Paper deep dive

Faithful Chart Generation for Multimodal Deep Research: Frame-Evidence Co-Adaptation

Yuxin Yue, Yingchen Zhang, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Xueqi Cheng

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 69

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%

Last extracted: 10/2/2026, 8:54:25 AM

Summary

The paper introduces Frame-Evidence Co-Adaptation (FECA), a framework for multimodal deep research that generates numerically faithful analytical charts. Unlike traditional 'plan-then-check' methods that often result in unsupported visualizations, FECA iteratively adapts visual frames based on retrieved quantitative evidence. This bidirectional interaction ensures that every value in a generated chart is grounded in actual web evidence, significantly improving numerical fidelity while maintaining report quality.

Entities (6)

FECA · method · 98%Multimodal Deep Research · domain · 95%Numerical Fidelity · concept · 92%Data-Frame Theory · theory · 90%Visual Frame · component · 88%Plan-then-Check · baseline-method · 85%

Relation Signals (5)

FECA → improves → Numerical Fidelity

confidence 95% · Experiments on 100 real-world research topics show that FECA substantially improves numerical fidelity while preserving report quality and chart utility.

FECA → inspiredby → Data-Frame Theory

confidence 90% · Inspired by the bidirectional sensemaking process in Data-Frame Theory, FECA models chart generation as an iterative interaction between visual frames and retrieved evidence.

FECA → replaces → Plan-then-Check

confidence 88% · FECA shifts chart generation from fixed-plan verification to adaptive evidence-grounded visual reasoning.

Retrieved Evidence → determines → Visual Frame Adaptation

confidence 85% · retrieved evidence determines whether the frame should be accepted, revised, or dropped before rendering.

Visual Frame → guides → Evidence Acquisition

confidence 85% · Each visual frame is adaptive: the frame guides evidence acquisition, while retrieved evidence determines whether the frame should be accepted, revised, or dropped before rendering.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Analytical charts in multimodal deep research encode quantitative claims, requiring every visualized value to be faithfully grounded in supporting evidence. Unlike retrieved images that mainly provide contextual information, charts require numerical fidelity: visualized values should not only match retrieved evidence quantitatively but also preserve its original meaning and scope. However, achieving such fidelity remains challenging because current systems usually construct visualization plans before knowing what quantitative evidence can actually be retrieved from the web. As a result, predefined plans may require entities, temporal ranges, or comparison dimensions that the retrieved evidence only partially supports. Existing approaches mainly address this issue through post-hoc verification after chart plans are fixed, enabling unsupported values to be identified but leaving the underlying visual frames unchanged. To address this challenge, we propose Frame-Evidence Co-Adaptation (FECA), an evidence-adaptive visual planning framework for multimodal deep research. Inspired by the bidirectional sensemaking process in Data-Frame Theory, FECA models chart generation as an iterative interaction between visual frames and retrieved evidence. Each visual frame is adaptive: the frame guides evidence acquisition, while retrieved evidence determines whether the frame should be accepted, revised, or dropped before rendering. By coupling visualization planning with evidence availability, FECA shifts chart generation from fixed-plan verification to adaptive evidence-grounded visual reasoning. Experiments on 100 real-world research topics show that FECA substantially improves numerical fidelity while preserving report quality and chart utility.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

65,397 characters extracted from source content.

