Minimizing KL(q||p) = maximizing rewards (aka policy gradient) - a KL penalty

Twist functions