2026 workshop program
The Platonic Universe: Do Foundation Models See the Same Sky?
The work
Abstract
We test the Platonic Representation Hypothesis (PRH) and its Aristotelian refinement (ARH) using diverse astronomical datasets to measure representational convergence across foundation models.
We propose astronomy as a natural testbed for this question: the historical success of astrophysics is itself evidence that a compact, modality-invariant description of galaxy observables exists, meaning that representational convergence toward reality should be measurable against the physical parameters astronomers already use.
Within this framework, we evaluate eleven foundation model families spanning supervised classification, self-distillation, joint-embedding prediction, masked autoencoding, vision-language pre-training, and astronomy-specific architectures ranging from approximately 10 million to 10 billion parameters. The models are tested on crossmatched JWST, HSC, and Legacy Survey imaging together with DESI spectroscopy.
All models are evaluated in frozen form, without astronomy-specific fine-tuning.
We probe redshift, stellar mass, and specific star formation rate using linear probes, as well as local (MKNN) and global (CKA) embedding geometry within families, between modalities, and across architectures.
We find that physics performance scales predictably with model capacity; probe directions align consistently with expected astrophysical correlations and selection effects; and local, though not global, embedding alignment tracks physics performance, including between DESI spectra and HSC imagery, modalities that share essentially no low-level statistics.
Our results support the ARH over the strict PRH and suggest that astro-foundation models can effectively build on general-purpose pre-trained architectures, benefiting from the broader machine learning community’s existing computational investment.