DataMagic: Transforming Tabular Data into Data Insight Video Yupeng Xie
Chen Ma
Zhenyang Wang
Liangwei Wang
Jiayi Zhu
HKUST (GZ)
China Unicom
HKUST (GZ)
HKUST (GZ)
HKUST (GZ)
Chuxuan Zeng
Zhouan Shen
Boyan Li
Yuyu Luo∗
China Unicom
HKUST (GZ)
HKUST (GZ)
HKUST (GZ)
arXiv:2606.20388v1 [cs.HC] 18 Jun 2026
ABSTRACT Data videos integrate dynamic charts, voice narration, and synchronized animations to communicate data insights as temporal narratives, making them an effective medium for improving data consumption efficiency in the data management lifecycle. However, producing high-quality data videos requires expertise spanning data analysis, narrative design, and video production. Existing approaches fall short: static visualization tools (e.g., BI dashboards) lack narrative logic and animation; authoring tools require users to pre-prepare visualizations rather than working from raw data; pixel-level video generation models cannot guarantee data fidelity or provenance. We demonstrate DataMagic, an end-to-end interactive system that transforms raw tabular data and natural language queries into narrative data-insight videos. To ensure data fidelity, DataMagic introduces the declarative specification DVSpec, which binds visual and animation elements to underlying data fields through data-driven semantic references. To address the combinatorial explosion of the design space, DataMagic adopts a Generatethen-Orchestrate multi-agent architecture that generates candidate scenes in parallel and then optimizes narrative coherence through global orchestration. Leveraging DVSpec’s decoupling of logic and rendering, the system further supports three interaction modes and structured provenance-based data Q&A, transforming one-way videos into explorable interactive data interfaces. Evaluation on 109 real-world samples validates the effectiveness of the DataMagic.
1
INTRODUCTION
Data videos integrate dynamic charts, voice narration, and synchronized animations to communicate data insights as temporal narratives, attracting growing attention from both academia and industry [1, 24, 34]. However, producing high-quality data videos requires cross-domain expertise spanning data analysis, narrative design, and video production [24], and existing approaches all fall short of end-to-end automation from raw data to data video [11, 27, 31–33]. Static visualization tools (e.g., BI dashboards, HAIChart [36], DeepEye [15]) output static charts, lacking narrative logic and animation. Authoring tools (e.g., Data Playwright [23]) focus on adding animation effects to existing charts, without extracting insights from raw data or handling multi-scene narrative orchestration. Pixel-level video generation models (e.g., Sora [13]) can synthesize ∗ Yuyu Luo is the corresponding author.
This work is licensed under the Creative Commons BY-NC-ND 4.0 International License. Visit https://creativecommons.org/licenses/by-nc-nd/4.0/ to view a copy of this license. For any use beyond those covered by this license, obtain permission by emailing [email protected]. Copyright is held by the owner/author(s). Publication rights licensed to the VLDB Endowment. Proceedings of the VLDB Endowment, Vol. 14, No. 1 ISSN 2150-8097. doi:XX.XX/XXX.XX
videos, but their black-box nature frequently produces numerical hallucinations and cannot map visual elements back to underlying data records. Our key observation is that effective data videos are fundamentally structured narratives rather than simple assemblies of visual elements [24]. This motivates us to model end-to-end generation as a hierarchical content orchestration problem. Two core challenges emerge: (1) how to design a structured intermediate representation that precisely describes heterogeneous components and their temporal relations while ensuring data fidelity and provenance; (2) how to efficiently search the vast design space for solutions that balance local scene quality and global narrative coherence. To this end, we demonstrate DataMagic1 , an end-to-end interactive system from raw tabular data to narrative data videos. For challenge (1), we introduce DVSpec (Data Video Specification), a declarative specification that binds visual and animation elements to underlying data fields through data-driven semantic references and narration-index triggering, ensuring full provenance. For challenge (2), DataMagic adopts a Generate-then-Orchestrate multiagent architecture [39] that generates diverse candidate scenes in parallel and then applies global orchestration to optimize scene selection, ordering, and narrative coherence. Leveraging DVSpec’s decoupling of logic and rendering, the system further supports three interaction modes and structured provenance-based data Q&A, enabling users to refine videos and explore data directly. The main contributions are as follows: (1) We demonstrate DataMagic, introducing the declarative specification DVSpec and the Generate-then-Orchestrate multi-agent strategy, which mechanistically ensure data fidelity and provenance. (2) We design three progressive demonstration scenarios covering automated generation, multimodal editing, and provenance-based data Q&A, and validate system effectiveness on 109 real-world samples.
