Pith. sign in

REVIEW 2 major objections 1 minor 1 cited by

A hierarchical multi-agent system generates research reports that interleave text with factually aligned visuals and traceable sources.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-28 14:43 UTC pith:6EICXGLO

load-bearing objection TVIR adds a new benchmark of 100 tasks and a hierarchical agent for text-visual interleaved reports, which fills a clear gap but stays incremental. the 2 major comments →

arxiv 2606.02320 v1 pith:6EICXGLO submitted 2026-06-01 cs.CL

TVIR: Building Deep Research Agents Towards Text--Visual Interleaved Report Generation

Xinkai Ma , Zhiqi Bai , Dingling Zhang , Pei Liu , Yishuo Yuan , He Zhu , Jiakai Wang , Qianqian Xie
show 16 more authors
Yifan Zhao Xinlong Yang Hao Cong Zhiheng Yao Fengxia Xie Zihao Xu Haoran Xu Zhaohui Wang Minghao Liu Shirong Lin Yingshui Tan Yuchi Xu Wenbo Su Zhaoxiang Zhang Bo Zheng Jiaheng Liu
This is my paper
classification cs.CL
keywords deep research agentsmultimodal report generationtext-visual interleaved reportsmultimodal benchmarkmulti-agent frameworkvisual assessmentevidence-driven generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that existing deep research agents and their benchmarks treat visuals as optional add-ons rather than integrated evidence. It builds TVIR-Bench, a set of 100 tasks where images or charts must advance specific analytical goals, and TVIR-Agent, a multi-agent system that plans outlines, pulls images, creates sourced charts, and writes reports in sequence while keeping context. Dual evaluation measures both the text and the visual elements for reliability and alignment. When the new agent is tested against nine other systems, it records strong overall results, indicating that explicit multimodal architecture matters for producing reports that readers can trust.

Core claim

TVIR-Agent, a hierarchical multi-agent framework, constructs outlines, retrieves images, generates charts with traceable sources, and composes reports through context-aware sequential writing; when evaluated on TVIR-Bench using separate textual and visual assessment paths, it achieves strong overall performance across nine compared systems.

What carries the argument

TVIR-Agent, the hierarchical multi-agent framework that coordinates outline construction, image retrieval, sourced chart generation, and sequential report composition.

Load-bearing premise

Current benchmarks and agents stay mostly text-only and do not check whether added visuals are accurate and properly tied to the surrounding claims.

What would settle it

If TVIR-Agent scores no higher than text-only systems on the visual assessment metrics of TVIR-Bench, the claim that explicit multimodal design is required would not hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Visual elements in reports must be checked for factual reliability and alignment with the text analysis.
  • Benchmarks for deep research agents need tasks that require specific visual contributions rather than treating images as optional.
  • Report generation pipelines benefit from separate image-retrieval and chart-generation agents that maintain source traceability.
  • Sequential, context-aware writing improves coherence when text and visuals are produced together.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same separation of concerns could apply to domains such as scientific literature review where figures must support specific claims.
  • Without dedicated visual modules, general-purpose models may continue to insert misaligned or untraceable images even when prompted to use visuals.
  • Future evaluation suites could add a third path that measures whether a reader can verify a claim using only the cited visual.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The paper introduces TVIR (Text--Visual Interleaved Report Generation) to address the text-centric limitation of existing deep research agents. It presents TVIR-Bench, a benchmark of 100 expert-curated multimodal tasks where visual elements must serve specific analytical sub-goals, and TVIR-Agent, a hierarchical multi-agent framework for building outlines, retrieving images, generating traceable charts, and composing reports via context-aware sequential writing. A dual-path evaluation (Textual Assessment + Visual Assessment) is proposed, and experiments on nine systems are reported to show that TVIR-Agent achieves strong overall performance, underscoring the value of explicit multimodal design for evidence-driven reports.

