2026 workshop program
Foundation Models for Science: Lessons from the Transient Sky
The work
Abstract
Astronomical surveys now routinely generate petabytes of imaging data across the electromagnetic spectrum. Hidden within them are variable and explosive phenomena: stellar eruptions, core-collapse supernovae, and tidal disruptions of stars by massive black holes. Characterizing these sources requires models that learn from noisy, irregularly sampled, multimodal observations, a regime where standard foundation-model recipes built for dense and abundantly labeled data break down. In this talk, I will distill lessons from building foundation models for time-domain astrophysics: that tokenization is a modeling decision rather than a preprocessing step; that cross-modal structure provides a supervision signal where labels are missing; and that compression provides a measurable test of physical understanding. I will end by outlining how domain-specific models are becoming the instruments of an agentic ecosystem for science, and what new modes of automated discovery this makes possible.
The speaker
Biography
Alex Gagliano is a Member of Technical Staff at Physical Superintelligence, where he develops agentic systems for physics simulation and scientific discovery. He previously held an NSF IAIFI Fellowship at MIT and the Harvard–Smithsonian Center for Astrophysics, where he built machine learning models for transient science and studied the final years of massive stars.