During reinforcement learning (RL) post-training, training trajectories are generated on the fly by running inference with the current policy model.