Yao Cui, Zhen Yang, Yuandong Wang, Yangfu Zhu, Tianyu Wo, Zhenzhou Shao
ReSI-Math introduces a framework that identifies recoverable capability frontiers in large language models, where failed reasoning trajectories can be corrected through evolutionary search guided by self-judgment. It enables data-efficient reinforcement learning with verifiable rewards by adaptively selecting frontier problems, improving math reasoning performance on Qwen3-VL-8B by over 1.4 points on MATH500 after three RLVR stages.
This work challenges the assumption that model failure equals lack of capability, offering a practical path to unlock latent reasoning abilities without costly retraining. By focusing on recoverable failures, ReSI-Math significantly improves efficiency in refining mathematical reasoning, making RLVR more scalable and effective for real-world deployment.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.