Dataset Card for StructViz-Bench

Dataset Summary

StructViz-Bench is a benchmark for evaluating Multimodal Large Language Model (MLLM) reasoning over visualized structured data. It systematically measures visualization sensitivity -- the degree to which a model's reasoning performance changes when the same underlying data is presented in different visual formats.

The benchmark encompasses three primary data modalities (tabular, time-series, and graph), 14 distinct visualization types, 4 question difficulty levels (1-hop, 2-hop, 3-hop, and counterfactual), and a diverse set of reasoning tasks. It contains 3,795 base items, which expand to 18,315 rendered instances per model. Released evaluations cover four core models (GPT-4o, Gemini Flash, Qwen2.5-VL-7B, Claude Sonnet) for a total of 73,260 instances, plus a supplementary same-family scale evaluation on Qwen2.5-VL-32B for testing whether parameter scale eliminates format dependence.

Supported Tasks

• Visual Question Answering (VQA): Answering natural language questions grounded in visualized structured data, requiring extraction, comparison, aggregation, and multi-step reasoning.
• Multimodal Reasoning: Evaluating the capacity of models to perform multi-hop inference and counterfactual reasoning over visual inputs derived from structured data.
• Visualization Robustness Evaluation: Quantifying the consistency of model performance across different visual representations of identical underlying data.
• Cross-Modal Reasoning: Assessing model performance on composite items that require simultaneous reasoning over two distinct data modalities (e.g., tabular + time-series, tabular + graph, time-series + graph).

Languages

All questions, answers, and metadata are in English (en).

Dataset Structure

Data Instances

A representative example from base_items.jsonl:

{
  "question_id": "tabular_000003::difficulty=1-hop",
  "question": "Which 'ticker' has the highest 'close_price' value?",
  "answer": "BET",
  "modality": "tabular",
  "data_id": "finance_stock_prices",
  "task": "ranking",
  "difficulty": "1-hop",
  "viz_methods": ["bar_chart", "heatmap", "table_image", "scatter_plot", "text_only"],
  "metadata": {"domain": "finance", "description": "Monthly stock activity for a set of listed companies."},
  "data_format": "dataframe_records",
  "data": [ ... ],
  "data_columns": ["ticker", "period", "close_price", "trade_volume_k", "daily_return_pct", "market_segment"]
}

Data Fields

Field	Type	Description
question_id	string	Unique identifier encoding modality, item index, and difficulty level.
question	string	The natural language question posed to the model.
answer	string	The ground-truth answer, generated via programmatic solvers.
modality	string	The data modality: tabular, timeseries, graph, mixed_tab_ts, mixed_tab_graph, or mixed_ts_graph.
source	string	Provenance of the underlying data: synthetic, scitabalign, ett, or networkx_realworld.
viz_methods	list[string]	Applicable visualization types for the item (e.g., bar_chart, heatmap, line_plot, node_link, text_only).
difficulty	string	Reasoning depth: 1-hop, 2-hop, 3-hop, or counterfactual.
task	string	The reasoning task category (e.g., value_extraction, ranking, comparison, trend_analysis, aggregation, filtering, correlation, forecasting, peak_identification, pattern_classification, anomaly_detection, connectivity, shortest_path, community, degree_query).
data_id	string	Identifier for the underlying data instance.
data	object	The raw structured data used to generate the visualization and formulate the question.
metadata	object	Domain-specific metadata including domain name, description, and column or feature descriptions.
data_format	string	Format descriptor for the data field (e.g., dataframe_records, json_dict).
data_columns	list[string]	Column names for tabular data (when applicable).

Data Splits

StructViz-Bench is an evaluation-only benchmark and does not provide training or validation splits. The data is organized into three files:

File	Items	Description
base_items.jsonl	1,500	Synthetic-only single-modality items (500 tabular + 500 time-series + 500 graph).
realworld_test.jsonl	3,795	Full benchmark: all 1,500 synthetic items plus 2,295 real-world items (1,265 SciTabAlign tabular + 870 ETT time-series + 160 NetworkX graph). This is the primary evaluation file used for all paper results.
mixed_items.jsonl	600	Cross-modal composite items (200 tab+ts, 200 tab+graph, 200 ts+graph).

