How would you design a reward function for an RL agent tasked with optimizing the inference latency of a transformer model?
