acceptodds
Withdrawn from ICLR 2027

The Platonic Universe: Do Foundation Models See the Same Sky?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

We investigate when foundation models converge towards shared representations, and how this convergence depends on model capacity, training regime, and model architecture. We take a `science-for-AI' approach, using astronomy as an experimental instrument to test the Platonic Representation Hypothesis (PRH) and its Aristotelian refinement (ARH) against an external physical reference. The historical success of astrophysics is evidence that a compact, modality-invariant description of galaxy observables exists, and so representation convergence toward reality should be measurable against the physical parameters astronomers already use. Given 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 from (10M)(10B) parameters) on crossmatched JWST, HSC, and Legacy Survey imagery, and DESI spectroscopy. All models are evaluated frozen, with no astronomy-specific fine-tuning. We probe redshift, stellar mass, and specific star formation rate via linear probes, and local (MKNN) and global (CKA) embedding geometry within families, between modalities, and across architectures. We find that physics performance scales predictably with capacity; probe directions align consistently with expected astrophysical correlations and selection effects; and local (but 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, demonstrate astronomy's value as an experimental framework for neural representation learning, and suggest that astro-foundation models can build on general-purpose pre-trained architectures, capitalizing on the broader open machine learning community's already-spent computational investment.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
foundation models, astronomy, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RHHpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper empirically tests the Platonic Representation Hypothesis (PRH) and its Aristotelian refinement (ARH) for representational convergence in foundation models, using multi astronomical observations as physical reference frame. The work systematically analyzes how model capacity, training regimes, and architectural types shape patterns of representational convergence. The study covers 11 foundation model families spanning major training paradigms, supervised classification, self-supervised knowledge distillation, masked autoencoding, vision-language pre-training, and astronomy-specific domain data, with parameter scales ranging from 10M to 10B. The authors use linear probing to test models’ ability to encode three core astrophysical parameters: redshift, stellar mass, and specific star formation rate. They further quantify similarity across representation spaces using two metrics: Mutual k-Nearest Neighbor (MKNN), which captures local neighborhood alignment, and Centered Kernel Alignment (CKA), which measures global geometric alignment. Key findings include: (1) Larger models encode more physics knowledge; (2) Physics performance is correlated between modalities; (3) only local representational alignment strongly correlates with physics performance, and this pattern holds even for cross-modal scenarios such as DESI spectra vs. HSC images. The results suggest that astro-foundation models can build on general-purpose pre-trained models, leveraging the substantial computational investment already made in the broader machine learning community.

Critical strengths
  1. By leveraging observations of the same galaxies from multiple astronomical instruments and using physical parameters independent of the training distribution as the benchmark, this work addresses a limitation of prior PRH/ARH studies: their reliance on circular validation confined to the natural image domain and the corresponding lack of an independent ground-truth reference. This design improves the objectivity of the conclusions. However, the novelty of this work remains ambiguous due to the absence of a dedicated related work section.

  2. The study covers mainstream training paradigms and architectural families of foundation models, and controls for confounding variables across model families via ANCOVA.

Critical weaknesses
  1. Lacks a dedicated Related Work section, leaving the academic contribution poorly contextualized. The paper only mentions related studies sporadically in the introduction, and does not systematically synthesize the research landscape of representational alignment metrics, representational convergence, and foundation models in astronomy. Readers cannot clearly delineate how this work advances beyond the original PRH study, the ARH refinement, and prior work on astronomy models, making it difficult to accurately identify its specific innovations and scholarly value.

  2. Research is primarily descriptive, with limited mechanistic explanation and practical guidance. The core conclusions of the paper remain at the level of phenomenological statistics, without in-depth investigation of the underlying mechanisms. For example, it does not explain why convergence manifests only in local neighborhood structure rather than global embedding geometry (Sec. 3.2), why distillation training causes models to deviate from the general convergence trajectory, or why the language backbone impairs encoding of physical representations from visual inputs (L316-323). Furthermore, the study does not derive actionable recommendations for architectural optimization or training strategy adjustments based on its findings, limiting both the academic impact and applied value of the work.

