RL Training

A large-model training process based on reinforcement learning, such as PPO or GRPO, that optimizes a policy through a sampling-encoding-forward pass-update loop. Inconsistency in message handling between training and inference stages causes input distribution shift and training failure.