Reinforcement learning with verifiable rewards (RLVR) has become an effective approach for improving mathematical reasoning in large language models. However, a fundamental question remains: when a model fails to solve a problem, does it truly lack the underlying capability? In this work, we reveal that direct failure does not necessarily imply inability. Some reasoning failures can be recovered by exploring the model's own trajectories, exposing a recoverable capability frontier: a policy-dependent region where solutions are not reliably generated through direct decoding but can be elicited through trajectory-level search. Based on this observation, we introduce ReSI-Math, a framework that explores recoverable capability frontiers and converts recovered capabilities into data-efficient RLVR supervision. ReSI-Math uses Self-Judge-guided evolutionary search to identify recoverable failures under a fixed policy and adaptively selects frontier problems for sequential RLVR as the policy evolves. On Qwen3-VL-8B, correct-incorrect trajectory crossover produces 272 verifier-confirmed correct offspring from 621 attempts (43.8%). After three sequential one-shot RLVR stages, the MATH500 and WE-MATH average improves from 78.71 to 80.16. Under matched single-problem RLVR, random selection reaches only 53.31 versus 54.16 for frontier-guided Q1.
Universal Abstraction: Harnessing Frontier Models to Structure Real-World Data at Scale
Microsoft Research(微软研究院);Johns Hopkins University School of Medicine(约翰霍普金斯大学医学院);Providence Portland Medical Center(普罗维德波特医疗中心);Earle A. Chiles Research Institute(埃勒·A·奇尔斯研究所);Providence Research Network(普罗维德研究网络);Providence Genomics(普罗维德基因组学);The Oregon Clinic(俄勒冈诊所)
Constellation University of Wisconsin-Madison(威斯康星大学麦迪逊分校星座研究所);Constellation Georgia Institute of Technology(佐治亚理工学院星座研究所);UK AI Security Institute(英国人工智能安全研究所)
Desert Camels and Oil Sheikhs: Arab-Centric Red Teaming of Frontier LLMs
MBZUAI(穆罕默德·本·扎耶德人工智能大学);University of Khartoum(喀土穆大学);Universität des Saarlandes(萨尔大学);University of British Columbia(不列颠哥伦比亚大学);Invertible AI(可逆人工智能公司);Vector Institute(向量研究所)