Significance. If the results and evaluation hold, the work is significant because it shifts the focus of deep research agents from purely textual long-form generation to interleaved multimodal reports with verifiable visuals. The new benchmark and dual-path evaluation framework could become a standard for testing factual alignment between text and visuals, filling a clear gap in current benchmarks and systems.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'TVIR-Agent achieves strong overall performance' and that experiments 'underscore the importance of explicit multimodal design' cannot be assessed because the abstract supplies no quantitative metrics, ablation results, task examples, or details on the nine compared systems. This is load-bearing for the paper's contribution claim.
  2. [Abstract] TVIR-Bench description: it is stated that the 100 tasks 'require visual elements to serve specific analytical sub-goals,' but without concrete task examples or curation criteria in the visible text, it is impossible to verify whether the benchmark genuinely tests factual reliability and alignment rather than decorative visuals.
minor comments (1)
  1. [Abstract] The dual-path evaluation is introduced but its concrete implementation (e.g., how Visual Assessment scores factual reliability or alignment) is not described in the provided material.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on the abstract. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'TVIR-Agent achieves strong overall performance' and that experiments 'underscore the importance of explicit multimodal design' cannot be assessed because the abstract supplies no quantitative metrics, ablation results, task examples, or details on the nine compared systems. This is load-bearing for the paper's contribution claim.

    Authors: We acknowledge that the abstract is high-level and omits specific metrics and details. The full paper reports quantitative results, including overall scores for TVIR-Agent versus the nine systems, in Section 4 along with ablations. To strengthen the abstract, we will revise it to include key performance highlights and a brief description of the evaluated systems. revision: yes

  2. Referee: [Abstract] TVIR-Bench description: it is stated that the 100 tasks 'require visual elements to serve specific analytical sub-goals,' but without concrete task examples or curation criteria in the visible text, it is impossible to verify whether the benchmark genuinely tests factual reliability and alignment rather than decorative visuals.

    Authors: The abstract summarizes the benchmark at a high level. Section 3.1 of the manuscript provides concrete task examples and curation criteria showing how visuals address analytical sub-goals (e.g., generating sourced charts for data verification). We will revise the abstract to include a short illustrative example of an analytical visual requirement. revision: yes

Circularity Check

0 steps flagged

No significant circularity; new artifacts evaluated empirically

full rationale

The paper introduces TVIR-Bench (100 expert-curated tasks) and TVIR-Agent (hierarchical multi-agent framework) as new constructs, then reports empirical performance across nine systems using a dual-path evaluation. No equations, fitted parameters, predictions derived from inputs, or self-citation chains appear in the abstract or described structure. The central claim concerns the utility of explicit multimodal design on a newly created benchmark, which is self-contained and externally falsifiable via the released artifacts rather than reducing to a definitional loop or renamed fit. This matches the default expectation of non-circularity for papers focused on new systems and benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no details on free parameters, axioms, or invented entities; ledger is empty by necessity.

pith-pipeline@v0.9.1-grok · 5778 in / 984 out tokens · 20053 ms · 2026-06-28T14:43:03.928597+00:00 · methodology

0 comments
read the original abstract

Deep Research Agents have shown strong capability in multi-step information retrieval, reasoning, and long-form report generation, but existing benchmarks and systems remain predominantly text-centric, with limited evaluation of whether visual elements are factually reliable and well aligned with the surrounding analysis. To address this gap, we introduce TVIR (Text--Visual Interleaved Report Generation), which includes TVIR-Bench, a benchmark of 100 expert-curated multimodal deep research tasks that require visual elements to serve specific analytical sub-goals, and TVIR-Agent, a hierarchical multi-agent framework that serves as a strong baseline for constructing outlines, retrieving images, generating charts with traceable sources, and composing reports through context-aware sequential writing. We further develop a dual-path evaluation framework that combines Textual Assessment and Visual Assessment. Experiments across nine deep research systems show that TVIR-Agent achieves strong overall performance, underscoring the importance of explicit multimodal design and evaluation for evidence-driven report generation.

Figures

Figures reproduced from arXiv: 2606.02320 by Bo Zheng, Dingling Zhang, Fengxia Xie, Hao Cong, Haoran Xu, He Zhu, Jiaheng Liu, Jiakai Wang, Minghao Liu, Pei Liu, Qianqian Xie, Shirong Lin, Wenbo Su, Xinkai Ma, Xinlong Yang, Yifan Zhao, Yingshui Tan, Yishuo Yuan, Yuchi Xu, Zhaohui Wang, Zhaoxiang Zhang, Zhiheng Yao, Zhiqi Bai, Zihao Xu.

