Multimodal QUD: Inquisitive Questions from Scientific Figures

Yating Wu, William Rudman, Venkata S Govindarajan, Alexandros G. Dimakis, and Junyi Jessy Li · 2026 preprint

A dataset of 1,250 questions evoked by scientific figures, with answers from the papers.

paper Dataset Citation Examples
Three point clouds with increasingly uneven variance across two dimensions
MQUD question · IsoScore, Figure 3 What are the implications of the middle panel's shape on dimensional utilization? Read the example and explanation

Overview

Questions evoked by scientific figures

A figure in a scientific paper can raise questions that its caption and the preceding text have not yet answered. We extend Questions Under Discussion (QUD) to study these questions and how they connect a figure to the paper’s argument.

We introduce MQUD, a dataset of 1,250 questions from 56 scientific papers, including 708 questions annotated by the papers’ original authors. The questions concern visual observations whose scientific relevance is explained later in the paper. Each is paired with an answer and supporting text.

We ask models to generate questions with three properties, defined in the paper:

01 Inquisitive

The question is not already resolved by the preceding context.

02 Salient

It concerns the paper’s research claims and is addressed later in the paper.

03 Visually grounded

It arises from an observation in the figure.

Representative questions

Questions from figure QA benchmarks

These examples illustrate the kinds of questions in ChartQA, SPIQA, and MQUD.

Worked example

Reading the IsoScore figure

Model experiments

In our experiments, open-source vision-language models such as Qwen 3.5 mostly ask questions answerable from the figure alone. Fine-tuning on MQUD helps the model generate inquisitive questions about the figure’s role in the paper’s argument. The paper describes the training and evaluation; the dataset card documents the released data.

Related projects

QSalience predicts how much a question’s answer would help a reader understand a text. QUDeval evaluates generated QUDs, and our QUD parser represents dependencies between sentences through questions and answers.

The dataset

Download the JSONL for a flat table of questions, figures, and paper passages. The Hub viewer groups the same information into nested fields.

Questions
question, answer, question_type, and difficulty. Each answer is extracted from the paper.
Figures
image gives the image path. paper_title, paper_abstract, and figure_caption supply the context.
Evidence
source_text and source_paragraphs contain the supporting passages. Use paper_id to group questions from the same paper.
Split
All 1,250 examples are in the Hub’s train split; the release has no separate test split.
Read one questionPython

Uses the Python standard library.

import json
from urllib.request import urlopen

revision = "295b4efc36226ad19429edcd4bccc151b72971d0"
url = (
    "https://huggingface.co/datasets/lingchensanwen/mqud/"
    f"resolve/{revision}/data/mqud.jsonl"
)
with urlopen(url) as response:
    row = json.loads(response.readline())

print(row["id"], row["paper_id"])
print(row["question"])
print(row["answer"])
print(row["image"])

The image value is a path in the dataset repository.

Source

Dataset card · Release counts · Dataset viewer

Download BibTeX.

Project links

Paper Data
BibTeX 
@misc{mqud2026,
  title={{Multimodal QUD: Inquisitive Questions from Scientific Figures}},
  author={Wu, Yating and Rudman, William and Govindarajan, Venkata S and Dimakis, Alexandros G. and Li, Junyi Jessy},
  year={2026},
  eprint={2604.23733},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2604.23733}
}