2
SYSTEM ARCHITECTURE
As shown in Figure 1, DataMagic takes raw tabular datasets and natural language queries as input, processes them through a multiagent engine to produce a DVSpec configuration, and compiles it into a complete narrative video. The core design revolves around three components: DVSpec declarative specification (Section 2.1), the multi-agent generation pipeline (Section 2.2), and provenancebased interaction and exploration (Section 2.3).
2.1
DVSpec: Declarative Data Video Specification
Existing declarative specifications (e.g., Vega-Lite [21], Canis [7], ChartMark [5]) perform well for static charts or single-chart annotations and animations, but have not been extended to the unified 1 DataMagic: https://datamagic-home.github.io/
Stage 1: Candidate Scene Generation
Input Data Query
Story Planner Task Decomposition
Data Manager
Visual Designer
Data Processing Planning
Chart Specification
Parallel Execution
Insight Summary
“Analyze Q4 revenue
peaks, regional marketing efficiency, and top product profit contributors ....”
Candidate Scenes
S1
subquery analytic type
S2
Python/Pandas
...
data field type priority
Sn
DVSPEC
S1
chart type style
S2
data binding insight
...
...
Sn
Stage 2: Global Narrative Orchestration Animation Coordinator
Narration Director Scene Re-rank Narration Generation
title narration Marketing Spend vs. Return Rate by Region
Narration-Synchronized Animation
entrance emphasize leave
...we examine how our regional marketing spend drives performance results.
We start by identifying our Laptop Pro X ... 55% of total core profit drivers among our ... 4K Monitor ... around 20%. product lineup. The other three together ...
Output Data Video
Interactive Refinement
Script Editing
NL Command
DVSpec field
Direct Manipulation
DVSpec fields DVSpec field
......
What can I do for you? Swap scene 2 and 3
Figure 1: System architecture of DataMagic. DVSpec Structural Components
description of cross-modal content and temporal coordination required for multi-scene data videos [3, 4]. To fill this gap, we design DVSpec (Data Video Specification), a declarative specification that fully decouples logical description from rendering implementation, serving as a structured intermediate representation between the multi-agent generation engine and the rendering engine: the generation stage writes analysis results to DVSpec, while interactive editing changes are also mapped to local DVSpec updates. DVSpec itself is rendering-library agnostic; the current system renders charts using D3.js and synthesizes videos with Remotion. As shown in Figure 2(a), DVSpec formalizes a data video as a combination of metadata 𝑀 and an ordered scene sequence 𝑆 = ⟨𝑠 1, . . . , 𝑠𝑛 ⟩: 𝑉 := (𝑀, 𝑆). Each scene 𝑠𝑖 is defined as a four-tuple 𝑠𝑖 := (type, content, narration, animation), where content encapsulates visualization configuration (chart type, data bindings, style parameters), narration is an ordered list of narration segments, and animation is a list of animation effects. The scene-based design stems from research on data video narrative structures [1, 22]. DVSpec introduces two key mechanisms to ensure data fidelity and audio-video synchronization: (1) Data-driven semantic references. Visual elements are referenced through data attribute values (e.g., {"company": "Nvidia"}) rather than hard-coded identifiers, ensuring that references remain valid when data is updated or chart types change, and allowing every visual element to be precisely traced back to the underlying tabular data. Minor naming inconsistencies (e.g., casing or whitespace differences) are resolved through fuzzy matching at render time, and the pipeline applies a retry policy on code-execution failures.
Scene Sequence S
[s1,s2,...,sn]
s
Scene Unit i Type Content Narration Animation chart, chart type, [n1,n2,...,nk] [a1,a2,...,ak] stat_cards... data, style...
Specify & Construct {
}
"meta": { Meta Config (M) },
"scenes": [
{
"id": "scene_1",
Scene1 "type": "opening",
"content": { Title, Background, Style },
"narration": [ { Narration Unit 1 } ],
"animations": [ ]
},
{
"id": "scene_2",
Scene2 "type": "chart",
"content": { Chart Type, Data, Binding, Style },
"narration": [ { Narration Unit 1 },
{ Narration Unit 2 }, ],
"animations": [ { Entrance Anim },
{ Emphasis Anim (Target: "MQ")},
{ Emphasis Anim (Target: "AA")}, ]
},
// ... more scenes (e.g. stat_cards, closing) ...
