Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AION-1: Omnimodal Foundation Model for Astronomical Sciences
Authors: Liam H. Parker, Francois Lanusse, Jeff Shen, Ollie Liu, Tom Hehir, Leopoldo Sarra, Lucas Meyer, Micah Bowles, Sebastian Wagner-Carena, Helen Qu, Siavash Golkar, Alberto Bietti, Hatim Bourfoune, Pierre Cornette, Keiya Hirashima, Geraud Krawezik, Ruben Ohana, Nicholas Lourie, Michael McCabe, Rudy Morel, Payel Mukhopadhyay, Mariel Pettee, Kyunghyun Cho, Miles Cranmer, Shirley Ho
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Trained on over 200M astronomical objects, AION-1 demonstrates strong performance across regression, classification, generation, and object retrieval tasks. Beyond astronomy, AION-1 provides a scalable blueprint for multimodal scientific foundation models that can seamlessly integrate heterogeneous combinations of real-world observations. |
| Researcher Affiliation | Academia | 1University of California, Berkeley, 2Flatiron Institute, 3New York University, 4Lawrence Berkeley National Laboratory, 5Université Paris-Saclay, Université Paris Cité, CEA, CNRS, AIM, 6Princeton University, 7University of Southern California, 8University of Cambridge, 9, Univeresity of Oxford, 10IDRIS, CNRS, 11RIKEN Center for i THEMS, 12University of Wisconsin, Madison |
| Pseudocode | No | The paper describes methods and architectures in prose and refers to figures, but does not contain explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our model release is entirely open source, including the dataset, training script, and weights: https://github.com/Polymathic AI/AION. Code is included in supplemnetal material. |
| Open Datasets | Yes | AION-1 is pretrained on the publicly available data from the Multimodal Universe (hereafter MMU) [66], a large-scale dataset of ML-ready, multimodal astronomical data. We refer the reader to Appendix A for full details on the pretraining data. |
| Dataset Splits | Yes | The head is trained on 80% of the sample with class-stratified splits and evaluated on the remaining 20%. |
| Hardware Specification | Yes | To achieve a batch size of 8192 in all cases, we train AION-1-B using 64 H100 GPUs for 1.5 days, AION-1-L using 100 H100 GPUs for 2.5 days, and 288 H100 GPUs for 3.5 days. |
| Software Dependencies | No | The paper mentions software like Adam W optimizer and PyTorch's FSDP ZeRO-2 strategy, but does not provide specific version numbers for these or other key software components. |
| Experiment Setup | Yes | We train three model versions Base (300M), Large (800M), and XLarge (3B) using the Adam W [37] optimizer (β1 = 0.9, β2 = 0.95, weight decay 0.05) for 205k steps with a global batch size of 8096. We use a linear warmup and cosine decay schedule, with a peak learning rate of 2 10 4. We adopt an input budget of 256 tokens, and output budget of 128 tokens for all our models during pretraining. |