Expand or collapse full text

Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation Yuxin Yue ∗ State Key Laboratory of AI Safety Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China yueyuxin25s@ict.ac.cn Yingchen Zhang ∗ State Key Laboratory of AI Safety Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China zhangyingchen23s@ict.ac.cn Ruqing Zhang † State Key Laboratory of AI Safety Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China zhangruqing@ict.ac.cn Jiafeng Guo State Key Laboratory of AI Safety Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China guojiafeng@ict.ac.cn Maarten de Rijke University of Amsterdam Amsterdam, The Netherlands m.deRijke@uva.nl Xueqi Cheng State Key Laboratory of AI Safety Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China cxq@ict.ac.cn Abstract Analytical charts in multimodal deep research encode quantitative claims, requiring every visualized value to be faithfully grounded in supporting evidence. Unlike retrieved images that mainly provide contextual information, charts require numerical fidelity: visualized values should not only match retrieved evidence quantitatively but also preserve its original meaning and scope. However, achieving such fidelity remains challenging because current systems usually construct visualization plans before knowing what quantitative evi- dence can actually be retrieved from the web. As a result, predefined plans may require entities, temporal ranges, or comparison dimen- sions that the retrieved evidence only partially supports. Existing approaches mainly address this issue through post-hoc verifica- tion after chart plans are fixed, enabling unsupported values to be identified but leaving the underlying visual frames unchanged. To address this challenge, we propose Frame–Evidence Co-Adaptation (FECA), an evidence-adaptive visual planning framework for mul- timodal deep research. Inspired by the bidirectional sensemaking process in Data–Frame Theory, FECA models chart generation as an iterative interaction between visual frames and retrieved evi- dence. Each visual frame is adaptive: the frame guides evidence acquisition, while retrieved evidence determines whether the frame should be accepted, revised, or dropped before rendering. By cou- pling visualization planning with evidence availability, FECA shifts chart generation from fixed-plan verification to adaptive evidence- grounded visual reasoning. Experiments on 100 real-world research topics show that FECA substantially improves numerical fidelity while preserving report quality and chart utility. ∗ Both authors contributed equally to this research. † Corresponding author. CCS Concepts • Information systems→Information retrieval; Web search- ing and information discovery;• Human-centered computing → Visualization. Keywords Multimodal deep research, Trustworthy visual evidence 1 Introduction Deep research systems [3,4,38,40] aim to automate end-to-end knowledge discovery by combining web search [12,34,37], rea- soning [17], and information synthesis [2] to produce long-form reports with supporting evidence. Early systems mainly focused on textual research reports, improving evidence retrieval, infor- mation organization, and citation-aware generation [18,19,30]. Recent multimodal deep research systems extend this capability by automatically generating analytical charts that summarize quan- titative information collected from the web [21,32,35]. Unlike retrieved images that provide illustrative context, charts function as structured visual evidence: each plotted value implicitly makes a quantitative claim that supports the analytical conclusions of the report. Thus, trustworthy multimodal deep research requires not only textual correctness but also reliable grounding of generated visual evidence. Numerical fidelity: A missing dimension of visual trustwor- thiness. Existing evaluations of deep research systems primarily focus on textual report quality and reliability, including informa- tion coverage, factuality, and citation correctness [7,16,31]. Recent multimodal extensions further consider visual quality and chart– text consistency [21,32,35]. However, charts introduce a unique demand of factuality beyond textual claims: every visualized value arXiv:2610.00374v1 [cs.HC] 30 Sep 2026 Yue and Zhang, et al. (a) Plan-then-Check TopicOutline Fixed visual frame Retrieve Candidate charts Check (b) FECA TopicOutline Accept Final report ✓ Numerically faithful charts ✓ High-quality multimodal report Limited Verification Failure Mode 1 Pass check Final report Unsupported values remain Strict Filtering Too few charts retained Failure Mode 2 Final report Revise Retrieve Analytical Intent Align Adaptive visual frame Evidence Figure 1: Unlike fixed visual frames that can only reject un- supported charts, our proposed FECA adapts visual frames ac- cording to retrieved evidence to produce evidence-grounded visualizations. corresponds to an implicit numerical assertion that should be trace- able to supporting evidence. We define this property as numerical fidelity: every value in a generated chart should be grounded in retrieved quantitative evi- dence while preserving the meaning and scope of its source. Despite its importance for trustworthy visual evidence [1,9], numerical fidelity remains largely unexplored in multimodal deep research. The limitation of fixed visualization planning. Current multi- modal deep research systems typically follow a plan–then–check paradigm for chart generation [21,32,35]. As illustrated in Fig- ure 1(a), they first instantiate visualization intent as a fixed visual frame, retrieve information required by the predefined frame, and subsequently check whether the retrieved evidence or generated output satisfies that frame. This paradigm implicitly assumes that a visual frame can be determined before evidence acquisition. However, web research is inherently uncertain: retrieved evidence may only partially cover the entities, temporal ranges, or comparison dimensions required by a predefined visual frame. When such a mismatch occurs, post-hoc verification can revise incorrect values using the available evidence or reject unsupported charts, but cannot adapt the visual frame to the evidence that is actually available. Diagnosing plan–evidence mismatches. To understand this lim- itation, we analyze a representative plan–then–check pipeline in Section 4.1. We find that 80.8% of its plotted values remain unsup- ported by the routed evidence after verification. Standard verifica- tion achieves only a limited reduction in the unsupported-value rate. We further construct a stricter verifier that retains only charts with fully supported values. Although verification improves numer- ical fidelity, it removes more than three quarters of the generated charts, substantially reducing the visual utility of the resulting re- ports. These results reveal a fundamental limitation of verification- based approaches: when a visual frame conflicts with available evidence, stricter checking can only filter invalid outputs rather than discover evidence-supported alternatives. This leads to a key research question: How can multimodal deep research systems gen- erate trustworthy visual evidence when retrieved evidence does not fully satisfy the initial visual frame? Our method: Frame–evidence co-adaptation. We argue that trustworthy chart generation requires a paradigm shift from ver- ifying fixed visual frames to adapting them according to the evi- dence actually available. As illustrated in Figure 1(b), we propose Frame–Evidence Co-Adaptation (FECA), in which the current vi- sual frame guides evidence acquisition, while retrieved evidence can in turn accept, revise, or drop the frame before chart generation. Inspired by the bidirectional sensemaking process in Data–Frame Theory [13], FECA realizes this interaction as an iterative frame– evidence co-adaptation process. Data–Frame Theory, a cognitive theory of sensemaking, describes how humans interpret the world under uncertain information through an iterative process: an initial frame guides evidence collection, while newly acquired evidence can confirm, refine, or revise the frame when it conflicts with prior assumptions. Following this principle, FECA treats each visual frame as adap- tive rather than fixed. Each visual frame specifies a concrete re- alization of the analytical intent in terms of its target measure, organizing dimensions, and data scope. The frame guides evidence acquisition, while retrieved evidence determines whether the frame should be accepted, revised, or dropped before chart generation. Through this bidirectional frame–evidence interaction, FECA en- ables visual evidence generation that remains aligned with both analytical intent and available web evidence. Specifically, FECA initializes a candidate visual frame for each analytical intent identified in the report outline. During evidence acquisition, it evaluates the compatibility between retrieved evi- dence and the current frame. If the evidence sufficiently supports the frame, the chart is directly generated. If the evidence reveals an alternative visualization that preserves the analytical intent, FECA adapts the frame accordingly. Otherwise, the unsupported frame is dropped. By coupling visualization planning with evidence avail- ability, FECA restricts chart generation to visual frames that have been aligned with retrieved quantitative evidence. Experimental results. Experiments on 100 real-world research topics show that FECA substantially improves numerical fidelity over existing systems while maintaining report and chart quality. Ablation and iteration analyses further confirm the importance of evidence-guided frame revision and characterize its effectiveness– cost trade-off. Overall, our findings show that trustworthy visual evidence generation requires iterative alignment between visual frames and available evidence, rather than verification alone. 