AstroAI Lunch Talks - January 12, 2026 - Kshitij Duraphe
12 Jan 2026 - Joshua Wing
The video can be found here: https://www.youtube.com/watch?v=NIf-QQikukE
Speaker: Kshitij Duraphe (UniverseTBD/Stealth)
Title: The Platonic Universe: Do foundation models see the same sky?
Abstract: A growing subset of astronomers has begun to experiment with foundation models trained on heterogeneous data, yet it remains unclear whether these different models learn compatible internal representations of the same underlying astrophysics. In this talk, I present an empirical study of representational convergence in both general-purpose and astronomy-specific foundation models: when models differ in architecture, training objective, and input modality, do their learned feature spaces nonetheless align in a way that suggests a shared latent description of astrophysics? In general-purpose text and vision models, such convergence has been reported as the Platonic Representation Hypothesis (PRH), but analogous tests in astronomy remain limited. Building on our prior work showing early evidence of this trend across general-purpose and astronomy-specific models, we reproduce a scaling pattern consistent with PRH for several self-supervised and astronomy-specialized models—representational alignment often increases with model capacity. However, this behavior is not universal: contrastively trained models frequently weaken or break the expected convergence signal. To characterize these differences, we introduce complementary metrics that probe alignment from multiple geometric perspectives and analyze whether, where, and at what depths convergence emerges. Finally, to connect representational similarity to scientific utility, I go beyond global scores and examine local distributional structure around individual representation elements in both general-purpose and astronomy-specific models, testing whether neighborhoods and activations reflect stable, physically meaningful factors of variation. Together, these results give some insight as to when representational convergence should be expected in astronomical foundation models and motivate diagnostics that more directly target the emergence of physics in learned embeddings.