====================================================================
CALL FOR PARTICIPATION
AstroCLIMB: Astronomy Citation Linking from Illustrations, a Multimodal Benchmark
Shared task of the 4th Workshop on Artificial Intelligence for
Scientific Publications (WASP 2026), co-located with IJCNLP-AACL 2026
Workshop: November 9-10, 2026 (online)
Task page: https://ui.adsabs.harvard.edu/WIESP/2026/shared_task [1]
Kaggle competition: https://www.kaggle.com/competitions/astroclimb/overview [2]
Dataset: https://huggingface.co/datasets/adsabs/AstroCLIMB [3]
====================================================================
OVERVIEW
Scientists rely on figures to communicate their results, yet the information locked in those figures is hard to parse, archive, and search. Recent multimodal models promise to change that, but they have seen little testing on real scientific figures. AstroCLIMB, built in partnership with astroexplorer.org [4], offers a benchmark to probe exactly this capability: can a model reconstruct part of the citation graph of astronomy papers from their figures and captions alone?
The dataset comprises over 100K figure-and-caption pairs drawn from recent open-access astronomy papers.
TASK
Multi-class, single-label classification. Each input is a pair of objects: two figures, a figure and a caption, or two captions. The relation is symmetric, and each pair has exactly one label:
- same figure: the two objects come from the same scientific figure (figure-caption pairs only)
- same paper: the two objects come from the same paper (same DOI)
- related papers: one object's paper cites the other's
- unrelated papers: none of the above
A note on scope: AstroCLIMB is meant as a benchmark for multimodal models rather than an end in itself. Participants are not expected to train large models from scratch for it; computing basic statistics or folding it into a larger multimodal training pipeline is very much in keeping with the intent.
DATA
Two versions of the same dataset are available:
- Full dataset (Hugging Face): over 94K rows, each with a figure, its caption, and metadata (source DOI, title, authors, cited and citing DOIs) describing an astronomy citation graph as an adjacency list. Pairs are not enumerated; you generate them from the adjacency list.
https://huggingface.co/datasets/adsabs/AstroCLIMB [3]
- Evaluation dataset (Kaggle): 10K enumerated pairs each in the train and test splits, with labels. Many more pairs can be generated from the full dataset.
https://www.kaggle.com/competitions/astroclimb/overview [2]
In the Kaggle data, some objects are English-language strings, and others are images encoded as strings; a helper notebook for decoding them to images is provided on Kaggle.
EVALUATION
Systems are evaluated on Kaggle using macro-F1 over the four classes. Submissions are CSV files with per-class indicators; a sample submission is available on the competition page.
HOW TO PARTICIPATE
1. Register and join the competition on Kaggle to access the data and score submissions:
https://www.kaggle.com/competitions/astroclimb/overview [2]
2. Optionally, write up your system description as a paper using the ACL LaTeX template
(https://github.com/acl-org/acl-style-files [5]) and submit it via OpenReview:
https://openreview.net/group?id=aclweb.org/AACL-IJCNLP/2026/Workshop/WASP [6]
System description papers undergo light peer review and, if accepted, are published in the WASP 2026 proceedings in the ACL Anthology. We encourage a submission regardless of leaderboard position.
TIMELINE (subject to change; all deadlines 11.59 pm UTC-12h, anywhere on Earth)
- 1st CfP + Registration opens: July 27, 2026
- Train and scoring portion of test data release: July 27, 2026
- Registration ends: September 12, 2026
- System run and output submission: September 13, 2026
- System paper submission: September 14, 2026
- Result announcement: October 1, 2026
- Camera-ready submission: October 12, 2026
- Full data release: November 2, 2026
- Workshop: November 9-10, 2026
CONTACT
For inquiries, contact Felix Grezes at felix.grezes@cfa.harvard.edu
We look forward to your participation.
Links: ------ [1] https://ui.adsabs.harvard.edu/WIESP/2026/shared_task [2] https://www.kaggle.com/competitions/astroclimb/overview [3] https://huggingface.co/datasets/adsabs/AstroCLIMB [4] http://astroexplorer.org/ [5] https://github.com/acl-org/acl-style-files [6] https://openreview.net/group?id=aclweb.org/AACL-IJCNLP/2026/Workshop/WASP