2 Our Method In this section, we present Frame–Evidence Co-Adaptation (FECA), a framework for multimodal deep research where visual frames and retrieved web evidence are iteratively co-aligned to produce numerically faithful charts. 2.1 Task formulation and overview Given an open-ended research topic푞, a multimodal deep research system aims to generate a reportRconsisting of textual analysis T(R)and visual evidenceV(R). While existing systems primarily optimize the informativeness, coherence, and overall utility ofR, Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation #1 Visual Frame Initialization #2 Evidence-Driven Frame Adaptation Visual Frame Retrieve Retrieved Evidence #3 Grounded Visualization and Report Generation Report Recomposition and Generation Multimodal Report Evidence E₀: ITU Facts and Figures 2024 Group Subscriptions per 100 inhabitants World112 High-income economies 138 ALIGNED Matched measure • entities • year Every global region has seen a steep rise in mobile phone subscriptions. Topic Visualization Generation Mobile-cellular subscriptions per 100 people Selected developed economies – 2023 178.4 124.7 122.8 116.7 112.4 Japan Germany United France United Kingdom States Report-level source Report Outline Examine mobile- subscription levels relative to population across developed economies. Analytical Intent Frame F₀: Average mobile-cellular subscriptions per household in France, Germany, Japan, the UK, and the US, 2024. Frame F₁: Mobile-cellular subscriptions per 100 people in France, Germany, Japan, the UK, and the US, 2024. Frame F₂: Mobile-cellular subscriptions per 100 people in France, Germany, Japan, the UK, and the US, 2023. NOT ALIGNED Per household ≠ per 100 inhabitants No country-level values Country 2023 2024 France 116.688 — Germany 124.663 129.151 Japan 178.432 — the UK122.782 121.643 the US112.411 113.190 Evidence E₁: World Bank WDI Country 2023 France116.688 Germany124.663 Japan178.432 the UK122.782 the US112.411 Evidence E2: World Bank WDI Retrieve Retrieve Align Revise Revise Align Align NOT ALIGNED Incomplete 2024 coverage: France and Japan are missing Figure 2: Overview of FECA. Starting from analytical intents, FECA reformulates chart generation as an evidence-adaptive planning process, where visual frames and retrieved evidence iteratively co-evolve: evidence availability determines whether a frame is accepted, revised, or dropped before generating the final evidence-grounded chart. we introduce an additional requirement for trustworthy visual evi- dence: every generated chart should be grounded in quantitatively supported web evidence. Let푉 푖 ∈V(R)denote a generated chart and퐸 푖 its corresponding supporting evidence. Conceptually, our objective is to maximize the analytical utility of the report while requiring every generated chart to be faithful to its supporting evidence: R ∗ = arg max R 푄 analysis (R) s.th. Faithful(푉 푖 ,퐸 푖 )= 1, ∀푉 푖 ∈V(R), (1) where푄 analysis (R)measures the usefulness of the generated report for answering the research question.Faithful(푉 푖 ,퐸 푖 )is a binary indicator that equals 1 when every quantitative value displayed in 푉 푖 is traceable to 퐸 푖 and is consistent with the source evidence. To realize the above objective, FECA represents each candidate visualization by an analytical intent and an adaptive visual frame. For visualization푖, let퐼 푖 denote the analytical intent identified from the report outline, which specifies what the visualization is intended to investigate and remains fixed throughout the generation process. At iteration푡,퐹 (푡) 푖 denotes the corresponding visual frame, which specifies a concrete way to realize퐼 푖 as a chart in terms of its target measure, organizing dimensions, and data scope. Let 퐸 (푡) 푖 denote the quantitative evidence retrieved under the current frame. During frame–evidence co-adaptation,퐹 (푡) 푖 guides evidence acquisition, while퐸 (푡) 푖 determines whether the current frame can be accepted, should be revised, or must ultimately be dropped. Thus, the analytical intent remains fixed, whereas its visual frame can adapt to the quantitative evidence available on the web. As illustrated in Figure 2, FECA follows three stages. It first con- structs the report outline and initializes candidate visual frames from analytical intents; it then iteratively retrieves evidence and adapts each frame according to the evidence available; finally, only accepted frame–evidence pairs proceed to grounded chart genera- tion and report recomposition. 2.2 Visual frame initialization Before retrieving quantitative evidence for individual charts, a re- search system should first establish the report’s analytical objec- tives. Otherwise, available data sources may bias the analytical process by prematurely determining what can be analyzed and vi- sualized. FECA therefore begins with report-level textual research to derive an initial structure, analytical intents, and candidate visual frames that guide subsequent evidence acquisition. Report-level source collection. Given a research topic푞, we retrieve relevant webpages from the open web and extract their textual content to construct a report-level source setS. This source set provides broad contextual knowledge for organizing the report and identifying key aspects of the research topic. We assess whether Ssufficiently covers the major dimensions of푞and identify missing aspects that require additional investigation. For under-covered Yue and Zhang, et al. aspects, we perform an additional retrieval pass and incorporate the newly collected content intoS. Outline and frame initialization. Based on푞andS, we construct an initial multimodal report outlineO (0) . The outline organizes the report into sections and identifies analytical intents that can potentially benefit from quantitative visual evidence. For each ana- lytical intent 퐼 푖 , we initialize a visual frame 퐹 (0) 푖 , which specifies a preliminary chart design for realizing the intended analysis. During subsequent frame–evidence co-adaptation, the analytical intent퐼 푖 remains fixed, while the corresponding visual frame퐹 (0) 푖 can be revised or dropped according to the evidence retrieved for it. At this stage, FECA has established what each visualization is intended to analyze, but the feasibility of realizing these visual frames remains uncertain because the availability of supporting quantitative evidence on the open web is unknown. The follow- ing stage resolves this uncertainty by adapting each visual frame according to the evidence that can be effectively retrieved. 2.3 Evidence-driven frame adaptation FECA treats each visual frame as an adaptive analytical frame rather than a fixed visualization specification. Given an initial frame퐹 (0) 푖 and its corresponding analytical intent퐼 푖 , it performs a bounded Retrieve–Align–Revise loop to iteratively adapt the frame accord- ing to the available quantitative evidence. At each iteration, the current frame guides evidence acquisition, and the retrieved evi- dence determines whether the frame should be accepted, revised, or dropped. Throughout this process, the analytical intent퐼 푖 remains unchanged, while the visual frame 퐹 (푡) 푖 may be revised. The adaptation process has three outcomes: (i) Accept, when the current frame is sufficiently supported by retrieved evidence; (i) Revise, when the evidence reveals a mismatch but an alternative realization of the same intent may exist; or (i) Drop, when no evidence-compatible frame is found within the iteration budget. Frame-guided evidence retrieval. A visual frame specifies the quantitative evidence required to realize an analytical intent, but the existence and availability of such evidence on the open web remain unknown. At each iteration, Retrieve acquires evidence conditioned on the current frame 퐹 (푡) 푖 . Specifically, it searches for candidate data sources that match the target measure, organizing dimensions, and data scope specified by the frame. Previously discovered sources are ranked and reused when possible; otherwise, new queries are generated to address the remaining evidence gaps. The retrieved sources are then processed to extract quantitative evidence: 퐸 (푡) 푖 = Retrieve(퐹 (푡) 푖 ).