Reinforcement Learning with Verifiable Rewards Enhanced by Rare-Event Amplification

A new arXiv paper proposes a method to improve reinforcement learning with verifiable rewards (RLVR) for training large language models on deterministic reasoning tasks. The authors argue that effective prompt selection should provide both reliable positive anchors and explicit negative learning signals from rare failures. They introduce positive-negative pairing, sampling a hard-but-solvable prompt and an easy-but-brittle prompt, and Weighted GRPO to reweight binary outcomes. This approach aims to stabilize optimization and improve transfer performance.