Tencent Huan Yuan team redefines batch size for large model reinforcement learning to boost efficiency. By recalibrating critical batch size and adjusting learning rates, their approach optimizes online RL algorithms like PPO and GRPO. This method resolves mismatches between rollout generation and training scales, ensuring each response learns effectively without data dilution. Real-world tests show a 2.29x throughput increase for PPO and a 29% time reduction for GRPO on fixed hardware. This strategy enables training teams to maximize GPU utilization, significantly cutting costs and accelerating model development without requiring larger clusters or thicker data.