Difference between the single-turn training and expanded sing-turn training

[threegold116]

Is the difference that the trajectory of the single-turn comes from data, while that of the expanded single-turn comes from rollout?