Figure 1
Figure 1. Figure 1: Comparison of representative deep research bench [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Domain taxonomy of TVIR-BENCH. Task Design and Complexity Task construction is guided by five core design principles: role￾driven, demand-oriented, deep research, frontier￾focused, and multimodal integration. These prin￾ciples ensure that tasks are grounded in realistic user needs, require substantive analytical synthe￾sis rather than simple information retrieval, and incorporate explicit multimodal elemen… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of TVIR-BENCH, including data construction pipeline and evaluation framework. 3.3 Dataset Statistics TVIR-BENCH comprises 100 high-quality multimodal deep research tasks, including 50 Chinese tasks and 50 English tasks. They span 10 major domains and are proportionally balanced across the three predefined complexity levels. The sub-questions within these tasks cover eight high-level functional typ… view at source ↗
Figure 4
Figure 4. Figure 4: Overall architecture of TVIR-AGENT for text–visual interleaved report generation. are assessed at the report level. WQ evaluates coherence and organization, clarity and readability, conciseness, and stylistic consistency. ADB assesses whether the report goes beyond surface-level description through explanatory reasoning, sustained analysis, critical evaluation, forward-looking insight, and broad thematic c… view at source ↗
Figure 5
Figure 5. Figure 5: IA and ADB performance across task complexity [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Tool usage distribution of TVIR-AGENT variants across major components. contrast, TVIR-Agent (Claude-4.5-Sonnet) adopts a more balanced tool usage profile, maintaining a relatively high AEC of 86.14 while achieving the highest chart fulfillment rate of 94.61%. Overall, these results suggest that the relative strengths of TVIR-AGENT variants depend not only on backbone model capability, but also on how tool… view at source ↗
Figure 7
Figure 7. Figure 7: Structural error distributions across evaluated [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Average Effective Citations (AEC) across evaluated deep research systems. [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗
Figure 3
Figure 3. Figure 3: . . ”), and it meaningfully discusses key elements of the figure, so the reader is [PITH_FULL_IMAGE:figures/full_fig_p055_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents

    cs.AI 2026-08 conditional novelty 6.0

    Fetch-then-Explore, which stores fetched pages in a per-question workspace and extracts evidence on demand with grep/read, beats visit-and-read and browsing baselines on BrowseComp across three LLM backbones.

Reference graph

Works this paper leans on

45 extracted references · 2 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Deepresearch bench ii: Diagnosing deep research agents via rubrics from expert report, 2026a

    URLhttps://openreview.net/pdf?id=hQ0K2Hhq7H. Ruizhe Li, Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. Deepresearch bench ii: Diagnosing deep research agents via rubrics from expert report, 2026b. URL https://arxiv. org/abs/2601.08536. Peizhou Huang, Zixuan Zhong, Zhongwei Wan, Donghao Zhou, Samiul Alam, Xin Wang, Zexin Li, Zhihao Do...

    work page arXiv 2026

  2. [2]

    MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

    URLhttps://arxiv.org/abs/2511.11793. Google. Gemini deep research. 2025a. URLhttps://gemini.google/overview/deep-research. xAI. Grok deepsearch. 2025. URLhttps://grokipedia.com/page/Grok_DeepSearch. Anthropic. Introducing claude sonnet 4.5. 2025. URL https://www.anthropic.com/news/claude-sonne t-4-5. Perplexity AI. Introducing perplexity deep research. 20...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Head of R&D at a biopharmaceutical company

    Role-Driven Each task should be centered on aclearly identified user with a concrete role or professional identity(e.g., policymaker, clinician, investment analyst, scientist, student), reflecting authentic research needs and decision contexts in a specific domain. • Each task must explicitly specify the user’s role (e.g., “Head of R&D at a biopharmaceuti...

  4. [4]

    What do you think?

    Demand-Oriented Each task should pose a clear and structured research need within a specific domain. Itmay include multiple related sub-questions, provided that these are operational and help the model organize its output in a way that supports subsequent evaluation. • A task may include multiple sub-questions, but they must remain within the same domain ...

  5. [5]

    • Tasks should go beyond fact listing and require the model to establish logical chains of reasoning

    Deep Research Tasks should require the model to go beyond superficial aggregation of facts, instead demanding multi-source synthesis, causal reasoning, evidence-based argumentation, and critical analysis, thereby demonstrating deeper understanding and integration. • Tasks should go beyond fact listing and require the model to establish logical chains of r...

  6. [6]

    • Topics shouldfocus on new technologies, new policies, emerging markets, or scientific breakthroughs that have appeared within the past 2–3 years

    Frontier-Focused Task topics should cover recent developments and key challenges across domains, encouraging the model to use up-to-date literature, data, and multimodal resources that reflect current frontier knowledge. • Topics shouldfocus on new technologies, new policies, emerging markets, or scientific breakthroughs that have appeared within the past...

