Efficient Resource-Shiftable RL Post-training with Generative Rewards
PivotRL is an RL post-training system that dynamically moves computational resources between different roles based on real-time workload demands to maximize throughput. It solves the problem of static resource allocation in RL post-training, where reward evaluation is computationally expensive. The system achieves this through a global parameter storage, a shift-aware scheduler, and a planner that decides optimal role transitions.