Reproduce “Weight-Space Programming Doesn’t Need Transformers” 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

Hyunjong Ok, Jaeho Lee

The paper studies WSP, where an LLM writes another model’s weights directly, without gradient training. The authors let LLMs choose model structures; three writers independently produce sparse fact tables. Claude Opus 5.5’s first-generation table reaches 50.8% ARC accuracy with 0.94 million parameters, about 1,800 times fewer than the 1.72-billion-parameter Qwen3-1.7B model written by that writer. The finding is limited to untrained models and multiple-choice recall.

Summary by alphaXiv.

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