]
(a) Declarative Grammar for Data Videos
nguage-Specific Converters
La
Visualization
Animation
ECharts
GSAP
.js
Anime.js
D3
ega-Lite
emotion
V
R
...
...
Rendered Video
...
(b) Composition & Rendering Flow
Figure 2: DVSpec structure and rendering flow.
(2) Narration-index declarative triggering. Animation trigger timing is declared using narration segment indices rather than absolute timestamps. During rendering, the system automatically aligns animations based on the actual audio duration generated by 2
A
B
PREVIEW
Analyze Q4 revenue peaks, regional marketing efficiency, and top pro...
video script export
D
Al Edit
N
atural language
A A A IP ATI
D1
C NV S
M N UL ON
cene 4. change to tree
ma and highlight top
roduct @S p p
one. Scene 4 updated to
treemap. Laptop Pro X
highlighted in blue. D
ic roduct has the ig est to Q4 growth rate Wh h p h h Q3 ?
Analysis · 7 steps. Done 4 Monitor 2. Runner up Laptop Pro X K
: +5
-
+5
2.2%
:
Add a new scene showing monthly revenue trend for the top 2 growth products
0s
Total Revenue by Product... 54s-1:05 4 Total Revenue by Product (Q4 Performance)
1:
bar chart · 40s-52s
NARRATION Finally, we look at total revenue to confirm our peak Q4 performance leaders. grow bars · -- Bar chart entrance
ew scene generated
and inserted as Scene 8.
-- Highlight Laptop Pro X
3 more rows
@Scene 4
· chart
D
Analysis · 7 steps
one
ound dataset Resolve Spec binding columns product_name onceptual roll-up same logic as code) GROUP B product_name calendar uarter sale_date ) SUM total_amount ) filter year 2025, uarter 3, 4 LS Highe t re en e gro th B
:
DV
:
C
(
:
,
Y
_q
(
→
(
;
q
∈ {
U
s
=
}
ION
Q3→Q4
v
u
w
(2025): 4K
at nner p
Monitor Ru
-u
+52.5% ($177K → $270K).
: Laptop Pro X
($496K → $754K).
o rce
S
u
:
D3
at
+52.2%
sales with_products.csv
D
N
grow bars fade in scale in sequential ask AI
The Laptop Pro X remains the clear revenue leader. pulse
+
30
tyle Editing
D2 DATA-DRIVEN Q&A
CONC
C SCENE TIMELINE STRIP
Product_name Total_amount Laptop Pro X 754320 4K Monitor 269920
5% 1
S
--
ata Q&A
E E E ATI
SC NE
G N R ON
S
cene Generation
Figure 3: DataMagic web interface with the main UI components (A-D), (D1-D3) highlight key interaction flows. text-to-speech (TTS), so that when users modify narration text, animation synchronization relationships are automatically maintained without manual keyframe adjustment [22].
2.2
(e.g., “macro-to-micro”, “phenomenon before cause”). It generates narration for each scene conditioned on adjacent scenes’ context, ensuring narrative coherence across scene transitions. The Animation Coordinator uses DVSpec’s semantic reference mechanism to bind data entities mentioned in the narration to corresponding visual elements, achieving end-to-end audio-video synchronization. The generated DVSpec is compiled into the final video by the rendering engine, with narration text synthesized into speech via TTS. The system adopts a model-agnostic design to support different LLM backends.
Multi-Agent Generation Pipeline
Data video generation involves complex dependencies among data processing, visual design, and narrative logic. Single-stage approaches struggle to ensure both per-scene data accuracy and overall narrative coherence. DataMagic adopts a Generate-then-Orchestrate two-stage architecture, where parallel generation of candidate scenes enables the orchestration stage to perform global optimization rather than making greedy step-by-step decisions:
2.3
Stage 1: Data-driven candidate generation. The Story Planner decomposes the user’s high-level query into several independent analytical sub-tasks based on the analytical dimensions involved (e.g., temporal trends, category comparisons, regional distributions). For each sub-task, the Data Manager plans the data processing workflow and automatically generates Python code to extract, filter, and aggregate relevant data slices from the source table. The Visual Designer then designs the visualization scheme (chart type, data bindings, and style configuration), extracts key insights, and populates the content field of DVSpec. This stage generates a pool of candidate scenes in parallel, decoupling low-level data processing from high-level narrative construction.
