policy

The policy model in reinforcement learning, which evolves continuously during RL post-training; changes in its distribution cause model-based drafters to become mismatched.