensional evaluation.Source: §A.5 · p. 28
• In the DeepResearch Bench case, DeepWeaver connects sports tutoring applications to signal-, feature-, and decision-level data fusion.Examples include fusing IMU and video data to build a 3D motion model for posture feedback and combining biomechanical and physiological signals for performance evaluation.Source: §A.5 · p. 28

B Ethics Statement

• The authors release only limited, approximately 1,000-token OCR-derived excerpts per book with annotations and metadata for non-commercial academic research, reproducibility, and evaluation.They state that the books were accessed through licensed institutional collections, the excerpts are not a substitute for or means to reconstruct the originals, and disputed excerpts will be reviewed and removed upon copyright-holder request; users remain responsible for legal and license compliance.Source: B Ethics Statement · p. 28
• The authors state that the data derive from published books and contain no private user data, personal communications, or sensitive personal information.Source: B Ethics Statement · p. 28

C Dataset License and Terms of Use

• Annotations, metadata, and evaluation code use Apache-2.0; data use CC BY-NC 4.0, and copyrighted book excerpts remain the property of their respective rights holders.Source: C Dataset License and Terms of Use · p. 28

Case Study on LoQA

• Treatment goals shift from discharge compliance to high-quality reclaimed water for varied uses, enabled by source separation, differentiated treatment, cascading reuse, and deep treatment.The proposed pathway includes clean–wastewater separation, treatment by water quality, cascading use, and dual-membrane methods.Source: Case Study on LoQA · p. 31
• The technical goal is to move beyond pollutant removal toward closed-loop recovery of water, energy, and materials, including biogas from high-concentration organic wastewater, nutrients, and valuable industrial byproducts.Anaerobic digestion, biological nitrogen and phosphorus removal, and membrane separation support efficient water reuse, recovery of energy and resources, pollutant minimization, and ecological, low-carbon processes.Source: Case Study on LoQA · p. 32
• The proposed integrated resource-recovery process package combines pretreatment, anaerobic biogas production, nitrogen and phosphorus removal, membrane treatment, and concentrate treatment, with source sorting of black and grey water.The case study invokes decentralized treatment and recycling (DESAR) to make recovery more targeted.Source: Case Study on LoQA · p. 32
• Evaluation should cover water and material recovery, ecological impacts, economic feasibility, system robustness, and intelligent coordination rather than treatment efficiency alone.Illustrative measures include reuse and recovery rates, energy use and emissions, shock-load resilience and effluent stability, life-cycle cost, and automation.Source: Case Study on LoQA · p. 32
• Economic assessment combines life-cycle costs and cost-benefit analysis with scale-up and marketability; high treatment costs require innovation, product standards, subsidies, and green finance.Costs include investment, operation, depreciation, and are weighed against recovery revenue and water- and discharge-fee savings.Source: Case Study on LoQA · p. 33
• Ecological design pairs source pollution reduction and industrial resource exchange with low-energy natural treatment, while toxic-pollutant pretreatment helps protect reuse and recovered products.Examples include industrial ecology parks, constructed wetlands, and stable ponds; enhanced hydrolysis or advanced oxidation can address persistent contaminants and secondary pollution.Source: Case Study on LoQA · pp. 34–35
• A comprehensive evaluation framework should move beyond traditional removal-rate and discharge-compliance measures, assessing resource recovery, process efficiency, ecological and environmental impacts, economic feasibility, system robustness, and coordinated intelligent management.Examples include water, energy, and nutrient recovery; energy and chemical use; carbon footprint; lifecycle cost; and shock-load resistance.Source: Case Study on LoQA · p. 36
• LoQA identifies lifecycle economics and ecological safety as implementation conditions, including cost-benefit analysis, policy and market support, and pretreatment to prevent toxic contaminants from entering recovery streams.The passages cite high costs for advanced oxidation and membrane processes, and call for targeted pretreatment to prevent toxic substances from accumulating or causing secondary pollution.Source: Case Study on LoQA · pp. 36–37
• Industrial water treatment’s goal is framed as shifting from discharge compliance toward efficient, tiered water reuse that produces high-quality reclaimed water for different uses.The described pathways include enterprise- or campus-scale separation, graded treatment, cascading use, and ultrafiltration–reverse-osmosis deep treatment.Source: Case Study on LoQA · pp. 37–38
• LoQA extends reuse into coordinated recovery of energy and valuable materials, using integrated treatment processes and source separation to target different wastewater streams.The described resource-recovery package combines pretreatment, anaerobic biogas production, nitrogen and phosphorus removal, membrane separation, and concentrate treatment; DESAR-style sorting separates blackwater and greywater.Source: Case Study on LoQA · pp. 36, 38