(2) All quantitative values in퐸 (푡) 푖 are directly copied from their cited sources without modification. Frame–evidence alignment. The existence of a relevant data source does not necessarily imply that the current frame can be realized. Retrieved evidence may differ from the frame in terms of target measures, organizing dimensions, or analytical scope. Therefore, Align evaluates whether퐸 (푡) 푖 sufficiently supports퐹 (푡) 푖 : 푎 (푡) 푖 = Align(퐹 (푡) 푖 ,퐸 (푡) 푖 ), 푎 (푡) 푖 ∈ Accept, Revise.(3) We implement Align as a structured LLM-based evaluation over the frame and retrieved evidence. The evaluator independently exam- ines whether the target measure, organizing dimensions, and data scope required by퐹 (푡) 푖 are supported by퐸 (푡) 푖 . The frame is accepted only when all required aspects are supported without introducing unsupported quantitative assumptions. Otherwise, Align returns Revise together with the identified evidence gaps. Importantly, a revision decision does not indicate that the analyt- ical intent is infeasible. Instead, the retrieved evidence may support an alternative visual realization of the same intent. Evidence-guided frame revision. When the retrieved evidence cannot support the current frame, Revise determines whether the mismatch can be resolved without changing the underlying analyt- ical intent. It adapts the current frame according to the retrieved data and previous frame attempts: 퐹 (푡+1) 푖 = Revise 퐹 (푡) 푖 ,퐸 (푡) 푖 ,H 푖 | 퐼 푖 ,(4) whereH 푖 records the previous unsuccessful frame attempts for the same analytical intent. An LLM revises퐹 (푡) 푖 by adjusting its target measure, organizing dimensions, and data scope according to what is supported by퐸 (푡) 푖 , while preserving the analytical intent퐼 푖 . The historyH 푖 prevents the model from repeatedly proposing frames that have already failed. If a distinct and evidence-compatible frame can be produced, it becomes퐹 (푡+1) 푖 and guides the next retrieval round. If no new frame consistent with퐼 푖 can be found, Revise returns∅, and the candidate visualization is dropped. Termination and output. For each candidate visualization, FECA terminates when one of three conditions is met: (i) Align returns Accept, producing an accepted frame–evidence pair(퐹 ∗ 푖 ,퐸 ∗ 푖 ) (Sec- tion 2.3); (i) Revise returns∅, indicating that no alternative frame consistent with퐼 푖 can be constructed (Section 2.3); or (i) the it- eration limit퐵is reached without an Accept decision. The latter two cases drop the visual frame, while all accepted frame–evidence pairs form the setC ∗ . 2.4 Grounded visualization and report generation Once the set of accepted frame–evidence pairsC ∗ has been ob- tained, the task shifts from deciding what can be supported by the available evidence to deciding how that evidence should be visual- ized. Detailed visualization specification therefore begins only after a frame–evidence pair has been accepted. Subsequent specification and rendering operate on fixed퐹 ∗ 푖 and퐸 ∗ 푖 and cannot alter either the frame or its evidence. After rendering, the initial outline is re- composed according to the report-level textual sources, accepted frame–evidence pairs, and successfully rendered visualizations. Visualization generation. For each accepted frame–evidence pair (퐹 ∗ 푖 ,퐸 ∗ 푖 ) ∈ C ∗ , the visualization planner produces an executable specification푃 푖 that defines the chart form, data bindings, and visual encoding. The renderer then translates푃 푖 into ECharts [15] code: 퐹 ∗ 푖 ,퐸 ∗ 푖 Planner −→ 푃 푖 Renderer −→ 푉 푖 .(5) All data bindings in푃 푖 are derived directly from퐸 ∗ 푖 . The planner de- termines how the accepted evidence is visually encoded, but cannot Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation modify its values or introduce quantitative values not contained in 퐸 ∗ 푖 . A render supervisor then validates the generated chart for speci- fication conformance, successful execution, and visual readability. Rendering failures are returned to the planner or renderer for re- tries. These retries may correct the specification or implementation, but퐹 ∗ 푖 and퐸 ∗ 푖 remain unchanged throughout the process. Charts that pass supervision form the final visualization setV ∗ . Report recomposition and generation. Because frame revision and rendering may change which visualizations are ultimately avail- able, the initial outlineO (0) may no longer accurately reflect the evidence and visualizations produced by FECA. We therefore re- compose the outline using the report-level textual sourcesS, the accepted frame–evidence pairs, and the successfully rendered visu- alizations: O ∗ = Recompose O (0) ,S, 퐹 ∗ 푖 ,퐸 ∗ 푖 ,푉 푖 푉 푖 ∈V ∗ .(6) Recompose updates the initial narrative claims according to the report-level textual sources, evidence, and rendered visualizations. In particular, it revises the corresponding section claim to reflect 퐸 ∗ 푖 and places each visualization alongside the claim it supports. Finally, the report is generated as R= Write O ∗ ,S, 퐹 ∗ 푖 ,퐸 ∗ 푖 ,푉 푖 푉 푖 ∈V ∗ .(7) The resulting report follows the recomposed narrative and inte- grates each successfully rendered visualization with its accepted evidence. 3 Experimental Setup In this section, we describe the dataset, evaluation metrics, baselines, and implementation details used in our experiments. 3.1 Dataset We evaluate all systems on the 100 research topics in MultimodalRe- portBench [35], curated from expert-authored multimodal reports published by Pew Research, Our World in Data, and the Open Knowledge Foundation. MultimodalReportBench starts from open-ended topics and re- quires systems to independently research, organize, and visualize information from the web. Topics involve quantitative trends, com- parisons, and distributions across heterogeneous sources, making it suitable for evaluating report utility and numerical fidelity of gener- ated charts. For each topic, we generate a report with each system using its report-generation procedure and evaluate the resulting reports and charts. 3.2 Evaluation metrics We evaluate the generated reports along three dimensions: numeri- cal fidelity, report and chart quality, and numerical richness. Numerical fidelity. Analogous to atomic-claim evaluation in long- form text [22], we treat each numeric unit displayed in a chart as an independently verifiable quantitative claim and conduct a two-stage audit. The first stage checks each unit against its associated evidence 퐸 푖 for both value agreement and semantic consistency in measure, unit, entity, and temporal scope. A chart satisfiesFaithful(푉 푖 ,퐸 푖 ) only if all its numeric units pass this audit. Units that cannot be re- solved against their associated evidence proceed to a second-stage targeted open-web audit, which searches for external evidence matching the same measure, unit, entity, and temporal scope. This stage distinguishes values that are externally verifiable but unsup- ported by the evidence originally associated with the chart from values for which no matching evidence can be recovered. At either stage, after normalizing units and scales, a unit is con- sidered correct only when these semantic attributes match and its relative error from the source value does not exceed 10%. Each audited unit is assigned one of five mutually exclusive outcomes: Given-Source Correct, Given-Source Error, Open-Search Correct, Open-Search Error, or Not Found. At the chart level, Provided mea- sures source availability, while Valid measures whether the pro- vided source is accessible and relevant to the chart. Report and chart quality. Following prior evaluations of multi- modal research reports [21,35], we use an LLM evaluator to assess report-level quality along four dimensions: Coverage and Depth, Coherence and Organization, Visual Integration, and Overall Re- port Quality. We introduce four chart-level dimensions that assess whether individual charts function effectively as visual evidence: (i) Chart Contribution measures whether the chart provides mean- ingful quantitative evidence for the local analysis. (i) Chart Self- -Documentation measures whether the chart contains sufficient information to be interpreted independently. (i) Chart Design eval- uates whether its visual encoding and presentation communicate the underlying data clearly. (iv) Figure–Text Consistency measures whether the chart is connected to and supports the surrounding discussion. All report- and chart-quality metrics are scored on a 1–5 scale. Numerical richness. To determine whether higher fidelity is ob- tained by displaying less quantitative information, we report Raw Numerical Richness, the average number of numeric units per chart, and Correct Numerical Richness, the average number verified as correct through either audit stage. 