MaRVL-QA Dataset

MaRVL-QA is a benchmark for evaluating the mathematical and spatial reasoning capabilities of Multimodal Large Language Models (MLLMs). The dataset is designed to test these models' ability to reason directly from images of mathematical surface plots, isolating the reasoning task from the semantic noise often found in natural images.

---

Dataset Structure

The dataset is divided into two subsets, each tailored to a specific reasoning task.

Topological Counting

This task challenges models to identify and enumerate topological features, such as local maxima, within a single plot.

Data Fields:

• image: An image of the mathematical function's plot.
• question_type: The type of question being asked about the plot (e.g., "maxima", "minima").
• answer: The correct numerical answer to the question.
• func_id: A structured identifier for the function, formatted as tier.family_id.function_id. This allows for tiered evaluation as described in the paper.
• family: The family of the function (e.g., "Wave Surface").
• equation_str: The mathematical equation of the function.
• plot_range_x: The range of the x-axis for the plot.
• plot_range_y: The range of the y-axis for the plot.
• plot_type: The type of plot used (e.g., "heatmap", "contour_heatmap").
• color_map: The color map used in the plot (e.g., "magma", "inferno").

Transformation Recognition

This task requires the model to recognize the geometric transformation applied to an original plot (Plot A) to produce a new one (Plot B). The input image for this task contains both Plot A (left) and Plot B (right).

Data Fields:

• image: The image containing both the original and transformed plots.
• answer: The correct option number corresponding to the applied transformation.
• func_id: A structured identifier for the base function, formatted as tier.family_id.function_id.
• family: The family of the base function.
• equation_str: The mathematical equation of the base function.
• plot_range_x and plot_range_y: The plot range for both the original and transformed plots. As a controlled variable, the plot range remains consistent across the transformation.
• plot_type: The type of plot used.
• color_map: The color map used in the plot.
• options: A list of possible transformation choices presented to the model.
• transformation_params: The parameters of the ground truth transformation that was applied.

---

Citation

If you find this dataset useful in your research, please cite the following paper:

@misc{pande2025marvlqabenchmarkmathematicalreasoning,
      title={MaRVL-QA: A Benchmark for Mathematical Reasoning over Visual Landscapes}, 
      author={Nilay Pande and Sahiti Yerramilli and Jayant Sravan Tamarapalli and Rynaa Grover},
      year={2025},
      eprint={2508.17180},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.17180}, 
}