Reproduce “A Mechanistic Interpretability Study of an Astronomy Foundation Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rachel C. Zhang, Alberto Bietti, Wes Gurnee, Alessandro Favero, Miles Cranmer
The authors investigate how image information improves photometric redshift predictions in a frozen AION-1 encoder. On the same held-out DESI Bright Galaxy Survey sample, a readout using nine photometry scalars and a four-band image reaches R²=0.931 versus 0.869 with photometry alone. Interventions show spatial extent is distributed across heads, while Head 3 matters for dust-attenuated galaxies; its function remains unidentified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.