3.3 Baselines We compare FECA with six representative deep research systems from two complementary families. Multimodal deep research. We compare against three multi- modal deep research systems specifically designed to generate text–visual or text–chart interleaved reports. (i) Multimodal Deep- Researcher (MMDR) plans and generates text–chart interleaved reports from open-web research [35]. (i) CogGen recursively re- fines report structures and visual plans and audits generated chart coordinates against its collected evidence [32]. (i) TVIR uses spe- cialized agents to retrieve images, generate charts, and compose text–visual interleaved reports [21]. General-purpose deep research systems. We further compare against three general-purpose deep research systems that are not specifically designed for multimodal report generation but can con- duct open-web research, data analysis, and report synthesis. (i) Evo- Scientist is a general-purpose multi-agent research system whose current open-source implementation coordinates planning, web research, code execution, data analysis, visualization, and research writing through specialized agents [20]. (i) OpenAI Deep Research (OpenAI-DR) conducts multi-step web research and synthesizes retrieved information into detailed reports with inline citations [25]. Yue and Zhang, et al. 0 20 40 60 80 100 Numeric units (%) 11.10 88.90 19.20 80.80 Before verificationAfter verification GroundedUnsupported Figure 3: Effect of post-hoc verification on the unsupported- value rate in a plan–then–check pipeline. (i) Perplexity Deep Research (Perplexity-DR) performs extensive web search and synthesizes information from multiple sources into detailed reports with source citations [28]. 3.4 Implementation details Across all locally executed systems, including MMDR [35], Cog- Gen [32], TVIR [21], and EvoScientist [20], we use GPT-4.1-mini for search-query formulation and GPT-4.1 for all other LLM calls in the report-generation pipeline [26]. Web search and webpage extraction are performed through the Firecrawl API [8]. For FECA, MMDR, CogGen, and TVIR, we standardize the initial evidence- collection budget to two retrieval rounds, with three queries issued per round and three results retrieved per query. We preserve each system’s original report-generation procedure except for necessary task-specific adaptations. For TVIR, we retain its chart-generation branch but disable its image-retrieval branch, since retrieved images fall outside the scope of our evaluation. For FECA, the maximum frame–evidence co-adaptation budget is set to 퐵= 4. OpenAI-DR [25] and Perplexity-DR [28] are evaluated through their official APIs and retain their native models, decoding con- figurations, and retrieval backends. For numerical-fidelity evalu- ation, we use GPT-5.4 [27] to extract numerical units from chart images, assess the validity of the provided sources, and verify the extracted units against the corresponding evidence. Firecrawl is used to search, retrieve, and parse cited and open-web sources during verification. 4 Experimental Results We organize our experiments around four research questions (RQs). (RQ1) What limits the plan–then–check paradigm? (RQ2) Does FECA improve chart-level numerical fidelity without sacrificing report utility? (RQ3) Which components of FECA contribute to effective frame–evidence co-adaptation? (RQ4) How does FECA balance adaptation effectiveness and computational cost? 4.1 Limitations of fixed plan–then–check pipelines To answer RQ1, we first investigate whether the numerical fidelity issues in plan–then–check pipelines can be addressed through ver- ification alone. We evaluate two settings: a practical LLM-based verifier and an idealized oracle filtering mechanism. Verification Table 1: Oracle provenance filtering eliminates unsupported values but reveals the fidelity–utility trade-off of fixed plan– then–check pipelines. MetricFixed PlanOracle FilterΔ Unsupported rate (%)↓80.80.0 −80.8 p Charts per report↑6.001.45 −4.55 Visual integration↑2.821.73 −1.09 Overall quality↑3.092.91 −0.18 can identify unsupported outputs, but cannot resolve mismatches between predefined visual frames and available evidence. Practical verification provides limited fidelity improvement. Following the verification design of CogGen [32], we implement a representative plan–then–check pipeline that generates charts un- der fixed visual frames and verifies plotted values against their cor- responding evidence. The verifier can correct unsupported values when valid supporting evidence is already available; however, it nei- ther modifies the underlying visual frame nor performs additional evidence acquisition. We randomly sample 20 reports generated by this pipeline and compare charts before and after verification in terms of unsupported-value rate. As shown in Figure 3, verification increases the proportion of supported values from 11.1% to 19.2%, but the unsupported-value rate remains high at 80.8%. These results indicate that post-hoc ver- ification provides only limited improvements in numerical fidelity, leaving most plotted values unsupported by the routed evidence. Oracle verification reveals the limits of fixed frames. The limited improvement above raises the question of whether stricter verification could resolve the issue. To isolate the effect of verifica- tion quality, we construct an oracle provenance gate that retains a chart only when every plotted value has a complete provenance chain from the retrieved evidence to the final visualization. As shown in Table 1, oracle filtering reduces the unsupported- value rate to zero by construction, but removes more than three quarters of generated charts and substantially decreases Visual Integration. This result reveals a fundamental limitation of the plan–then–check paradigm: when visual frames are fixed before ev- idence acquisition, verification can only reject unsupported charts; it cannot adapt the analytical frame to recover evidence-compatible alternatives. 4.2 Evidence fidelity and report quality To answer RQ2, we evaluate whether FECA improves evidence- grounded chart generation while preserving the analytical utility of multimodal deep research reports. We compare FECA with repre- sentative baselines in terms of numerical fidelity, report and chart quality, and further conduct a blinded human evaluation to assess whether the observed improvements align with human judgments. Numerical fidelity. We first evaluate whether quantitative claims encoded in generated charts are faithfully grounded in supporting evidence. Following the audit protocol described in Section 3.2, we verify each numeric unit by first checking the source provided with the chart and, when unresolved, conducting targeted open-web searches. We report source availability and validity at the chart level, as well as the distribution of audit outcomes at the value level. Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation Table 2: Source availability and numerical fidelity of generated charts. Bold and underlined values denote the best and second- best results on fidelity-related metrics, respectively. System# Charts Source Availability (%)Given-Source Audit (%)Open-web Audit (%) Provided↑Valid↑Correct↑Error↓Correct↑Error↓Not Found↓ FECA86399.2999.2982.670.213.411.3712.34 MMDR27582.8167.928.481.021.2512.4476.81 CogGen43326.40100.005.561.023.78 8.6281.02 TVIR325100.0098.1532.0111.035.386.8144.77 OpenAI-DR265100.0099.1560.73 10.932.434.4521.46 Perplexity-DR663100.0098.6755.994.822.585.3731.24 EvoScientist21251.72100.0028.430.97 0.7220.2449.64 Table 3: Report and chart quality evaluation. C&D: Coverage and Depth; C&O: Coherence and Organization; VI: Visual Integration. Contrib.: Chart Contribution; Self-Doc.: Chart Self-Documentation; Consist.: Figure–Text Consistency. Bold and underlined values denote the best and second-best results, respectively. Report QualityChart QualityNumerical Richness SystemC&D C&OVIOverall Contrib. Self-Doc. Design Consist.RawCorrect FECA3.94 4.02 3.783.893.893.923.634.3721.0518.12 MMDR3.974.003.37 3.903.583.823.152.7713.811.34 CogGen3.963.892.973.403.043.203.113.4518.841.76 TVIR3.943.973.633.873.603.593.744.316.062.26 OpenAI-DR3.713.993.663.823.713.073.604.424.222.67 Perplexity-DR 4.044.013.453.783.463.043.374.214.022.35 EvoScientist3.183.962.863.433.313.843.672.7714.314.17 As shown in Table 2, we observe three findings: (i) Combining values verified as correct from the provided sources and through open-web search, FECA achieves the highest verified-correct rate of 86.08% and the lowest Not Found rate of 12.34%. Its verified-correct rate exceeds those of TVIR and OpenAI-DR, the strongest open- source and commercial baselines, by 48.69 and 22.92 percentage points, respectively. (i) The advantage of FECA cannot be attrib- uted solely to better source availability. Several baselines achieve comparable Provided and Valid scores but substantially lower value- level correctness, indicating that retrieving