Reproduce “Zero-Shot Weight-Space Programming for Prediction and Classification of Target Variables in Structured Datasets” 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

Eduardo C. Garrido-Merchán

The paper tests whether language models can infer predictors from labeled examples and write their network parameters directly, without code, gradients, optimization, or pre-submission evaluation. Claude Opus 5’s eight-unit ReLU network scored 1.0000 on 120 held-out s3_xor rows, versus 0.9662 for gradient descent. This finite-test score does not establish accuracy above the true Bayes-optimal rate.

Summary by alphaXiv.

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