Tencent Hunyuan Research Improves LLM RL Training Efficiency by 29%
Tencent Hunyuan's research revisited critical-batch-size theory for online LLM RL, finding that learning-rate retuning preserves learning per response over a range of batch sizes. Scaling batch size improved PPO generation throughput by 2.29× and GRPO reached validation target 29% faster.