Reproduce “A JEPA Recipe for Tabular Foundation Models” 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
Mingyu Jeon, Suwan Cho, Jae Young Suh
The authors add a JEPA-style latent-prediction objective to a tabular foundation model, using an EMA target encoder while letting the value head read the encoder field rather than the predictor. In a 20,000-step test, this avoided collapse: MSE was .5513 versus .9364 and effective rank 98.1 versus 1.5. At plateaus, JEPA lost to value-only prediction on both tasks in this setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.