Case Study on DeepResearch Bench

• SITLS fuses wearable, physiological, and video data to assess movement and performance, provide immediate corrective feedback, and adapt training plans to individual learners.The case describes IMUs, pressure sensors, cameras, and physiological sensors, alongside a dynamic student model updated from performance and preferences.Source: Case Study on DeepResearch Bench · p. 41
• SITLS is described as detecting potential injury risks from long-term biomechanical patterns and real-time physiological signals, while supporting real-time technical correction in golf and gymnastics.Examples include flagging joint-loading asymmetry or fatigue spikes and using edge AI to correct golf swings or gymnastics technique on the spot.Source: Case Study on DeepResearch Bench · pp. 41–42
• For team sports, SITLS combines wearable and positioning data to assess coordination and tactics, giving coaches team insights and learners personalized feedback.The described outputs include temporal and spatial synchronization, collective movement patterns, workload, and injury risk, as well as progress feedback for learners.Source: Case Study on DeepResearch Bench · p. 42
• SITLS supports skill acquisition by breaking complex techniques into parts and providing targeted real-time feedback.The case study describes this as reducing the learning curve.Source: Case Study on DeepResearch Bench · p. 43
• By combining biomechanical, physiological, and movement metrics at the feature level, SITLS provides a multidimensional evaluation of performance, efficiency, and fatigue.The resulting representation supports real-time assessment and longitudinal training analysis.Source: Case Study on DeepResearch Bench · pp. 43, 48
• SITLS also tracks rehabilitation progress with biomechanical and physiological metrics and supports coach development through data-driven assessments and personalized recommendations.The case study links rehabilitation tracking to safer return-to-play decisions and coach recommendations to teaching behavior analysis.Source: Case Study on DeepResearch Bench · pp. 43–44
• SITLS generates personalized, adaptive training pathways from athlete profiles and performance data, with recommendations updated in real time.The described pathways account for individual responses and preferences.Source: Case Study on DeepResearch Bench · p. 45
• SITLS predicts injury risk by combining biomechanical stress and physiological fatigue signals to detect early warnings such as asymmetrical joint loading and declining HRV.The case study frames these signals as enabling intervention before clinical injury.Source: Case Study on DeepResearch Bench · pp. 44, 46
• In precision sports such as golf and gymnastics, SITLS detects movement faults and delivers immediate multimodal feedback for technical correction.The examples include swing mechanics and air awareness.Source: Case Study on DeepResearch Bench · pp. 44, 46
• For team sports, SITLS maps multi-player sensor and video data to quantify coordination, synchronization, team dynamics, and formation weaknesses.The case study says these patterns can be hidden from traditional observation.Source: Case Study on DeepResearch Bench · pp. 44, 46
• SITLS fuses raw IMU and synchronized video data at the signal level to model athletic movement and detect deviations for immediate feedback.The case study says this reduces noise and motion artifacts and supports a dynamic 3D movement model.Source: Case Study on DeepResearch Bench · pp. 47–48
• Across n refinement rounds, DeepWeaver samples residual evidence, builds subordinate TBCs, commits useful claims, and discards redundant, irrelevant, or weakly supported blocks.Residual evidence is retrieved by identifying evidence not covered by the current main TBC.Source: Case Study on DeepResearch Bench · p. 49

M , RETRIEVECOVERED(E, H, T (n)

• DeepWeaver links evidence to each final TBC block using top-k embedding retrieval, generates a section per block, and appends the sections sequentially into the answer.Source: M , RETRIEVECOVERED(E, H, T (n) · p. 49