Reproduce “Scaling Latent Reasoning via Looped Language Models” 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
Rui-Jie Zhu, Zixuan Wang, Kai Hua, Tianyu Zhang, Ziniu Li, Haoran Que, Boyi Wei, Zixin Wen and 25 more
Ouro reuses the same Transformer layers across recurrent steps, allowing hidden states to be refined without adding distinct weights. It learns when to stop, providing an internal computation scaling axis. On MATH500, its 1.4B model at four steps scores 82.40%, versus 59.60% for Qwen3-4B. This comparison does not control for training data or compute, and does not establish faster inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.