Reproduce “Procedural Adversarial Learning” 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
Jack Nugent, Jia Deng
Simple-PAT trains models to reduce discovered worst-case errors on procedurally generated perception tasks. It searches controls with CMA-ES, generates labeled examples, and trains on them; black-box evaluation searches the same parameter spaces after training. Across four PAL-Dojo generators, it reduces discovered worst-case error by a geometric-mean 2.63× versus the strongest robust baseline per task, within finite searches, not real-world guarantees.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.