Reproduce “ISO: An RLVR-Native Optimization Stack” 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
Hanqing Zhu, Wenyan Cong, Zhizhou Sha, Sagnik Mukherjee, Xinyuan Song, David González-Martínez, Xiaoxia Wu, Yuandong Tian and 3 more
The authors propose Isospectral Optimization (ISO), which fixes a model’s singular-value spectrum while adapting both input and output singular frames for RLVR. On Qwen3-4B-Base, ISO-AdamW matches AdamW’s 400-update average@16 math accuracy by update 260 and finishes at 46.98 versus 44.82. In the profiled run, it reaches the matched score 124 hours sooner, a result limited to this setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.