腾讯混元研究将经典临界批次大小理论扩展到在线 LLM 强化学习，发现通过重新调整学习率，可在一定批次范围内保持每响应学习效果。在固定硬件上，增大批次使 PPO 生成阶段吞吐量最高提升 2.29×，最佳 GRPO 配置达到相同验证目标的时间减少 29%。
⚡️ As LLM reinforcement learning scales to larger GPU clusters and more training data, training efficiency becomes a first-order concern.
Our new research revisits classical critical-batch-size theory and extends it to online LLM RL, where the model generates its own training data and rollout generation and training scale differently.
Across GRPO and PPO, we find that learning-rate retuning can preserve learning per response over a bounded range of batch sizes.
On fixed hardware, scaling up the batch size improves PPO generation-stage throughput by up to 2.29×, while our best measured GRPO configuration reaches the same validation target in 29% less time. 🚀
Read the full research:
https://hy.tencent.ai/research/100116
来源：Tencent Hy · x.com