Reproduce “ReSI-Math: Exploring Recoverable Capability Frontiers for Efficient Reasoning Improvement” 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
Yao Cui, Zhen Yang, Yuandong Wang, Yangfu Zhu, Tianyu Wo, Zhenzhou Shao
ReSI-Math tests whether a math model’s failed answer can be repaired without changing the model. It searches and recombines the model’s sampled attempts, with an external verifier checking offspring, then selects problems for one-problem RLVR and re-profiles after updates. On Qwen3-VL-8B, 272 of 621 correct–failed offspring were verified correct (43.8%); this is an offspring success rate, not a problem-recovery rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.