Reproduce “Loop the Loopies!” 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

Zitian Gao, Yilong Chen, Yihao Xiao, Xinyu Yang, Ran Tao, Joey Zhou, Bryan Dai

Loopie tests whether layer-loop recurrence improves training under a matched pre-training budget. It applies each Attention/MoE layer twice before advancing, using freed memory to increase microbatch size. Against a Qwen3-like 30B-A3B baseline, Loopie-20B-A2B initially trails, then passes it after roughly 600 billion tokens on an eight-benchmark mean; this reflects matched realized training cost, not equal theoretical FLOPs or inference efficiency.

Summary by alphaXiv.

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