metadata
task_categories:
- other
When to Align, When to Predict: A Phase Diagram for Multimodal Learning
This repository contains pretrained encoders, cached features, and phase diagnostics data associated with the paper "When to Align, When to Predict: A Phase Diagram for Multimodal Learning".
The paper characterizes when contrastive (Cross-Alignment, CA) vs. predictive (Cross-Prediction, CP) self-supervised objectives recover shared signals in multimodal data, particularly for scientific domains.
Project Website | GitHub Repository
Dataset Description
The data stored here supports experiments including:
- Astro Data: Pretrained encoders and cached features for LAMOST × Kepler/TESS spectra.
- Phase Diagnostics: Data used to locate real-world datasets in the phase diagram regimes (Both, CA only, CP only, or Neither).
Sample Usage
You can download the checkpoints, cached features, and phase diagnostics data using the huggingface_hub library:
from huggingface_hub import snapshot_download
snapshot_download("Ilayk/mm_align_vs_pred", repo_type="dataset", local_dir="hf_data")
After downloading, you can point your environment to the data for the astrophysical experiments:
export MULTIDESA_ROOT=hf_data
Citation
@article{kamai2026align,
title = {When to Align, When to Predict? A Phase Diagram for Multimodal Self-Supervised Learning},
author = {Kamai, Ilay and Van Assel, Hugues and Regev, Aviv and Perets, Hagai B. and Balestriero, Randall},
journal = {arXiv preprint arXiv:2606.11190},
year = {2026}
}