Provenance-Based Interaction and Exploration
Prior work has established data provenance as a key enabler for interactive visualizations [19]. DataMagic extends this idea to the data video setting: DVSpec records the binding relationships from visual elements to underlying data fields during the generation stage. This structured data provenance information permeates the system’s downstream interactions, supporting two core capabilities. Parametric Editing. Since DVSpec explicitly records the data fields and narration segments corresponding to each visual and animation element, any modification by the user can be precisely located and applied locally, without regenerating the entire video [26, 30]. The system provides three interaction modes: canvas direct manipulation (e.g., switching chart types, adjusting data mappings), declarative script editing (directly modifying narration or animation parameters), and natural language commands (e.g., “@Scene 4:
Stage 2: Global narrative orchestration. The Narration Director selects scenes from the candidate pool based on insight value and query coverage, and plans a playback order following narrative logic 3
(2) Declarative script editing. The audience directly modifies narration text or chart parameters in the narration editor (region C). When text changes cause audio duration to change, the system automatically re-aligns animation trigger timing without manual keyframe adjustment. (3) Natural language commands. The audience enters commands in the AI chat panel, e.g., “@Scene 4: change to treemap and highlight top product” (as shown in Figure 3 D1). The system parses the command and automatically updates the corresponding scene’s DVSpec configuration, triggering real-time incremental re-rendering.
change to treemap and highlight top product”). All three share the same DVSpec state, maintaining context consistency across mode switches. Data Q&A. The provenance relationships preserved by DVSpec enable the system to map users’ natural language questions to specific data binding fields: the system uses an LLM to parse the question [17, 25, 35], identifies relevant data fields from the current scene’s DVSpec bindings, and constructs structured query operations (e.g., filtering, aggregation, extremum retrieval) to execute directly on the underlying tabular data [10], rather than relying on visual models to infer content from pixels. Users can also transform insights discovered through Q&A into new video scenes, completing the full loop from data exploration to narrative expansion.
3
3.3
This scenario demonstrates how DataMagic transforms one-way data videos into explorable interactive data interfaces. Traditional video models output pure pixel streams and cannot understand the data meaning behind the visuals, whereas DVSpec preserves the complete semantic and data context, enabling precise structured retrieval and data provenance rather than relying on visual models to infer content from pixels. The audience asks questions about the currently playing video scene (e.g., “Which product has the highest Q3 to Q4 growth rate?”). As shown in Figure 3 D2, the system locates the data binding fields for the current scene in DVSpec, jointly queries the original dataset, and returns an accurate answer (e.g., 4K Monitor: +52.5%, Runnerup: Laptop Pro X +52.2% ↑). The audience can also ask global questions, experiencing multi-granularity data exploration from scenelevel to video-level. Furthermore, the audience can instruct the system to transform Q&A insights into new video scenes (e.g., “Add a new scene showing monthly revenue trend for the top 2 growth products”). As shown in Figure 3 D3, the system automatically generates the corresponding DVSpec configuration and inserts the new scene into the video sequence, completing the full loop from data exploration to narrative expansion.
DEMONSTRATION SCENARIOS
We design three scenarios to showcase DataMagic’s capabilities in end-to-end data video generation, multimodal editing, and provenance-based data Q&A, with quantitative evaluation to validate generation quality. The demonstration is conducted through a web interface (Figure 3), which includes a data and history panel (A), a real-time video preview panel (B), a scene timeline and narration editor (C), and an AI-assisted editing panel (D). Users interact with the system directly at each stage of the workflow.
3.1
Scenario 1: Data-Driven Automated Generation
This scenario demonstrates how DataMagic transforms raw data into a narrative video. The user uploads a business CSV file (e.g., technology company financial reports) and enters a query (e.g., “Analyze Q4 revenue peaks, regional marketing efficiency, and top product profit contributors”). The user can observe the multi-agent engine executing in real time: query decomposition, data slicing, chart design, and narrative orchestration. Once generation is complete, the video plays in the embedded player (Figure 3 B). Unlike traditional black-box generation, the system transparently displays the corresponding scene timeline, narration text, animation labels, and underlying data tables in region C, allowing the audience to intuitively understand the strict binding between visual content and underlying data fields. The history panel reveals intermediate outputs for traceable generation.
