Reflection AI trained Beam using intensive reinforcement learning that penalized the model whenever it generated unnecessary tokens to complete a task