relevant sources does not guarantee that the quantitative values visualized in charts are actually supported by those sources. (i) Additional open-web veri- fication improves FECA’s verified-correct proportion by only 3.41 percentage points, suggesting that most correct numeric units are already directly grounded in the evidence retrieved during the FECA process. Report and chart quality. We next investigate whether FECA’s fidelity improvement compromises report utility or quantitative richness. Using the report- and chart-level metrics defined in Sec- tion 3.2, we evaluate both the overall reports and individual vi- sualizations. We further distinguish raw numerical richness from correct numerical richness to determine whether improved fidelity is achieved by simply reducing quantitative content. The results in Table 3 show that: (i) At the report level, FECA achieves the best Coherence, Organization, and Visual Integration scores, while remaining within 0.01 and 0.10 points of the best- performing systems on Overall Report Quality and Coverage and Depth, respectively. This demonstrates that improved numerical fidelity does not come at the cost of reduced analytical coverage Table 4: Human pairwise evaluation over 20 randomly sam- pled topics. We report the percentage of topics where FECA wins, ties, or loses against each baseline. Baseline FECA Win / Tie / Lose (%) FidelityIntegrationOverall MMDR90/5/565/15/2045/20/35 TVIR75/10/1550/25/2545/25/30 OpenAI-DR35/40/2555/15/30 60/10/30 or report quality. (i) At the chart level, FECA achieves the highest Chart Contribution and Chart Self-Documentation scores, remains within 0.05 points of the best system in Figure–Text Consistency, and maintains competitive Chart Design quality. These results in- dicate that FECA-generated charts are not only evidence-grounded but also informative, interpretable, and well integrated with textual analysis. (i) FECA provides 11.7% more numeric units per chart and 4.35×more verified-correct numeric units than the strongest baseline, showing that fidelity improvements result from better evidence alignment rather than reduced quantitative content. Human evaluation. To examine whether the improvements ob- served in automatic evaluation correspond to human preferences, we conduct a blinded pairwise evaluation on 20 randomly sampled topics from the test set. We compare FECA against two representa- tive open-source systems (MMDR and TVIR) and the commercial OpenAI-DR. For each topic, three annotators independently com- pare two anonymized reports presented in randomized order along three dimensions: Numerical Fidelity, Visual Integration, and Over- all Report Quality. Final preferences are determined by majority Yue and Zhang, et al. Table 5: Ablation study of iterative evidence retrieval and visual frame revision. Bold values indicate best results per column. ComponentsResults System Iterative Retrieval Frame Revision # Charts per Report Verified-Correct Rate (%) Visual Integration Overall Report Quality Chart Contribution Figure–Text Consistency Single-Pass ×3.1677.663.403.703.914.18 Fixed Frame✓×4.4670.263.523.823.814.23 FECA✓8.6386.083.783.893.894.37 voting, and we report the proportion of topics where FECA wins, ties, or loses against each baseline. As shown in Table 4, FECA is consistently preferred over MMDR and TVIR in numerical fidelity and achieves comparable perfor- mance with OpenAI-DR. Across all comparisons, FECA obtains more wins than losses in both visual integration and overall report quality. These human judgments further confirm that FECA im- proves numerical grounding while preserving the usefulness and readability of generated research reports. 4.3 Ablation study of frame–evidence co-adaptation To answer RQ3, we isolate the contributions of iterative evidence retrieval and evidence-guided frame revision, to understand the sources of FECA’s improvements. We compare the full FECA model with two progressively simplified variants: (i) Single-Pass, which performs only one retrieval round without frame revision; and (i) Fixed Frame, which enables iterative retrieval but keeps the initial visual frame unchanged. The full FECA model incorporates both mechanisms to enable iterative frame–evidence co-adaptation. Table 5 shows the impact of each component. (i) Compared with Single-Pass, Fixed Frame produces 41.1% more charts and modestly improves Visual Integration and Overall Report Quality. However, its verified-correct rate decreases by 7.40 percentage points, and its Chart Contribution also declines. This suggests that additional retrieval can increase the amount of visualizable evidence, but does not by itself resolve mismatches between a fixed visual frame and the retrieved evidence. (i) Adding evidence-guided frame revision to Fixed Frame increases the number of charts per report by 93.5% and the verified-correct rate by 15.82 percentage points, accompa- nied by consistent improvements in all report- and chart-quality metrics. Frame revision therefore enables the system to transform newly retrieved evidence into both more abundant and more re- liable visual evidence. (i) Full FECA achieves the best result on five of the six evaluation metrics and remains within 0.02 points of the best Chart Contribution score. Taken together, these results show that FECA’s gains come from combining iterative retrieval with evidence-guided adaptation of the visual frame, rather than from performing additional searches alone. 4.4 Adaptation efficiency and frame resolution dynamics To answer RQ4, we further analyze the efficiency and behavior of FECA during frame–evidence co-adaptation. Specifically, we study how the iteration budget퐵affects the quality–cost trade-off and 1234 Iteration Budget B 3 4 5 6 7 8 9 Charts per Report Charts per ReportVisual IntegrationOverall Quality 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4.0 4.1 Quality Score Figure 4: Effect of the iteration budget퐵on the number of evidence-supported charts and report quality. examine how initialized visual frames are resolved throughout the adaptation process. Effect of iteration budget on adaptation efficiency. We exam- ine how the iteration budget affects FECA’s adaptation effectiveness and computational cost. We vary the maximum number of itera- tions퐵from 1 to 4 while keeping all other settings unchanged. For each budget, we measure the number of charts per report, Visual Integration, Overall Report Quality, and token consumption. Figure 4 shows how the number of evidence-supported charts, Vi- sual Integration, and Overall Report Quality change as퐵increases. Together with the corresponding token measurements, we observe that (i) Increasing퐵consistently produces more evidence-supported charts. The chart count doubles from퐵=1 to퐵=2, while the gains in the third and fourth rounds decrease to 1.31 and 1.00 ad- ditional charts per report, respectively. Thus, the second iteration produces the largest increase in the number of evidence-supported charts. (i) Visual Integration continues to improve with additional iterations, whereas Overall Report Quality largely saturates after 퐵=2. Later iterations therefore primarily improve the availability and integration of visual evidence rather than the overall quality of the report. (i) Token consumption increases from 812K at퐵=1 to 1,046K, 1,199K, and 1,279K at퐵=2, 3, and 4, respectively. Increas- ing퐵from 1 to 4 raises token consumption by 57.5% but increases the number of charts per report by 173.1%. These results expose a configurable trade-off: smaller budgets reduce computational cost, whereas larger budgets recover more evidence-supported visual- izations. We use퐵=4 in the main experiments to prioritize the generation of evidence-supported charts. Frame resolution dynamics during co-adaptation. We analyze when initialized visual frames become evidence-compatible during the co-adaptation process. Each frame is categorized according Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation Accept@1 24.50% Accept@2 34.10% Accept@3 12.50% Accept@4 7.50% Dropped 21.30% Figure 5: Resolution stages of initialized visual frames during frame–evidence co-adaptation. to the iteration at which it is accepted, denoted as Accept@푏, or as Dropped if no compatible realization can be obtained within the iteration budget. Accept@1 corresponds to frames that can be directly supported by retrieved evidence; acceptance in later iterations requires one or more rounds of evidence-guided revision. Figure 5 summarizes the resolution process of initialized frames. We observe that: (i) only 24.5% of initialized frames are accepted in the first iteration, while evidence-guided adaptation enables an additional 54.1% to be accepted in later rounds. (i) among all accepted frames, 68.8% require at least one additional adaptation round, demonstrating that frame revision is a fundamental mecha- nism rather than an occasional fallback. (i) although most frames are accepted within the first two iterations, the third and fourth iter- ations recover an additional 20.0% of initialized frames. Meanwhile, unresolved frames are dropped, preventing unsupported visual claims from entering the final report. These results explain why increasing the adaptation budget leads to more evidence-grounded visualizations. 