Reproduce “GenPFN: An In-Context Tabular Foundation Model Approach for Augmenting Small Datasets in Engineering” 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

Noah J. Bagazinski, Cyril Picard, Faez Ahmed

GenPFN proposes engineering designs from labeled examples using a pretrained TabPFN classifier with fixed weights. Candidates start between desired designs, receive probability-gradient guidance below a threshold, and are varied with noise. Across fourteen problems, feasibility exceeded CTGAN by 69.54 percentage points. Evaluated additions reduced validation MSE by 62.1% versus context-only training, but guidance alone was not shown to explain the reduction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.