3.2
Scenario 3: Structured Provenance-Based Data Q&A
3.4
Quantitative Evaluation
We evaluated DataMagic on 109 real-world samples from DACompDA [9] and T2R-bench [38]. We first used four LLMs (DeepSeekV3.2 [6], Gemini-2.5-Pro [8], GPT-5 [18], and Claude-Sonnet-4 [2]) to directly generate data videos in one pass, probing current capability limits; we also integrated these models into the DataMagic framework to verify model-agnosticism. The evaluation covered execution rate (the fraction of samples for which the full pipeline completes without error and renders a playable video) and five quality dimensions (Intent Fulfillment, Data Insights, Narrative Quality, Animation Effectiveness, Aesthetic Quality; 1–5 scale) [14, 16, 28, 29, 37]. Gemini-2.5-Pro [8] served as an automated judge (Pearson 𝑟 =0.91 vs. human experts on 60 sampled videos). Results showed that even the strongest LLMs faced significant challenges in data video generation: average quality scores ranged from 1.91 to 2.22, with execution rates between 48–86%. The main issues were audio-visual temporal misalignment and lack of narrative structure: models frequently produced videos where narration was out of sync with visuals and scenes lacked logical transitions.
DVSpec-Based Multimodal Editing
Data reporting is inherently iterative. Traditional pixel-level videos are extremely difficult to modify after generation, whereas DataMagic enables parametric editing through DVSpec. The three interaction modes share the same DVSpec state, and the audience can freely switch between them based on the video generated in Scenario 1: (1) Canvas direct manipulation. The audience clicks to select a chart on the video canvas and the system displays a Property Inspector panel. Users can switch chart types via dropdown menus (e.g., convert bar to pie chart), modify data field mappings, or adjust color schemes, as well as drag elements or double-click to edit text. All interactions are mapped in real-time to local DVSpec updates. 4
DataMagic addressed these problems through DVSpec’s narrationindex triggering mechanism and the Generate-then-Orchestrate strategy, raising the average score to 3.89 with execution rates above 95%. The largest gains were in Animation Effectiveness (1.84→4.03, +120%) and Narrative Quality (2.00→3.43, +72%). On average, direct generation completes in ∼57 seconds; DataMagic totals ∼176 seconds (60 s configuration generation + 116 s rendering at parallelism = 3), reflecting a deliberate quality–latency trade-off. Ablation experiments confirm that both components are indispensable: removing the Story Planner degrades average quality by 11.6% (Narrative −15.1%), and removing Orchestration by 9.0% (Intent −12.4%). Training a dedicated data video model could further reduce API dependency and latency [12, 20].
[16] Yuyu Luo, Nan Tang, Guoliang Li, Chengliang Chai, Wenbo Li, and Xuedi Qin. 2021. Synthesizing natural language to visualization (NL2VIS) benchmarks from NL2SQL benchmarks. In Proceedings of the 2021 International Conference on Management of Data. 1235–1247. [17] Yuyu Luo, Nan Tang, Guoliang Li, Jiawei Tang, Chengliang Chai, and Xuedi Qin. 2021. Natural language to visualization by neural machine translation. IEEE Transactions on Visualization and Computer Graphics 28, 1 (2021), 217–226. [18] OpenAI. 2025. GPT-5 System Card. https://openai.com/index/gpt-5-systemcard/. [19] Fotis Psallidas and Eugene Wu. 2018. Provenance for interactive visualizations. In Proceedings of the Workshop on Human-In-the-Loop Data Analytics. 1–8. [20] Yanlin Qi, Xinhang Chen, Huiqiang Jiang, Qitong Wang, Botao Peng, and Themis Palpanas. 