5 Related Work Reliable chart generation in multimodal deep research lies at the intersection of two research directions: multimodal deep research that aim to improve the trustworthiness of generated reports, and evidence-grounded chart generation methods that study how visu- alizations can be generated from available data or documents. Multimodal deep research. Existing multimodal deep research systems apply reliability checks at different stages of chart gener- ation. Multimodal DeepResearcher uses visual feedback to refine generated charts [35]; TVIR verifies data authenticity and cross- source consistency before plotting [21]; and CogGen audits gen- erated chart values against its collected evidence [32]. Although operating at different stages, these approaches generally assume a fixed visual frame: the frame is determined before knowing whether the open web contains the quantitative evidence required for its realization, and subsequent checks only evaluate whether the re- trieved evidence or generated output satisfies the predefined frame. These methods improve reliability within a fixed visual frame but cannot resolve frame–evidence mismatch when the required evidence is unavailable. FECA addresses this limitation by enabling visual frames to adapt to retrieved evidence before chart generation. Evidence-grounded chart generation. Outside deep research, reliable chart generation has primarily been studied in settings where the input evidence is known in advance. Existing work can be broadly grouped into two settings: (i) table-to-chart generation, where a structured data table and visualization request are pro- vided, ranging from specification generation and constraint-based recommendation [6,23,24] to LLM-based methods that improve executable code and visual quality through feedback, debugging, or iterative refinement [5,10,14,29,33,36]; and (i) document-to-chart generation, where charts are generated from data-bearing docu- ments, either automatically [39] or according to a predefined chart request [11]. In both settings, the supporting data or source docu- ments are available before visualization planning and generation, making evidence acquisition outside the scope of the task. Open-web report generation introduces a different dependency between visualization planning and quantitative evidence acquisi- tion. A system must determine what a chart should communicate without knowing whether the required evidence exists or can be reliably recovered from the web. FECA extends evidence-grounded chart generation to this setting by aligning visual frames with evidence acquired during generation, rather than with evidence provided beforehand. 6 Conclusion In this work, we revisit chart generation in multimodal deep re- search from the perspective of evidence–visualization alignment. We identify a fundamental limitation of existing plan–then–check pipelines: fixed visual frames can only verify or reject generated charts, but cannot adapt when required evidence is unavailable. This suggests that trustworthy visualization requires not only vali- dating charts after generation, but also adapting visual frames to the evidence that can actually be obtained. To address this chal- lenge, we introduce Frame–Evidence Co-Adaptation (FECA), an evidence-adaptive framework that integrates evidence acquisition with visual frame adaptation. FECA preserves high-level analyti- cal intents while allowing their visual realizations to be accepted, revised, or dropped according to retrieved evidence. By enabling it- erative interaction between visual frames and supporting evidence, FECA reframes chart generation as an evidence-grounded reason- ing process rather than fixed-plan verification. Experiments on 100 real-world research topics demonstrate that FECA substantially im- proves numerical fidelity while maintaining the analytical quality and usability of generated reports. Despite these improvements, several challenges remain: (i) FECA still depends on the availability and reliability of external evidence, and may struggle when relevant quantitative information is sparse, conflicting, or incomplete; (i) FECA relies on LLM-based reason- ing for frame–evidence alignment and revision, which can still be affected by hallucinations, uncertainty, or imperfect judgments. As a result, unsupported numerical content may occasionally pass through the co-adaptation process even when the overall proce- dure operates correctly; and (i) FECA currently keeps the report outline and its analytical intents fixed during frame–evidence co- adaptation, leaving open the broader challenge of jointly adapting Yue and Zhang, et al. higher-level report plans, visual frames, and evidence during open- ended investigation. Future work will explore uncertainty-aware evidence reconcili- ation, robust numerical verification beyond LLM judgments, and hierarchical co-evolution of research plans, visual frames, and evi- dence in long-horizon research agents. Ethical Considerations This work relies on publicly accessible web sources for evidence retrieval and report generation. Third-party sources are used only for research and grounded synthesis, with their original links re- tained for attribution. We do not intentionally collect private or personally identifiable information. Despite its evidence-alignment mechanisms, FECA cannot guarantee that generated charts are objectively correct: retrieved sources may themselves be outdated, incomplete, or biased, and errors may still arise during retrieval, extraction, frame adaptation, or chart generation. Generated re- ports should therefore be treated as references requiring human verification rather than as authoritative evidence, particularly in high-stakes domains such as healthcare, finance, and public policy. Our human evaluation involved three graduate-student volunteers who were informed of the study purpose and evaluated anonymized system outputs. The evaluation did not require participants to dis- close personal information or intentionally expose them to harmful content. References [1]Mubashara Akhtar, Nikesh Subedi, Vivek Gupta, Sahar Tahmasebi, Oana Co- carascu, and Elena Simperl. 2024. ChartCheck: Explainable Fact-Checking over Real-World Chart Images. In Findings of the Association for Computational Lin- guistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 13921–13937. https://doi.org/10.18653/v1/2024.findings-acl.828 [2]Akari Asai, Jacqueline He, Rulin Shao, Weijia Shi, Amanpreet Singh, Joseph Chee Chang, Kyle Lo, Luca Soldaini, Sergey Feldman, Mike D’arcy, et al.2026. Synthe- sizing scientific literature with retrieval-augmented language models. Nature 650, 8103 (2026), 857–863. https://doi.org/10.1038/s41586-025-10072-4 [3]Mengyuan Chen, Chengjun Dai, Xinyang Dong, Chengzhe Feng, Kewei Fu, Jian- she Li, Zhihan Peng, Yongqi Tong, Junshao Zhang, and Hong Zhu. 2025. Dingtalk DeepResearch: A Unified Multi Agent Framework for Adaptive Intelligence in Enterprise Environments. arXiv:2510.24760 [cs.CL] https://arxiv.org/abs/2510. 24760 [4]Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. 2025. MindSearch: Mimicking Human Minds Elic- its Deep AI Searcher. In The Thirteenth International Conference on Learn- ing Representations. https://proceedings.iclr.c/paper_files/paper/2025/hash/ e01c431b83153632c0dcfaf8ccda0a-Abstract-Conference.html [5]Victor Dibia. 2023. LIDA: A Tool for Automatic Generation of Grammar-Agnostic Visualizations and Infographics using Large Language Models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Danushka Bollegala, Ruihong Huang, and Alan Ritter (Eds.). Association for Computational Linguistics, Toronto, Canada, 113–126. https://doi.org/10.18653/v1/2023.acl-demo.11 [6] Victor Dibia and Çağatay Demiralp. 2019. Data2Vis: Automatic Generation of Data Visualizations Using Sequence-to-Sequence Recurrent Neural Networks. IEEE Computer Graphics and Applications 39, 5 (2019), 33–46. https://doi.org/10. 1109/MCG.2019.2924636 [7]Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. 2025. DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents. arXiv:2506.11763 [cs.CL] https://arxiv.org/abs/2506.11763 [8]Firecrawl. 2026.Firecrawl Documentation.https://docs.firecrawl.dev/ introduction. Accessed: 2026-08-21. [9] Kung-Hsiang Huang, Mingyang Zhou, Hou Pong Chan, Yi Fung, Zhenhailong Wang, Lingyu Zhang, Shih-Fu Chang, and Heng Ji. 2024. Do LVLMs Understand Charts? Analyzing and Correcting Factual Errors in Chart Captioning. