MeRF includes a natural-language description of the reward specification in the training prompt. This gives the model explicit information about its objective while reinforcement learning supplies external feedback. Experiments study the benefit of aligning these two signals. The study also examines how the learner responds when the prompt and reward function disagree.