    2023

  7. [7]

    plot,” “show,

    Multimodal Integration Tasks should explicitly require the integration and presentation of multimodal information, asking the model to combine text, images, charts, and other modalities during analysis in order to improve clarity, explanatory power, and decision support. • Multimodal elements may include: images directlyretrieved from the web(e.g., model ...

    2024

  8. [8]

    de minimis

    Low Complexity About 200 Chinese characters (or about 130 English words); 1–3 multimodal requirements.The task may simply be described in the form of an outline; bullet points are not mandatory. • As the lead researcher at a prominent AI lab specializing in hybrid intelligence systems, I need to compile a comprehensive report for guiding our 2026 R&D road...

    2026

  9. [9]

    Medium Complexity About 400 Chinese characters (or about 260 English words); 2–4 multimodal requirements.It is recommended to describe the task in bullet points (3–4 points), with each point containing 1–3 sub-questions. • As a senior policy advisor in the UK Treasury’s Tax Policy Division, I need to draft a compre- hensive report on the role of inheritan...

    2026

  10. [10]

    Skills-First

    High Complexity About 600 Chinese characters (or about 390 English words); 3–5 multimodal requirements.The taskmustbe described in bullet points (4–5 points), with each point containing 2–4 sub-questions. • As a Workforce Development Policy Lead, I am evaluating the integration of micro- credentials into the national higher education framework to address ...

    1963

  11. [11]

    [3]”, “[4,5]

    Scope • Only extract citation facts that appear in themain text bodyand haveexplicit citation markers(e.g. “[3]”, “[4,5]”). • Ignore: – Citations related only tofigures(e.g. in figure captions or source notes around figures). –Anyunmarkedcitations (mentions of prior work without explicit numbered markers)

  12. [12]

    Expand context if needed for clarity

    Fact • Must be complete, understandable sentences or paragraphs containing the citation markers. Expand context if needed for clarity. • When one or more citations are associated with a table: – If the citations apply to theentire table(e.g., in the table caption or source notes around the table), the fact must contain theentire table content. – If the ci...

  13. [13]

    fact": "Text fragment from the report body containing the citation marker(s)

    ref_idx • Use the citation numbers exactly as they appear in the text. • If a fact cites multiple references, include all cited numbers in ref_idxs. Output Format Respond with a JSON list. Each item: 36 { "fact": "Text fragment from the report body containing the citation marker(s)", "ref_idxs": ["1", "2"] } USER PROMPT Please extract all citation facts f...

  14. [14]

    Figure 1: [to be inserted]

    Explicit Inclusion of Multimodal Elements • Valid multimodal elements include embedded or clearly shownfigures, or aconcrete, non- empty file path or URLthat clearly points to such an element. • The following donotqualify as valid multimodal content: –Placeholders (e.g., “Figure 1: [to be inserted]”, “image here”). – Any references without an actual figur...

  15. [15]

    information not provided,

    Focus on Substantive Data Delivery • The evaluation must be based strictly on the actual, concrete content delivered in the report; scoring is only permitted when the report explicitly includes substantive content fulfilling the required information. • The following donotqualify as satisfying an item: – Methodology-only text (plans, approaches, or descrip...

  16. [16]

    • Paragraph-level coherence: each paragraph has a clear main idea; sentences within a para- graph follow a logical progression

    Coherence & Organization Does the research_report have a logical, coherent flow at the document and paragraph level, with well-structured sections and smooth transitions? Focus: • Overall structure: clear sections and a reasonable order. • Paragraph-level coherence: each paragraph has a clear main idea; sentences within a para- graph follow a logical prog...

  17. [17]

    • Avoidance of overly convoluted, ambiguous, or unnecessarily complex sentences

    Clarity & Readability Is the research_report written in clear, readable language, avoiding unnecessary complexity and ambiguity? Focus: • Sentence-level clarity: proper grammar, clear syntax, and appropriate vocabulary. • Avoidance of overly convoluted, ambiguous, or unnecessarily complex sentences. • Technical terms and specialized concepts are used appr...

  18. [18]

    throat-clearing

    Conciseness & Redundancy Does the research_report avoid unnecessary repetition and filler, expressing ideas as succinctly as is reasonable for the task defined by the query? Focus: • Presence of repeated arguments, sentences, or paragraphs without added value. • Overuse of generic phrases, boilerplate, or “throat-clearing” that does not advance the conten...