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 730–749. https://doi.org/10.18653/v1/2024.findings-acl.41 [10]Mohammed Saidul Islam, Md Tahmid Rahman Laskar, Md Rizwan Parvez, Enamul Hoque, and Shafiq Joty. 2024. DataNarrative: Automated Data-Driven Story- telling with Visualizations and Texts. In Proceedings of the 2024 Conference on Em- pirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 19253–19286. https://doi.org/10.18653/v1/2024.emnlp-main.1073 [11]Akriti Jain, Pritika Ramu, Aparna Garimella, and Apoorv Saxena. 2025. Doc2Chart: Intent-Driven Zero-Shot Chart Generation from Documents. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Pro- cessing. Association for Computational Linguistics, Suzhou, China, 34948–34963. https://doi.org/10.18653/v1/2025.emnlp-main.1770 [12]Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan O Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning. (2025).https: //openreview.net/forum?id=Rwhi91ideu [13]Gary Klein, Jennifer K. Phillips, Erica L. Rall, and Deborah A. Peluso. 2007. A Data–Frame Theory of Sensemaking. In Expertise Out of Context: Proceedings of the Sixth International Conference on Naturalistic Decision Making, Robert R. Hoffman (Ed.). Lawrence Erlbaum Associates, 113–155. [14]Woosung Koh, Janghan Yoon, MinHyung Lee, Youngjin Song, Jaegwan Cho, Jaehyun Kang, Taehyeon Kim, Se-Young Yun, Youngjae Yu, and Bongshin Lee. 2025.퐶 2 : Scalable Auto-Feedback for LLM-based Chart Generation. In Pro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol- ume 1: Long Papers), Luis Chiruzzo, Alan Ritter, and Lu Wang (Eds.). Associ- ation for Computational Linguistics, Albuquerque, New Mexico, 4525–4566. https://doi.org/10.18653/v1/2025.naacl-long.232 [15]Deqing Li, Honghui Mei, Yi Shen, Shuang Su, Wenli Zhang, Junting Wang, Ming Zu, and Wei Chen. 2018. ECharts: A declarative framework for rapid construction of web-based visualization. Visual Informatics 2, 2 (2018), 136–146. https://doi.org/10.1016/j.visinf.2018.04.011 [16] Minghao Li, Ying Zeng, Zhihao Cheng, Cong Ma, and Kai Jia. 2025.Re- portBench: Evaluating Deep Research Agents via Academic Survey Tasks. arXiv:2508.15804 [cs.CL] https://arxiv.org/abs/2508.15804 [17]Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025. Search-o1: Agentic Search-Enhanced Large Reasoning Models. CoRR abs/2501.05366 (2025). https://doi.org/10.48550/ARXIV. 2501.05366 arXiv:2501.05366 [18] Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji- Rong Wen, Yutao Zhu, and Zhicheng Dou. 2025. WebThinker: Empower- ing Large Reasoning Models with Deep Research Capability. In Advances in Neural Information Processing Systems, Vol. 38. Curran Associates, Inc., 120091–120131. https://proceedings.neurips.c/paper_files/paper/2025/hash/ ae03bdef276132fae089692445725635-Abstract-Conference.html [19] Zijian Li, Xin Guan, Bo Zhang, Shen Huang, Houquan Zhou, Shaopeng Lai, Ming Yan, Yong Jiang, Pengjun Xie, Fei Huang, Jun Zhang, and Jingren Zhou. 2026. WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open- Ended Deep Research. In The Fourteenth International Conference on Learning Representations. https://openreview.net/forum?id=MtNCJjlrKt [20]Yougang Lyu, Xi Zhang, Xinhao Yi, Yuyue Zhao, Shuyu Guo, Wenxiang Hu, Jan Piotrowski, Jakub Kaliski, Jacopo Urbani, Zaiqiao Meng, Lun Zhou, and Xiaohui Yan. 2026. EvoScientist: Towards Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery. arXiv preprint arXiv:2603.08127 (2026). [21]Xinkai Ma, Zhiqi Bai, Dingling Zhang, Pei Liu, Yishuo Yuan, He Zhu, Jiakai Wang, Qianqian Xie, 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, and Jiaheng Liu. 2026. TVIR: Building Deep Research Agents Towards Text–Visual Interleaved Report Generation. arXiv:2606.02320 [cs.CL] https://arxiv.org/abs/2606.02320 [22]Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine- grained Atomic Evaluation of Factual Precision in Long Form Text Generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Singapore, 12076–12100. https://doi.org/10.18653/v1/2023.emnlp-main.741 [23]Dominik Moritz, Chenglong Wang, Greg L. Nelson, Halden Lin, Adam M. Smith, Bill Howe, and Jeffrey Heer. 2019. Formalizing Visualization Design Knowledge as Constraints: Actionable and Extensible Models in Draco. IEEE Transactions on Visualization and Computer Graphics 25, 1 (2019), 438–448. https://doi.org/10. 1109/TVCG.2018.2865240 [24] Arpit Narechania, Arjun Srinivasan, and John Stasko. 2021. NL4DV: A Toolkit for Generating Analytic Specifications for Data Visualization from Natural Language Queries. IEEE Transactions on Visualization and Computer Graphics 27, 2 (2021), 369–379. https://doi.org/10.1109/TVCG.2020.3030378 [25]OpenAI. 2025. Deep Research. OpenAI API Documentation. https://developers. openai.com/api/docs/guides/deep-research [26]OpenAI. 2025. Introducing GPT-4.1 in the API. https://openai.com/index/gpt-4-1/. Accessed: 2026-08-21. Faithful Chart Generation for Multimodal Deep Research: Frame–Evidence Co-Adaptation [27]OpenAI. 2026. Introducing GPT-5.4. https://openai.com/index/introducing-gpt- 5-4/. Accessed: 2026-08-21. [28]Perplexity. 2025. Sonar Deep Research. Perplexity API Documentation. https: //docs.perplexity.ai/docs/sonar/models/sonar-deep-research [29]Fatemeh Pesaran Zadeh, Juyeon Kim, Jin-Hwa Kim, and Gunhee Kim. 2024. Text2Chart31: Instruction Tuning for Chart Generation with Automatic Feed- back. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 11459–11480. https://doi.org/10.18653/v1/2024.emnlp-main.640 [30]Yijia Shao, Yucheng Jiang, Theodore Kanell, Peter Xu, Omar Khattab, and Monica Lam. 2024. Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 1: Long Papers). Association for Computational Linguistics, Mex- ico City, Mexico, 6252–6278. https://doi.org/10.18653/v1/2024.naacl-long.347 [31]Manasi Sharma, Chen Bo Calvin Zhang, Chaithanya Bandi, Clinton Wang, Ankit Aich, Huy Nghiem, Tahseen Rabbani, Ye Htet, Brian Jang, Sumana Basu, Aish- warya Balwani, Denis Peskoff, Marcos Ayestaran, Sean M. Hendryx, Brad Ken- stler, and Bing Liu. 2026. ResearchRubrics: A Benchmark of Prompts and Rubrics For Evaluating Deep Research Agents. In The Fourteenth International Conference on Learning Representations. https://openreview.net/forum?id=ErnvfmSX0P [32]Kuo Tian, Pengfei Sun, Zhen Wu, Junran Ding, and Xinyu Dai. 2026. CogGen: A Cognitively Inspired Recursive Framework for Deep Research Report Generation. In Findings of the Association for Computational Linguistics: ACL 2026, Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens (Eds.). Association for Computational Linguistics, San Diego, California, United States, 5961–5988. https://doi.org/10.18653/v1/2026.findings-acl.296 [33] Chenglong Wang, Bongshin Lee, Steven M. Drucker, Dan Marshall, and Jianfeng Gao. 2025. Data Formulator 2: Iterative Creation of Data Visualizations, with AI Transforming Data Along the Way. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, Article 677, 17 pages. https://doi.org/10.1145/3706598. 3713296 [34] Jinming Wu, Zihao Deng, Wei Li, Yiding Liu, Bo You, Bo Li, Zejun MA, and Ziwei Liu. 2026. MMSearch-R1: Incentivizing LMMs to Search. (July 2026), 2456–2487. https://doi.org/10.18653/v1/2026.acl-long.114 [35]Zhaorui Yang, Bo Pan, Han Wang, Yiyao Wang, Xingyu Liu, Luoxuan Weng, Yingchaojie Feng, Haozhe Feng, Minfeng Zhu, Bo Zhang, and Wei Chen. 2026. Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports from Scratch with Agentic Framework. Proceedings of the AAAI Conference on Artificial Intelligence 40, 40 (March 2026), 34368–34377. https://doi.org/10.1609/aaai.v40i40. 40734 [36]Zhiyu Yang, Zihan Zhou, Shuo Wang, Xin Cong, Xu Han, Yukun Yan, Zheng- hao Liu, Zhixing Tan, Pengyuan Liu, Dong Yu, Zhiyuan Liu, Xiaodong Shi, and Maosong Sun. 2024. MatPlotAgent: Method and Evaluation for LLM-Based Agen- tic Scientific Data Visualization. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 11789–11804. https://doi.org/10.18653/v1/2024.findings-acl.701 [37]Huanjin Yao, Qixiang Yin, Min Yang, Ziwang Zhao, Yibo Wang, Haotian Luo, Jingyi Zhang, and Jiaxing Huang. 2026. M-DeepResearch: A Simple and Effective Multimodal Agentic Search Baseline.arXiv:2603.01050 [cs.CV] https://arxiv.org/abs/2603.01050 [38]Fangda Ye, Kuicai Dong, Xie Zhifei, Yuxin Hu, Yihang Yin, Shurui Huang, Shikai Dong, Chen Zhang, Jianzhu Bao, and Shuicheng Yan. 2026. Deep-Reporter: Deep Research for Grounded Multimodal Long-Form Generation. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens (Eds.). Association for Computational Linguistics, San Diego, California, United States, 41137–41177. https://doi.org/10.18653/v1/2026.acl-long.1909 [39] Songheng Zhang, Lei Wang, Toby Jia-Jun Li, Qiaomu Shen, Yixin Cao, and Yong Wang. 2024. Chartifytext: Automated chart generation from data-involved texts via llm. arXiv preprint arXiv:2410.14331 (2024). [40]Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pen- grui Lu, and Pengfei Liu. 2025. DeepResearcher: Scaling Deep Research via Reinforcement Learning in Real-world Environments. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Suzhou, China, 414–431. https://doi.org/10.18653/ v1/2025.emnlp-main.22

Instant research discovery