Researchers have proposed a new method for scheduling reinforcement learning with verifiable rewards (RLVR) tasks. The approach uses a graph-based online difficulty estimator to assign exploration budgets based on the difficulty of each task. This helps to reduce redundant rollouts and improve overall performance. The framework can be integrated into existing sample-selection and rollout-allocation schedulers, making it a plug-and-play solution for RLVR tasks. --- Why it matters: This matters because efficient scheduling of RLVR tasks is crucial for large language models' reasoning capabilities. By adapting exploration budgets to task difficulty, this method can improve the performance of these models and reduce computational costs. Source: https://arxiv.org/abs/2608.17941