Note: realworld_test.jsonl is a superset of base_items.jsonl. The name reflects that it includes real-world sources; use the source field (synthetic, scitabalign, ett, networkx_realworld) to filter by provenance.

When expanded across all applicable visualization types, the base items yield 18,315 rendered instances per model. Across four evaluated models, this produces a total of 73,260 evaluated instances.

Dataset Creation

Curation Rationale

Existing multimodal benchmarks typically focus on a single data type (e.g., charts or tables) or employ a fixed visualization style per data instance. This design conflates the difficulty of the underlying reasoning task with the difficulty imposed by the specific visual encoding. StructViz-Bench was created to fill this gap by decoupling data content from visual representation, thereby enabling controlled measurement of visualization sensitivity in MLLM reasoning.

Source Data

The benchmark draws on both synthetic and real-world data sources:

• Synthetic data: Programmatically generated from parameterized templates across multiple domains (finance, healthcare, environmental science, etc.) to ensure controlled complexity and balanced coverage of reasoning tasks.
• Real-world tabular data: Derived from SciTabAlign, a collection of scientific tables with structured annotations.
• Real-world time-series data: Derived from the ETT (Electricity Transformer Temperature) dataset, containing multivariate sensor readings.
• Real-world graph data: Classic and empirical graph structures generated via NetworkX, including social networks and infrastructure topologies.

Annotations

All questions and ground-truth answers are generated programmatically using quality-validated templates and deterministic solvers. This approach ensures:

• Correctness: Answers are computed directly from the underlying data, eliminating human annotation errors.
• Scalability: The template-based system supports systematic generation across modalities, difficulty levels, and task types.
• Reproducibility: Given the same data and configuration, the benchmark can be regenerated identically.

Annotation Process

No manual annotation was performed. The question-answer generation pipeline consists of:

1. Data instantiation from parameterized templates or real-world sources.
2. Question generation via modality-specific and task-specific template functions.
3. Answer computation using deterministic programmatic solvers.
4. Quality validation through automated consistency checks and unit tests.

Considerations for Using the Data

Social Impact

StructViz-Bench is intended to advance the scientific understanding of MLLM capabilities and limitations in interpreting visualized structured data. It may contribute to improved model robustness and more reliable deployment of vision-language systems in data analysis workflows.

Biases

• Domain coverage: While the benchmark spans multiple domains (finance, healthcare, environmental science, infrastructure), it does not exhaustively represent all possible application domains.
• Visualization design: The 14 visualization types selected represent common and widely-used formats but do not cover all possible visual encodings (e.g., 3D visualizations, interactive dashboards).
• Language: All content is in English, limiting direct applicability to multilingual evaluation settings.

Limitations

• Evaluation-only: This benchmark is designed exclusively for evaluation. Using it for model training may lead to overfitting on the specific templates and visualization styles employed.
• Synthetic component: Although real-world data sources are included, the synthetic generation component may not fully capture the noise, irregularities, and complexity of all real-world visualization scenarios.
• Static visualizations: The benchmark evaluates static rendered images only and does not address interactive or dynamic visualization modalities.
• Answer format: Ground-truth answers are expressed as short strings (exact match evaluation), which may not capture the full spectrum of valid natural language responses.

Additional Information

Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).

Citation Information

@inproceedings{structvizbench2026,
  title={StructViz-Bench: A Unified Benchmark for Evaluating MLLM Reasoning over Visualized Structured Data},
  author={Anonymous},
  booktitle={NeurIPS 2026 Datasets and Benchmarks Track},
  year={2026}
}

Dataset Statistics

Statistic	Value
Total base items	3,795
Single-modality items (base_items)	1,500
Cross-modal items (mixed_items)	600
Real-world items (realworld_test)	3,795
Visualization types	14
Data modalities	3 (tabular, timeseries, graph)
Difficulty levels	4 (1-hop, 2-hop, 3-hop, counterfactual)
Rendered instances per model	18,315
Total evaluated instances (4 models)	73,260
Task categories	49

Visualization Types

Tabular modality (5 types):

• bar_chart, heatmap, table_image, scatter_plot, text_only

Time-series modality (5 types):

• line_plot, gaf (Gramian Angular Field), recurrence_plot, heatmap, text_only

Graph modality (4 types):

• node_link, adjacency_matrix, circular_layout, text_only

Contact

For questions or issues regarding this dataset, please open an issue in the associated repository.