  19. [19]

    ". Output Format Respond with a JSON object. Do not add markdown code blocks (like “‘json) around the output. {

    Stylistic & Referencing Consistency Is the writing style internally consistent in terms of formatting, including its use of tone, terminology, citations, references, and figures? Focus: • Tone: the level of formality is stable; the text does not jump needlessly between very informal and highly formal styles. • Terminology: key terms, labels, and names for...

  20. [20]

    what happened

    Causal Explanatory Reasoning To what extent does the report go beyond description and provide clear, causal or mechanism-oriented explana- tions for its main claims? Focus: 42 • Does the report move beyond simply stating “what happened” to explain “why” and “how it happens”? • Are there clear causal or mechanism links (e.g., linking causes, conditions, an...

  21. [21]

    Depth is concentrated on the most important points; added length is substantively analytical with minimal filler or repetition

    Analytical Depth & Development To what extent does the report develop its key points into sustained analysis rather than brief, list-like statements? Focus: • Are key claims unpacked with multi-step reasoning rather than asserted in a single step? • Are the most important points developed into coherent, connected paragraphs instead of fragmented one-line ...

  22. [22]

    It discusses important limitations of data and methods and thoughtfully considers how these constrain the conclusions

    Critical Evaluation & Assumptions To what extent does the report critically examine its assumptions, limitations, and alternative viewpoints instead of treating its own perspective as unquestioned? Focus: • Does the report identify and discuss key assumptions underlying its analysis, models, or narratives? 43 • Does it acknowledge limitations in data, met...

  23. [23]

    what to do or pay attention to next

    Actionable Forward-Looking Insight Does the report translate its analysis into meaningful, forward-looking implications or recommendations that help the reader understand what to do or watch for next? Focus: • Does the report derive concrete implications for action, decision-making, strategy, policy, or further work from its analysis? • Are recommendation...

  24. [24]

    ". Output Format Respond with a JSON object. Do not add markdown code blocks (like “‘json) around the output. {

    Thematic Breadth & Coverage How comprehensively does the report cover the key aspects of its topic, showing a sufficiently broad and well-chosen thematic scope without obvious blind spots? Focus: 44 • Does the report address the main dimensions that are clearly relevant to the core topic, rather than focusing narrowly on a single aspect? • Does it reasona...

  25. [25]

    increasing

    Factual contradictions Examples include, but are not limited to: • The same quantity (e.g., sample size, number of companies, growth rate, percentage, etc.) is given conflicting values with no explanation (e.g., no mention of different subsamples, filtering, or time points). • The same time period, region/market, or population/scope is described different...

  26. [26]

    unlikely to succeed,

    Logical contradictions Examples include, but are not limited to: 46 • The report states an explicit premise, assumption, or applicability condition, and later reasoning or conclusions clearly violate that premise under the same context. • The report gives a certain judgment/evaluation of a model, strategy, or view in one place and later uses an opposite c...

  27. [27]

    Multimodal Layout & Quantity Appropriateness Are multimodal elements placed, distributed, and used in an appropriate amount across the report’s stages, supporting a smooth reading flow without obvious gaps or overload? Focus: • Figures are placed near the sections where they are discussed, not arbitrarily clustered or detached. • The overall visual distri...

  28. [28]

    data_heavy

    Multimodal Variety & Richness Does the report demonstrate appropriate variety and richness in its multimodal elements based on the report’s topic and nature? Focus: •Source-type variety: – Retrieved external images: e.g., architecture diagrams, system schematics, flowcharts, sequence diagrams, environment photos, illustrations, etc. – Code-generated chart...

  29. [29]

    • Correct identification of subject, variables, and visible relationships

    Visual Accuracy Does the caption correctly identify and describe what is actually shown? Focus: • Alignment between caption and figure. • Correct identification of subject, variables, and visible relationships. Scoring: • 9-10 – Perfect match: The caption correctly identifies what the figure is (e.g., chart/dia- gram/photo), names the key entities/compone...

  30. [30]

    • Provide only theminimal guidance needed to read this figure.Do NOT requirethe caption to restate or list details already visible in the figure

    Minimum Necessary Information Using only the figure and this caption, can a reader correctly interpret what is shown? Does the caption provide theminimum necessaryinformation given the figure’s complexity (no less, no more)? Focus: • If a reader can understand the figure’s main point usingthe figure itself + this caption (without the main text), the capti...