  3. The authors have modified the in-text citation style, which deviates from the official specifications defined in the ICLR 2027 submission template.

Questions for authors
  1. Does the scaling relationship between model capacity and physics encoding ability still hold under random initialization? Sec. 3.1 concludes that larger models encode more physics knowledge, and that this effect is independent of pre-training data type (natural images, vision-language data, or astronomical data). It is recommended to supplement linear probing experiments with randomly initialized, frozen backbones, using the DINO-v3 ViT architectures from small version to 7B parameters, to verify whether the “larger models encode more physics” phenomenon persists without pre-training.
Reviewer_FKtWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

in this work, the authors perform an extensive study on evaluating the properties of hidden representations of deep neural networks (pretrained on some arbitrary vision tasks) by contrasting them against various known astrophysical properties. that is, they want to know if these large-scale models capture features that encode (or can decode out) astrophysical properties and whether astronomical relationships are encoded (either locally or globally) in these hidden representation space.

although i applaud their effort, i am having a difficult time reading their reported results (e.g. plots, etc.) and agreeing with their observations and lessons. for instance, consider Fig. 1: from which the authors drew the conclusion that there’s a good correlation between # of parameters and mean R2 (against three physical properties). sadly it’s really difficult for me to see that there is indeed a linear trend. for some models , i agree (e.g. ViT-MAE), but if you look at other models (e.g. PaliGemma-2), i can even say that there’s a negative linear trend.

this diversity in the trends across different models shows up across all plots and conclusions they present. this makes me think that the conclusion isn’t really about aristotelian or platonic but that it really matters which data we use to train which network to see any particular properties. since what data is used and how these networks are trained weren’t controlled, i would have expected to see randomly initialized models of these architecture families as well as prematurely trained ones, in order to draw any conclusion from this study.

with a bit more carefully controlled study and analysis, i believe there can be an interesting finding here, but from the current state of the manuscript, i cannot really follow the authors’ findings and reasoning.

Critical strengths

see my answer above.

Critical weaknesses

see my answer above.

Questions for authors

see my answer above.

Reviewer_mHYkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the convergence properties of foundation models in astronomy. The authors tested both the Platonic Representation Hypothesis (PRH) and the Aristotelian Representation Hypothesis (ARH) on multimodal astronomy data. Through a set of experiments across models and modalities, they show there is stronger evidence for consistent local neighbourhood structure in the representation space instead of a global shared embedding.

Critical strengths

This paper presents an interesting study on astronomy— to test if the platonic representation hypothesis (PRH) and its Aristotelian version (ARH) on the galaxy dataset. It summarises several surprising results and shows that the ARH serves as a better explanation for at least this astronomy dataset they test on. It was a fun read to understand the latent behaviour of the state-of-the-art foundation models in this specific galaxy dataset.

Critical weaknesses
  1. The biggest weakness is the venue fit and methodological novelty. I believe this paper is very interesting for domain experts in astronomy, but for the machine learning community, I do not immediately see new insights or inspirations from this study, mainly because it is a single study tested on one specific scientific dataset and scientific domain, based on two already published hypotheses (PRH and ARH).
  2. The paper itself is not very friendly for ML researchers to read, at least for me — I do not have a very good idea of how difficult it is to infer these astronomy-related physical quantities, I also cannot imagine the difference between different imaging modalities (JWST, HSC) — some visualisation examples in the appendix or a nice figure 1 would be very helpful. As a submission to a ML conference, it would be much better if this paper had provided a more thorough explanation on the dataset itself, where the difficulties lies in, why these quantities are important, what are the reasons, why do we want to estimate them etc. This would make the motivation of this paper much clearer and bring impact beyond testing two known hypotheses on a new dataset.
Questions for authors
  1. Which characteristic of this galaxy dataset could have likely led to this result? How generalisable the statement is across different domains? E.g. single cell images, satellite imaginary for tropical cyclones?
  2. Which new insights could we gain from this study? How would this work inspire better representation learning algorithms?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.