2026. ParisKV: Fast and Drift-Robust KV-Cache Retrieval for LongContext LLMs. arXiv preprint arXiv:2602.07721 (2026). [21] Arvind Satyanarayan, Dominik Moritz, Kanit Wongsuphasawat, and Jeffrey Heer. 2017. Vega-Lite: A Grammar of Interactive Graphics. IEEE Transactions on Visualization and Computer Graphics 23, 1 (2017), 341–350. [22] Leixian Shen et al. 2024. Data Player: Automatic Generation of Data Videos with Narration-Animation Interplay. IEEE Trans. Vis. Comput. Graph. 30, 1 (2024), 109–119. [23] Leixian Shen et al. 2025. Data Playwright: Authoring Data Videos With Annotated Narration. IEEE Trans. Vis. Comput. Graph. 31, 9 (2025), 5884–5897. [24] Leixian Shen, Haotian Li, Yun Wang, and Huamin Qu. 2025. Reflecting on Design Paradigms of Animated Data Video Tools. In CHI. ACM, 190:1–190:21. [25] Leixian Shen, Enya Shen, Yuyu Luo, Xiaocong Yang, Xuming Hu, Xiongshuai Zhang, Zhiwei Tai, and Jianmin Wang. 2022. Towards natural language interfaces for data visualization: A survey. IEEE transactions on visualization and computer graphics 29, 6 (2022), 3121–3144. [26] Shuyu Shen, Sirong Lu, Leixian Shen, and Yuyu Luo. 2026. Debugging Defective Visualizations: Empirical Insights Informing a Human-AI Co-Debugging System. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–24. [27] Tarique Siddiqui et al. 2016. Effortless Data Exploration with zenvisage: An Expressive and Interactive Visual Analytics System. Proc. VLDB Endow. 10, 4 (2016), 457–468. [28] Xiaoyan Su, Peijie Dong, Zhenheng Tang, Song Tang, Yuyao Zhai, Kaitao Lin, Liang Chen, Gai Yuhang, Yuyu Luo, Qiang Wang, et al. 2026. VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing. arXiv preprint arXiv:2605.15677 (2026). [29] Yinghao Tang, Xueding Liu, Boyuan Zhang, Tingfeng Lan, Yupeng Xie, Jiale Lao, Yiyao Wang, Haoxuan Li, Tingting Gao, Bo Pan, et al. 2026. IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation. arXiv preprint arXiv:2601.04498 (2026). [30] Yinghao Tang, Yupeng Xie, Yingchaojie Feng, Tingfeng Lan, and Wei Chen. 2026. sketch-plot: Progressive Editing for Text-to-Image Academic Figures. arXiv preprint arXiv:2606.09171 (2026). [31] Yinghao Tang, Yupeng Xie, Yingchaojie Feng, Tingfeng Lan, Jiale Lao, Yue Cheng, and Wei Chen. 2026. ViviDoc: Generating Interactive Documents through Human-Agent Collaboration. arXiv preprint arXiv:2603.27991 (2026). [32] Yinghao Tang, Yupeng Xie, Yingchaojie Feng, Jiale Lao, Tingfeng Lan, and Wei Chen. 2026. Demonstrating chart-plot: Closing the Last Mile of Academic Chart Generation. arXiv preprint arXiv:2606.09174 (2026). [33] Manasi Vartak et al. 2015. SEEDB: Efficient Data-Driven Visualization Recommendations to Support Visual Analytics. Proc. VLDB Endow. 8, 13 (2015), 2182–2193. [34] Liangwei Wang, Zhengxuan Zhang, Yi-Fan Cao, Fugee Tsung, and Yuyu Luo. 2026. TableTale: Reviving the Narrative Interplay Between Data Tables and Text in Scientific Papers. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. 1–17. [35] Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. 2024. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. In Findings of the Association for Computational Linguistics: EMNLP 2024. 12174–12200. [36] Yupeng Xie, Yuyu Luo, Guoliang Li, and Nan Tang. 2024. HAIChart: Human and AI Paired Visualization System. Proc. VLDB Endow. 17, 11 (2024), 3178–3191. [37] Yupeng Xie, Zhiyang Zhang, Yifan Wu, Sirong Lu, Jiayi Zhang, Zhaoyang Yu, Jinlin Wang, Sirui Hong, Bang Liu, Chenglin Wu, and Yuyu Luo. 2025. VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations. arXiv:2510.22373 [cs.CL] [38] Jie Zhang, Changzai Pan, Kaiwen Wei, et al. 2025. T2R-bench: A Benchmark for Generating Article-Level Reports from Real World Industrial Tables. CoRR abs/2508.19813 (2025). https://doi.org/10.48550/ARXIV.2508.19813 arXiv:2508.19813 [39] Yizhang Zhu, Liangwei Wang, Chenyu Yang, Xiaotian Lin, Boyan Li, Wei Zhou, Xinyu Liu, Zhangyang Peng, Tianqi Luo, Yu Li, et al. 2025. A Survey of Data Agents: Emerging Paradigm or Overstated Hype? arXiv preprint arXiv:2510.23587 (2025).