  31. [31]

    ". Output Format Respond with a JSON object. Do not add markdown code blocks (like “‘json) around the output. 53 {

    Clarity & Readability Is the caption written in clear and accessible language that is easy to understand? Focus: • Clear sentence structure and word choice. • Avoids unnecessary jargon or defines technical terms when necessary. • Easy to parse and understand on first reading. Scoring: • 9-10 – Highly clear: The caption uses straightforward, direct languag...

  32. [32]

    The place- holder <figure> indicates where the figure appears in the text

    Context: The paragraph(s) in the report that reference or surround this figure. The place- holder <figure> indicates where the figure appears in the text. Goal Judge therelationship between the Figure and the Context. Scoring Dimensions (Each 1–10 points)

  33. [33]

    • Whether the figure depicts the same topic, variables, phenomena, or cases mentioned in the surrounding text

    Contextual Relevance How well does the figure’s content match the specific topic and claims in the context? Focus: •Semantic alignmentbetween what the figure shows and what the context discusses. • Whether the figure depicts the same topic, variables, phenomena, or cases mentioned in the surrounding text. Scoring: • 9-10 – Highly Relevant: The figure dire...

  34. [34]

    As shown in Figure 3

    Narrative Coherence How smoothly is the figure integrated into the flow of the text? Focus: 54 • Whether the contextexplicitly or clearly implicitlypoints the reader to this figure. • How natural the transition is between text and the figure, and whether the discussion around it is sufficient. Scoring: • 9-10 – Seamless Integration: The text explicitly re...

  35. [35]

    ". Output Format Respond with a JSON object. Do not add markdown code blocks (like “‘json) around the output. {

    Visual Information Value Does the figure provide visual information that text alone cannot effectively convey? Focus: • Whether the figure offersunique visual valuethat would be difficult, inefficient, or impossi- ble to communicate through text alone. • Consider how much understanding would be lost if the figure were removed? Would readers struggle to gr...

  36. [36]

    ‘json) around the output. {

    References: A JSON list of cited source snippets associated with the chart. Treat this list as the only permissible evidence; do not assume access to external links or any information not contained in the provided snippets. Goal Decide whether the chart isconsistentwith the provided sources. Out Of Scope • Pure visual aesthetics issues (layout, fonts, col...

  37. [37]

    At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report

    The report must truly achieve rich illustration with text and images. At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report. Visualization forms include but are not limited to charts generated using tools like Matplotlib, such as bar charts, line charts, heatmaps, radar charts, and related f...

  38. [41]

    The report must be written inMarkdownand follow standard Markdown syntax 57 Prompt for Grok-4.1-Thinking DeepSearch You are responsible for generating well-structured, thoroughly argued, and richly illustrated multi- modal deep research reports that meet the following requirements:

  39. [42]

    At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report

    The report must truly achieve rich illustration with text and images. At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report. Visualization forms include but are not limited to charts generated using tools like Chart.js, such as bar charts, line charts, heatmaps, radar charts, and related for...

  40. [46]

    The report must be written inMarkdownand follow standard Markdown syntax Prompt for Claude-4.5-Sonnet w/Search and Genspark Deep Research You are responsible for generating well-structured, thoroughly argued, and richly illustrated multi- modal deep research reports that meet the following requirements:

  41. [47]

    At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report

    The report must truly achieve rich illustration with text and images. At key argumentation points, you shouldproactively and adequately insert visual elementsand embed them into the report. Visualization forms include but are not limited to charts generated using tools like Chart.js, such as line charts, heatmaps, radar charts, and related formats, as wel...

  42. [48]

    as shown in Figure 1

    Allimages and chartsin the report must have clear and uniformly formatted numbering and titles, and must be explicitly referenced and interpreted in the main text, for example, “as shown in Figure 1”

  43. [49]

    [1]

    Allfacts, data, viewpoints, images, and other contentobtained through online searches must be marked in the main text withstandardized reference citations, for example, “. . . [1]”; charts generated using tools must also be annotated according to this requirement with their underlying data sources

  44. [50]

    At the end of the report, complete reference entries corresponding to the numbering in the main text must be listed according to a unified academic format, including accessibleURLs

  45. [51]

    The report must be written inHTMLand follow standard HTML syntax 58