ACKNOWLEDGMENTS This paper was supported by the NSF of China (62402409); Youth S&T Talent Support Programme of Guangdong Provincial Association for Science and Technology (SKXRC2025461); the Young Talent Support Project of Guangzhou Association for Science and Technology (QT-2025-001); Guangzhou Basic and Applied Basic Research Foundation (2026A1515010269, 2025A04J3935, 2023A1515110545); Guangzhou-HKUST(GZ) Joint Funding Program (2025A03J3714); and Research and Development of Heterogeneous Computing Interconnection and Scheduling Software Stack (YF202400000003).
REFERENCES [1] Fereshteh Amini et al. 2015. Understanding Data Videos: Looking at Narrative Visualization through the Cinematography Lens. In CHI. ACM, 1459–1468. [2] Anthropic. 2025. Claude Sonnet 4. https://www.anthropic.com/news/claudesonnet-4. [3] Yiru Chen et al. 2025. Physical Visualization Design: Decoupling Interface and System Design. Proc. ACM Manag. Data 3, 3 (2025), 197:1–197:27. [4] Yiru Chen, Jeffrey Tao, and Eugene Wu. 2023. DIG: The Data Interface Grammar. In HILDA@SIGMOD. ACM, 7:1–7:7. [5] Yiyu Chen, Yifan Wu, Shuyu Shen, Yupeng Xie, Leixian Shen, Hui Xiong, and Yuyu Luo. 2025. ChartMark: A Structured Grammar for Chart Annotation. In 2025 IEEE Visualization and Visual Analytics (VIS). IEEE, 311–315. [6] DeepSeek-AI. 2025. DeepSeek-V3 Technical Report. arXiv:2412.19437 [cs.CL] [7] Tong Ge, Weiwei Cui, Bongshin Lee, Huamin Qu, and Dongmei Zhang. 2020. Canis: A High-level Language for Cohort-based Chart Animations. In Proceedings of the ACM Conference on Human Factors in Computing Systems. [8] Google DeepMind. 2025. Gemini 2.5: Our Most Intelligent Model Yet. https://blog.google/technology/google-deepmind/gemini-model-thinkingupdates-march-2025/. [9] Fangyu Lei, Jinxiang Meng, Yiming Huang, Junjie Zhao, Yitong Zhang, Jianwen Luo, Xin Zou, Ruiyi Yang, Wenbo Shi, Yan Gao, Shizhu He, Zuo Wang, Qian Liu, Yang Wang, Ke Wang, Jun Zhao, and Kang Liu. 2025. DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle. CoRR abs/2512.04324 (2025). https://doi.org/10.48550/ARXIV.2512.04324 arXiv:2512.04324 [10] Boyan Li, Yuyu Luo, Chengliang Chai, Guoliang Li, and Nan Tang. 2024. The dawn of natural language to sql: Are we fully ready? arXiv preprint arXiv:2406.01265 (2024). [11] Boyan Li, Yiran Peng, Yupeng Xie, Sirong Lu, Yizhang Zhu, Xing Mu, Xinyu Liu, and Yuyu Luo. 2026. Deepeye: A steerable self-driving data agent system. In Companion of the International Conference on Management of Data. 74–77. [12] Tianyi Lin et al. 2025. LEAD: Iterative Data Selection for Efficient LLM Instruction Tuning. arXiv preprint (2025). [13] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. 2024. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177 (2024). [14] Tianqi Luo, Chuhan Huang, Leixian Shen, Boyan Li, Shuyu Shen, Wei Zeng, Nan Tang, and Yuyu Luo. 2026. nvbench 2.0: Resolving ambiguity in textto-visualization through stepwise reasoning. Advances in Neural Information Processing Systems 38 (2026). [15] Yuyu Luo, Xuedi Qin, Nan Tang, and Guoliang Li. 2018. Deepeye: Towards automatic data visualization. In 2018 IEEE 34th international conference on data engineering (ICDE). IEEE, 101–112. 5